Research ArticleAn IoT-Based Motion Tracking System for Next-GenerationFoot-Related Sports Training and Talent Selection

Shanshan Lu1 Xiao Zhang 1 Jiangqing Wang1 Yufan Wang2

Mengjiao Fan2 and Yu Zhou 3

1Department of Computer Science South-Central University for Nationalities Wuhan 430074 China2Department of Industrial Engineering amp Management School of Mechanical Engineering Shanghai Jiao Tong UniversityShanghai 200240 China3College of Computer Science and Software Engineering Shenzhen University Shenzhen China

Correspondence should be addressed to Xiao Zhang xiaozhangmycityueduhk and Yu Zhou yuzhouszueducn

Received 4 March 2021 Revised 7 May 2021 Accepted 12 June 2021 Published 25 June 2021

Academic Editor Chi-Hua Chen

Copyright copy 2021 Shanshan Lu et al is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Motion tracking in different fields (medical military film etc) based on microelectromechanical systems (MEMS) sensingtechnology has been attracted by worldrsquos leading researchers and engineers in recent years however there is still a lack of researchcovering the sports field In this study we propose a newAIoT (AI + IoT) paradigm for next-generation foot-driven sports (soccerfootball takraw etc) training and talent selection e system built is cost-effective and easy-to-use and requires much fewercomputational resources than traditional video-based analysis onmonitoring motions of players during traininge system builtincludes a customized wireless wearable sensing device (WWSDs) a mobile application and a data processing interface-basedcloud with an ankle attitude angle analysis model Eleven right-foot male participators wore the WWSD on their ankle while eachperformed 20 instances of different actions in a formal soccer fielde experimental outcome demonstrates the proposed motiontracking system based on AIoT and MEMS sensing technologies capable of recognizing different motions and assessing theplayersrsquo skills e talent selection function can partition the elite and amateur players at an accuracy of 93 is intelligentsystem can be an emerging technology based on wearable sensors and attain the experience-driven to data-driven transition in thefield of sports training and talent selection and can be easily extended to analyze other foot-related sports motions (egtaekwondo tumble and gymnastics) and skill levels

1 Introduction

Recent trends in smart wearable technologies based on theInternet of ings (IoT) have opened up a large number ofapplications [1] which involve the recognition of sportsactivities [2ndash4] Soccer also known as football in somecountries is one of the most popular sports in the world withnumerous professionals and an even larger number ofnonprofessional practitioners Soccer players train a specificset of well-defined motions (eg shooting and passing) toconsolidate them into muscle memory and lower their re-action time during a game e correct execution of theexercises and training has a steep learning curve Quantityand quality information about soccer motions like shooting

and passing an exhaustive performance evaluation is in-dispensable for coaches and players in training sessions andcompetitions

Traditionally the training of professional soccerplayers is human-oriented which is done by subjectiveguidance based on the personal experience of coachingstaff or trainers However different coaches could havedifferent ideas upon their experience and there are manyfast and slight soccer motions during an action thatcannot be captured through their visual sense e needfor an objectively and digitized method rather than asubjective method is raised In the last few years therehave appeared lots of research efforts in the field ofprofessional soccer for a measurable and quantifiable

HindawiJournal of Healthcare EngineeringVolume 2021 Article ID 9958256 14 pageshttpsdoiorg10115520219958256

analysis of sports through information technology emanager and coaching staff use videography to monitorthe biomechanics of soccer actions and check the objectiveplayersrsquo performance [5] Despite this videography istypically restricted by equipment and environment andunable to give feedback to the coaching staff or soccerplayers in real time due to the significant amounts ofstorage and computational load

e rapid development of bluetooth low energy (BTLE)and microelectromechanical systems (MEMS) technologiesimplements durative motion data capture and communi-cation in real time ere is increasing research on motionrecognition by using inertial low-cost sensors worn on thebody For example the authors of [3] use inertial sensorsworn on the wrist to identify the stroke type of table tennisIn [6 7] wearable sensing devices (WSDs) on the wrist areused to classify different volleyball and badminton actionsHowever due to the nature of soccer the ankle movementsof players are complex and indicative of different soccermotions Moreover the high variability in the training andindividual execution of each exercise makes the classifi-cation of soccer motions extremely challenging A par-ticular challenge we address is to classify the mostfundamental soccer motions (ie shooting and passing)(the development process and methods we describe in thispaper can be reused in more motion recognition appli-cations) e existing approaches used in other sports arehardly applicable [8]

erefore an IoT system based on WSD is proposed inthis work to provide objective feedback to coaches andsoccer players after or during a training session to help themimprove their skills e proposed IoT system consists ofwearable devices a mobile device (eg a mobile phone ortablet) and a data processing platform based cloud A WSDof MEMS is used to collect raw data from soccer players Byusing BTLE technology the material is transmitted to thecloud-based data processing platform which analyzes thedata and outputs the results to a mobile device in real time Asupport vector machine (SVM) model classification algo-rithm with an ankle-based attitude angle model is proposedto recognize different motions and assess different skilllevels

is system is ideally suitable for the young player soccertraining A soccer training session typically includes exer-cises of passes crosses and shots For example in a soccerclub or training school there are dozens or hundreds ofyoung players conducting an exercise of shooting or passingIt is evident that a limited number of coaches are hard toassess all soccer playersrsquo individual execution of each ex-ercise Instead by using our developed system how manypassesshoots have been done and how professional will beshown on a coachrsquos mobile device e work we presentenables soccer players to keep track of the training exercisesthey perform e managers and coaches could get anoverview of the exercises during past days weeks ormonths which they could use to arrange future trainingsessions pointedly We intend to apply the system in arealistic training environment with a focus on the recog-nition of basic soccer motions is study demonstrates the

trend in using the IoT framework for a new era of soccertraining

Our key novelty andmain contributions are summarizedas follows

(i) We develop an IoT system for soccer motion rec-ognition and assessment in which a WSD withoverall dimensions of 2mm times 103mm times 87mm isdeveloped to capture inertial data a mobile appli-cation allows visualization of experimental outputsand transfers the data to the cloud-based processingplatform

(ii) We build a SVM classification algorithm with anattitude angle model to classify different soccermotions and make a distinction between the skilllevels between elite and amateur players

(iii) Our system is confirmed by the experimental resultsthat it enables recognizing different motions that isshooting and passing with an accuracy close to 90e talent selection distinguishes between the eliteand amateur players with an accuracy of 93

We organize the remainder main structure of this articleas follows we give a literature review in Section 2 en weillustrate the system composition and data acquisition inSection 3 Next Section 4 presents our algorithm for soccermotion recognition and assessment as well as the experi-mental results Finally Section 5 summarizes thecontribution

2 Related Work

Motion analysis and recognition by using informationtechnology have been a field of study for decades Duringthis period a variety of motion recognition applications andapproaches have been proposed to assess the performanceand correctness of a physical exercise of an athlete andprovide feedback to users about their actions

In early studies motion recognition was a kind ofcomputer vision field that recognizes human pose or actionin which video image acquisition and processing technologywas adopted to identify the human bodyrsquos motion behaviorIn vision-based action recognition the training method ismainly to collect videos images and other information toanalyze the moving process erefore it is necessary toplace cameras and other monitors in advance in the de-tection environment for data collection e popularity ofthe RGB camera has made it an effective auxiliary tool for thehuman motion recognition study and also leads to theappearance of several survey articles [9 10] discussingvarious features and classifiers for human action recogni-tionough image analysis technology can identify peoplersquosdaily moves more accurately image analysis technology stillsuffers from many shortcomings In nature a large numberof hardware conditions are required so as to run computervision algorithms and computationally concentrated imageprocessing Moreover a large amount of image data ac-quired suffers from insufficient storage leading to failure inmonitoring [11]

2 Journal of Healthcare Engineering

Several human activitiesrsquo recognition approaches havebeen presented in which wearable sensors with acceler-ometers and gyroscopes are used to acquire motion data[12 13] A much wider field of views can be obtained by thissensor technology Long-running recordings calculationand constant interaction are possible owing to the progressof inertial sensors in energy-consuming reduction and theimprovement of computational power Moreover 3D mo-tion data which consists of 3-axis angular velocities fromgyroscopes as well as 3-axis accelerations from accelerom-eters can be given by wearable inertial sensors [14 15]

Sensors are often of small size high precision andsensitivity and low environmental requirements andpower consumption as well as easy to wear A universalactivity recognition depth framework based on multimodalwearable sensors was proposed in [16 17] Shoaib et al [18]used inertial sensors to detect the activities that involvehand gestures (smoking eating drinking giving a chatetc) In the field of competitive sports most movementrecognition systems are dedicated to solving wrist move-ments or simple lower limb movements such as walkingand standing [19]

ere are a few studies on the recognition and assess-ment of soccer playersrsquo motions Reference [20] indicatedthat soccer playersrsquo shin guards instrumented with sensorscan measure ankle joint kinematics Meamarbashi et altested whether sensors can be used to examine kinematicparameters of soccer players in [21] but no further analysisof the data In [4] Haladjian et al developed a smart glovewith an inertial sensor and gave a machine learning algo-rithm to recognize the goalkeeperrsquos training exercises whichwas a good attempt to use WSD to recognize soccer motionAs far as we know this is the first work to use a single WSDattached to an ankle to recognize different soccer playersrsquomotions (not just the goalkeepers) and assess their skilllevels

3 System Design and Data Acquisition

oughWSDs have been widely used in motion recognitionfor soccer motions the existing approaches are hardly ap-plicable However the limited number of coaches are hard toassess all soccer playersrsquo individual performance us thisstudy presents a complete system in using the IoTframeworkfor soccer motion recognition By using the developedsystem the passesshoots and their skill levels can be shownon a mobile device hold by a coach e system enablessoccer players to keep track of the training exercises theyperform e specific details of hardware design interfacesoftware and data acquisition are given in this sectionFigure 1 shows an overview of our IoT system for soccermotion recognition and assessment which is composed ofwireless WSDs mobile devices and cloud-based data pro-cessing platform

By using the proposed IoT system BTLE sent mobiledevice data collected byWSD It transfers the raw data to theremote execution server via cloud computing technology atthe moment when the mobile device received the exercise

data Completing all data collection and processing userscan view the analysis results of soccer players through theplatform

31 Hardware and Software System We collect data using awireless WSD jointly developed by our team and thecompany AI Motion Sports (httpswwwaimotionsportscomen) Figure 2 shows the components of the deviceis wireless WSD is comprised of four major componentsan MEMS motion sensing chip with a 3-axis gyroscope andaccelerometer a microprocessing unit with Bluetoothwireless newsletter function a lithium battery and an ON-OFF switch on the back of the board

e 3-axis gyroscope and accelerometer in MEMS areused to sense the motions and transform the signal into rawdata DA14583 Programmed in Dialog SemiconductorReading UK baseband radio processor and entirely inte-grates radio transceiver for BTLE Not only is the micro-processor with chip nRF52832 characterized by low energyconsumption but also highly efficient transmission by usingBluetooth v50 BMI160 from BOSCH has a fitting sensorrange and tiny size (25mm times 30mm times 083mm) [22] toenable it to be adopted

e acceleration and orientation in three dimensions canbe obtained through the high-integration and low-powerBMI160 e entire WSD keep in size within 103mm times

87mm times 2mm using this IMU chip weighing only 2 g Wedevelop a mobile application for receiving and visualizingdata transmitted from WSD in real time e applicationdeveloped by the Java Script consists of three functionalmodules wireless connection real-time captured datademonstration and data synchronization to the cloud Alldata sensed are display synchronously and saved locally Acloud-based approach is also used to save data received to aremote server Once the data collection process is completeduser clients can access the analytical results

32 Data Collection We conduct the experiments at theChinese Football Association (CFA) Youth Academy(Wuhan) (Chinese National Football Team Training Center)and Sports Center of South-Central University for

Cloud

Mobile device

Soccer players

WSD

Figure 1 In the IoT system the motion sensor is attached to theright ankle of the soccer player e motion data is transmitted viaBluetooth to the Cloud platform for further process

Journal of Healthcare Engineering 3

Several human activitiesrsquo recognition approaches havebeen presented in which wearable sensors with acceler-ometers and gyroscopes are used to acquire motion data[12 13] A much wider field of views can be obtained by thissensor technology Long-running recordings calculationand constant interaction are possible owing to the progressof inertial sensors in energy-consuming reduction and theimprovement of computational power Moreover 3D mo-tion data which consists of 3-axis angular velocities fromgyroscopes as well as 3-axis accelerations from accelerom-eters can be given by wearable inertial sensors [14 15]

Sensors are often of small size high precision andsensitivity and low environmental requirements andpower consumption as well as easy to wear A universalactivity recognition depth framework based on multimodalwearable sensors was proposed in [16 17] Shoaib et al [18]used inertial sensors to detect the activities that involvehand gestures (smoking eating drinking giving a chatetc) In the field of competitive sports most movementrecognition systems are dedicated to solving wrist move-ments or simple lower limb movements such as walkingand standing [19]

ere are a few studies on the recognition and assess-ment of soccer playersrsquo motions Reference [20] indicatedthat soccer playersrsquo shin guards instrumented with sensorscan measure ankle joint kinematics Meamarbashi et altested whether sensors can be used to examine kinematicparameters of soccer players in [21] but no further analysisof the data In [4] Haladjian et al developed a smart glovewith an inertial sensor and gave a machine learning algo-rithm to recognize the goalkeeperrsquos training exercises whichwas a good attempt to use WSD to recognize soccer motionAs far as we know this is the first work to use a single WSDattached to an ankle to recognize different soccer playersrsquomotions (not just the goalkeepers) and assess their skilllevels

3 System Design and Data Acquisition

oughWSDs have been widely used in motion recognitionfor soccer motions the existing approaches are hardly ap-plicable However the limited number of coaches are hard toassess all soccer playersrsquo individual performance us thisstudy presents a complete system in using the IoTframeworkfor soccer motion recognition By using the developedsystem the passesshoots and their skill levels can be shownon a mobile device hold by a coach e system enablessoccer players to keep track of the training exercises theyperform e specific details of hardware design interfacesoftware and data acquisition are given in this sectionFigure 1 shows an overview of our IoT system for soccermotion recognition and assessment which is composed ofwireless WSDs mobile devices and cloud-based data pro-cessing platform

By using the proposed IoT system BTLE sent mobiledevice data collected byWSD It transfers the raw data to theremote execution server via cloud computing technology atthe moment when the mobile device received the exercise

data Completing all data collection and processing userscan view the analysis results of soccer players through theplatform

31 Hardware and Software System We collect data using awireless WSD jointly developed by our team and thecompany AI Motion Sports (httpswwwaimotionsportscomen) Figure 2 shows the components of the deviceis wireless WSD is comprised of four major componentsan MEMS motion sensing chip with a 3-axis gyroscope andaccelerometer a microprocessing unit with Bluetoothwireless newsletter function a lithium battery and an ON-OFF switch on the back of the board

e 3-axis gyroscope and accelerometer in MEMS areused to sense the motions and transform the signal into rawdata DA14583 Programmed in Dialog SemiconductorReading UK baseband radio processor and entirely inte-grates radio transceiver for BTLE Not only is the micro-processor with chip nRF52832 characterized by low energyconsumption but also highly efficient transmission by usingBluetooth v50 BMI160 from BOSCH has a fitting sensorrange and tiny size (25mm times 30mm times 083mm) [22] toenable it to be adopted

e acceleration and orientation in three dimensions canbe obtained through the high-integration and low-powerBMI160 e entire WSD keep in size within 103mm times

87mm times 2mm using this IMU chip weighing only 2 g Wedevelop a mobile application for receiving and visualizingdata transmitted from WSD in real time e applicationdeveloped by the Java Script consists of three functionalmodules wireless connection real-time captured datademonstration and data synchronization to the cloud Alldata sensed are display synchronously and saved locally Acloud-based approach is also used to save data received to aremote server Once the data collection process is completeduser clients can access the analytical results

32 Data Collection We conduct the experiments at theChinese Football Association (CFA) Youth Academy(Wuhan) (Chinese National Football Team Training Center)and Sports Center of South-Central University for

Cloud

Mobile device

Soccer players

WSD

Figure 1 In the IoT system the motion sensor is attached to theright ankle of the soccer player e motion data is transmitted viaBluetooth to the Cloud platform for further process

Journal of Healthcare Engineering 3

Nationalities We recruited 11 male soccer players including5 elites and 6 amateurs who are right-footed players Amongthem elite players had represented their club as participantsin more than ten 10 national games from CFA YouthAcademy (Wuhan) amateurs were beginners from uni-versity As soccer is a foot-based sport our small and lightWSD is attached to each right-footed subjectrsquo right anklewhen performing soccer basic training to make sure thatmajor inertial data of soccer motions can be capturedwithout obstruction During the experiment each subjectperformed 20 passings with the inside of the foot and 20shootings with foot-arch at the soccer field of CFA YouthAcademy (Wuhan) e subject performed the actions atthe same position Each subjectrsquos passings and shootingsare required to be performed within a certain range ofspeed and precision otherwise we considered it was aninvalid action and we did not count it Figure 3 illustratesour experimental scene

Figure 4 displays the six-axis synchronal raw datacaptured by the WSD from elites by performing differentsoccer motions e angular velocity as well as accelerationfor passing are displayed in Figures 4(a) and 4(b) whileFigures 4(c) and 4(d) show the related data information forshooting

4 Soccer Motion Recognition and Assessment

e process of the soccer motion recognition system asshown in Figure 5 can be divided into five steps datapreprocessing attitude angle modeling feature extractionand selection dimensionality reduction and classificationTo be specific the ankle-based attitude anglesrsquo features act apivotal part in improving the accuracy of classification sincesoccer as a type of foot-based sport is more complicated thanother sports like racket sports Principle component analysis(PCA) is adopted in our system to balance the complexityand the accuracy Finally we propose a SVM-based clas-sification algorithm to recognize the soccer motions

In data preprocessing a 3-point moving average filter isused to lessen the noisy interference from raw data Bylocating the peak of the signal the segmentation of themotion signal can be processed automatically e sliding-window algorithm [23] can be applied to process the signalsin real time

41 Angle Trajectory Model As shown in Figure 6 the in-tuitive signal of ωx ωy ωz (acceleration from three di-mensions x -axis y -axis and z -axis) and αx αy αz(angular velocity from three dimensions) are not significantWe present an attitude angle model for extracting moreuseful features

For sports motion recognition angular velocity andacceleration are important features Besides the anklesrsquorotations are specially important for soccer motion recog-nition and assessment [16] A three-dimensional rotationproblem is typically addressed by a rotation matrix Miningthe information of attitude angle can improve characterizingthe action e attitude angle is expressed by three Eulerangles yaw pitch and roll Figure 7 shows the specifictransformation

e quaternion [24] as the quotient of two directed linesin a three-dimensional space is used to solve angle in theangle trajectory model e representation of the quaternionis p pω + pxi + pyj + pzk with a real part pω and threeimaginary parts parameters px py and pz Calculate theinitial attitude angle

θ arcsin ax0( 1113857

φ arctanmagy0

magx0

ψ arctan minusay0

az01113888 1113889

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(1)

Lithiumbattery

(a)

DA14583

BMI160

(b) (c)

Figure 2 (a) A lithium battery on the back of the board (b) Circuit board of the WSD (c) e WSD is attached to the right ankle of asoccer player

4 Journal of Healthcare Engineering

where ax0 ay0 az0 represent the initial acceleration θ ψand φ represent the initial yaw roll and pitch respec-tively magx0 and magy0 are calculated by ax0 ay0 az0

specifically magx0 ax0ay0 + ax0az01 minus 2(a2y0 + a2

z0) andmagy0 ax0ay0 + ay0az01 minus 2(a2

x0 + a2z0) [25] e initial

four elements are calculated from the initial angle

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

WxWyWz

(a)

Ang

ular

vel

ocity

(degs

)

0

ndash1200

1200

0 100 200Times

AxAyAz

(b)

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

WxWyWz

(c)

Ang

ular

vel

ocity

(degs

)

0

ndash1200

1200

0 100 200Times

AxAyAz

(d)

Figure 4 Passing and shooting signal collected from an elite subject (a) Acceleration of a passing (b) Angular velocity of a passing(c) Acceleration of a shooting (d) Angular velocity of a shooting

BLE-based IoT sensor

(a)

Target location

Start position 1

Start position 2

(b)

Figure 3e sensor placement and experimental site setting for raw data collection (a)eWSD is attached to a subjectrsquos or playerrsquos rightankle (b) Positions of soccer players to perform passing and shooting

Journal of Healthcare Engineering 5

Acquire Preprocess Partition Modeling Clsssfier

Low-passfilter

segmentation Pendingdata

Labels

Trainingdataset

SVMAlgorithm

Recognition modelor

Assessment modelTest dataset

Angletrajectory

model

Featureextraction

PCA

Raw data

Labels

Figure 5 e process of soccer motion recognition and assessment system

Ang

ular

vel

ocity

(degs

)

0

ndash1200

1200

0 100 200Times

ProfessionalAmateur

(a)

ProfessionalAmateur

Ang

ular

vel

ocity

(degs

)0

ndash1200

1200

0 100 200Times

(b)

ProfessionalAmateur

Ang

ular

vel

ocity

(degs

)

0

ndash1200

1200

0 100 200Times

(c)

ProfessionalAmateur

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

(d)

ProfessionalAmateur

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

(e)

ProfessionalAmateur

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

(f )

Figure 6e elitesrsquo and amateursrsquo passing motion data comparison result (a) (b) and (c) are the contrast of acceleration on the 3-axes and(d) (e) and (f) are the contrast of angular velocity on 3-axes

6 Journal of Healthcare Engineering

p0

p1

p2

p3

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

sinψ2sin

θ2sin

φ2

+ cosψ2cos

θ2cos

φ2

sinψ2cos

θ2cos

φ2

minus cosψ2sin

θ2sin

φ2

cosψ2sin

θ2cos

φ2

+ sinψ2cos

θ2sin

φ2

cosψ2cos

θ2sin

φ2

minus sinψ2sin

θ2cos

φ2

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(2)

e RungendashKutta method [26] which mainly eliminatesthe complicated process of solving differential equationswhen the derivatives and initial value information of theequation are known is used to solve four elements

p0

p1

p2

p3

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

t+t

p0

p1

p2

p3

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

t

+t

2

minusωx middot p1 minus ωy middot p2 minus ωz middot p3

+ωx middot p0 minus ωz middot p2 minus ωy middot p3

+ωy middot p0 minus ωz middot p1 + ωx middot p3

minusωz middot p0 + ωy middot p1 minus ωx middot p2

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(3)

where t means the current time value and t represents theinterval of next data e following uses the quaternion to solvethe attitude anglee coordinate transformationmatrixD of theknownobject coordinate system to the Earth coordinate system is

D

1 0 00 1 00 0 1

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦ + 2 cos

θ2

0 minusn sinθ2

m sinθ2

n sinθ2

0 minusl sinθ2

minusm sinθ2

l sinθ2

0

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

+ 2

minus m2

+ n2

1113872 1113873sin2θ2

lm sin2θ2

nl sin2θ2

lm sin2θ2

minus l2

+ n2

1113872 1113873sin2θ2

mn sin2θ2

nl sin2θ2

mn sin2θ2

minus m2

+ l2

1113872 1113873sin2θ2

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(4)

m l and n mean the direction vector projected onto thegeographic coordinate along the direction of rotation vectore rotation of the object corresponds to the rotationaround the axis e trigonometry of a quaternion iscos θ2 + (l i

rarr+ m j

rarr+ n k

rarr)sin θ2 cos θ2 + u

rarr sin θ2 inwhich i

rarr jrarr

krarr

denote the unit direction vector in thegeographic graticule system

p0 cosθ2

p1 l sinθ2

p2 m sinθ2

p3 n sinθ2

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(5)

e constructed quaternion describes the fixed-pointrotation problem of the substance Correspondinglythe object system is formed by one-time equivalent ro-tation of the Earth system Equation (5) is substituted intoD1

D1

p20 + p

21 minus p

22 minus p

23 2 p1p2 minus p0p3( 1113857 2 p0p2 + p1p3( 1113857

2 p0p3 + p1p2( 1113857 p20 minus p

21 + p

22 minus p

23 2 p2p3 minus p0p1( 1113857

2 p1p3 minus p0p2( 1113857 2 p0p1 + p2p3( 1113857 p20 minus p

21 minus p

22 + p

23

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(6)

Let the unit vector in the coordinate system X be(ex1 ey1 ez1)

T and correspond to (ex2 ey2 ez2)T in Y and

the direction of the projection of the (ex2 ey2 ez2)T is Cn

b Suppose there is a vector R whose magnitude on X is(x y z)T and on Y is (x2 y2 z2)

T then coordinatetransformation is (x y z)T Cn

b(x2 y2 z2)T e attitude

angle transformation matrix is also converted into the co-ordinate transformation matrix form of object coordinatesystem to a geographic coordinate system and Cn

b is obtainedas follows

Z

Y

X

θ

Zprime

Yprime

Xprime

(a)

Z

YφZprime

Yprime

Xprime(X)

(b)

ZY

ψ

ZprimeYprime

X

Xprime

(c)

Figure 7 Attitude angle an exploded view of the process of the attitude angle rotation

Journal of Healthcare Engineering 7

Cnb Rot(zψ)Rot(y θ)Rot(xφ)

cos ψ minussin ψ 0

cos ψ cos ψ 0

0 0 1

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1 0 0

0 cos φ minussin φ

0 sin φ cos φ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

cos θ 0 sin θ

0 1 0

minussin θ 0 cos θ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

cos ψ cos θ minussin ψ cos φ + cos ψ sin θ sin φ sin ψ sin φ + cos ψ sin θ cos φ

sin ψ cos θ cos ψ cos φ + sin ψ sin θ sin φ minuscos ψ sin φ + sin ψ sin θ cos φ

minussin θ cos θ sin φ cos θ cos φ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(7)

Both D1 and Cnb are matrices transforming the object

coordinate system attitude into the geographic coordinatesystem that is

2 p1p3 minus p0p2( 1113857 minussin θ g1

2 p0p1 + p2p3( 1113857 sin φ cos θ g2m

p20 minus p

21 minus p

22 + p

23 cos θ cos φ g3

2 p0p3 + p1p2( 1113857 sin ψ cos θ g4

p20 + p

21 minus p

22 minus p

23 cos ψ cos θ g5

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎩

(8)

e equation for solving the attitude angle by thequaternion is

roll arctan2 p0p3 + p1p2( 1113857

1 minus 2 p20 + p

211113872 1113873

pitch arctan2 p1p3 minus p0p2( 1113857

4 p0p3 + p1p2( 11138572

+ 1 minus 2 p20 + p

211113872 11138731113872 1113873

2

yaw arctan2 p0p1 + p2p3( 1113857

1 minus 2 p22 + p

231113872 1113873

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(9)

e frequency of the sensor repetition rate is 100Hzen the angular velocity value is substituted into the so-lution attitude angle Anminusi Figure 8 shows the comparison ofattitude angle of elites and amateurs

By combining Anminusi with Snminusi as a part of Rnminusi we extractthe basic and morphology features fni of segments R(n)i asshown in Table 1

F(n) (f1 fmn) are merged into one largematrix F

of size n times m e updated data form is shown in Figure 9

42 Principle Component Analysis e cloud server benefitsfrom reducing memory requirements computing load andnecessary bandwidth during model implementation Torepresent identified variables in compact feature variableswe perform PCA processing e model with PCA canachieve higher accuracy than other nonlinear dimensionalityreduction methods [7] PCA is essentially a base transfor-mation that makes transformed data have the largest vari-ance which means the variance between one axis (spindle)and the point is minimized by the rotation of the coordinate

axis and the translation of the coordinate origin Suppose thematrix F with the dimension of n times m which means thatthere is total n samples in m-dimensions F can bedecomposed into UΣVT where U has the same size as F andthe orthogonal matrix V is m times m Σ is a diagonal matrix ofthe same size as V en YrUΣr F(ΣVT)(minus 1)Σr

We extracted a total of 34-dimensional features for avector characterizing where F [F1 F2 F34] AfterPCA the obtained new features can be expressed as Yr

[Yr1 Yr2 Yrm] and m represents the calculateddimension

Yrm F1aij + F2aij + middot middot middot + Fmaij (10)

where aij are eigenvalues of the covariance matrix Equation(10) can be simplified as follows

Yrm F1a1 + F2a2 + middot middot middot + Fmam (11)

43 Support Vector Machine SVM is widely applied in thedomain of machine learning and pattern recognition as atool to solve the classification problem e basic model ofSVM aims to maximize the distance of the closet sampleswhich is called support vector to the hyperplane on theeigenspace [27] In SVM a training data point is regarded asa $p$-dimensional vector the objective is to separate suchpoints using a $(p-1)$-dimensional hyperplane [19] at isthe training points are mapped to points in space to max-imize the width of the gap between the two and severalcategories by SVM Compared to other supervised algo-rithms such as logistics regression (LR) model or NaıveBayes algorithm [28] the SVM algorithm is more suitable todeal with high-dimensional and linear inseparable problemsby freely selecting the parametric model and only usesupport vector as the classification basis of hyperplane whichmeets our expectations as our problem is a small samplelinear inseparable problem

e training dataset is set as T (Xk yk)|X isin Rm yk isin1113864

0 1 nk1 Here are total n samples X represents an

m-dimensional matrix the value of the classification label yk

is 0 or 1 and the current sample is Xk erefore therepresentation of the optimization objective function is

minωb

12ω

2

st yk ω middot xk + b( 1113857 minus 1ge 0 k 1 2 N

(12)

8 Journal of Healthcare Engineering

Here ω is a vector on a hyperplane and the offset b of thesuperflat is along ω from the origin Problem (12) is a convexquadratic programming problem [29] According to theconvex optimization theory transform problem (12) into anunconstrained problem e optimization function can bedenoted as

L(ω b α) 12ω

2minus 1113944

N

k1αkyk ω middot xk + b( 1113857 + 1113944

N

k1αk (13)

where αk is a Lagrangian multiplierαk ge 0 (k 1 2 3 n) e original problem is expressedas

maxα

minωb

L(ω b α) (14)

αk is a Lagrangian multiplier whereαk ge 0 (k 1 2 3 n) Solving ω and b as a minimumproblem we can get the value of ω and b

ω 1113944N

k1αkykxk

1113944

N

k1αkyk 0

⎧⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎩

(15)

Substituting the obtained solution into the Lagrangianfunction for a minimize problem the following optimizationfunction can be obtained after substituting

stminα

12

1113944

N

k11113944

N

j1αiαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 0αk ge 0 k 1 2 N (16)

en introduce a slack variable ζk for (xk yk) for someabnormal sample points make the training set linearly

inseparable e penalty parameter Cge 0 e original issueis described as

Acc

eler

atio

n (g

)45

0

ndash450 100 200

Times

ProfessionalAmateur

(a)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(b)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(c)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(d)

Acc

eler

atio

n (g

)45

0

ndash450 100 200

Times

ProfessionalAmateur

(e)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(f )

Figure 8 Comparison of elitersquos and amateurrsquos attitude angular (a) is the contrast of pitch (b) is the contrast of roll (c) is the contrast ofyaw when passing (d) is the contrast of pitch (e) is the contrast of roll and (f ) is the contrast of yaw when shooting

Journal of Healthcare Engineering 9

Table 1 Basic and morphology features

Number Symbol Description1 Ray Root mean square (RMS) of y-axis acceleration2 Rwx RMS of x-axis angular velocity3 Rwy RMS of y-axis angular velocity4 Rwz RMS of z-axis angular velocity5 Dax Variance of x-axis acceleration6 Day Variance of y-axis acceleration7 Dwy Variance of y-axis angular velocity8 Maxxy Maximum of x-axis acceleration9 Maxay Maximum of y-axis acceleration10 Minay Minimum of y-axis acceleration11 Minwz Minimum of y-axis angular velocity12 Sax Skewness of x-axis acceleration13 Say Skewness of y-axis acceleration14 Saz Skewness of z-axis acceleration15 Swx Skewness of x-axis angular velocity16 Swy Skewness of y-axis angular velocity17 Swz Skewness of z-axis angular velocity18 IQRax Interquartile range of x-axis acceleration19 IQRay Interquartile range of y-axis acceleration20 IQRaz Interquartile range of z-axis acceleration21 IQRwx Interquartile range of x-axis angular velocity22 IQRwy Interquartile range of y-axis angular velocity23 IQRwz Interquartile range of z-axis angular velocity24 Stdax Standard deviation of x-axis acceleration25 Stdwy Standard deviation of y-axis angular velocity26 Mp Mean value of pitch27 Mr Mean value of roll28 My Mean value of yaw29 Rp Root mean square of pitch30 Rr Root mean square of roll31 Ry Root mean square of yaw32 Stdp Standard deviation of pitch33 Stdr Standard deviation of raw34 Stdy Standard deviation of yaw

Preprocess Dimension extended feature selection Matrices merge

Shooting matrices Fshoot

F1s

F2s

F11s

Passing matrices

f1ndash1

f2ndash1

f1ndash2

f2ndash2

f1ndashm1

F1p

F2s F11s

F1p

F2p

F11p

Fpass

f2ndash1f2ndash2

f2ndashm2

F2p

f11ndash1f11ndash2

f11ndashm11

F11p

f2ndashm2

f1ndash1f1ndash2

F1s

f1ndashm1

f11ndash1f11ndash2

f11ndashm11

Subject 1raw data

S1ndash1s S1ndash1p

S2ndash1s S2ndash1p

S11ndash1s S11ndash1p

R1ndash1s R1ndash1p

R2ndash1s R2ndash1p

R11ndash1s R11ndash1p

Subject 2raw data

Subject 11raw data

Angularmodel

6-dimetions

9-dimetions

Featureselect

9-dimetions

m-dimetions

Figure 9 With preprocessing we obtain effective data Snminusis for shooting and Snminusip for passing Add attitude angle for a suitable data formRnminusis for shooting and Rnminusip for passing Select key features Fis and Fip that contribute to maximizing classifier success rates Combine allsamples to obtain Fs and Fp

10 Journal of Healthcare Engineering

stminωbζ

12ω

2+ C 1113944

N

k1ζkyk ω middot xk + b( 1113857ge 1 minus ζk k 1 2 Nζk ge 0 k 1 2 N (17)

e ultimate function is obtained after conversion

stminα

12

1113944

N

k11113944

N

j1αkαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 00le αk leC k 1 2 N (18)

ere are two labels passing and shooting representedby 0 and 1 respectively in the action classification exper-iment We obtained 264 sets for passing and 250 sets forshooting e datasets are randomly divided into trainingand test sets 80 of them used for training and the rest usedas test data Randomly selected parameters are shown inTable 2e selectable parameter C ranges from 1 to 50000 crange is from 000001 to 005 and the kernel function in-cludes Linear RBF Sigmoid and Polynomial In order tosolve overfitting 3-fold cross is adopted Experimental re-sults show that C 1 c is 00001 and linear kernel functionis used in the optimal model

We use the proposed SVM-based classifier to recognizethe shooting and passing Decision Tree [30] and K-nearestneighbor algorithm (KNN) [31] are the competitors in ourexperiments e accuracy of different models for soccermotion recognition can be found in Table 3 Apparently ourproposed SVM-based algorithm has better performancethan other competitorse one with the attitude angle has abetter performance than the others

Table 4 shows the recognition and assessment results ofour proposed systeme recognition accuracies for passingand shooting are 857 and 885 which illustrated a cleardistinction between passing and shooting e accuracy ofclassifying different behaviors can be 871 on average isresult suggests our system works

44TalentSelectionSystem Similar to the above analysis thelabels are elites and amateurs in the talent selection ex-periment e selectable parameters are still randomly se-lected from Table 2 re-fold cross-validation is used againto prevent overfitting Finally the best performing classifiersare selected Experimental results show that C 1 and linearkernel function applies to all optimal models As for cdifferent from motion recognition the value is 000005

As shown in Table 5 the accuracies of shooting levelclassifiers are higher than others ere is a clear differencebetween different level participants when shooting eresults show that the SVMwith the attitude angle model alsohas satisfactory accuracy

e results in Table 6 are all inferior to others whichmeans there might be little difference between participantswhen passing As listed in Tables 3 5 and 6 comparing withKNN and Decision Tree SVM demonstrated superiority indealing with the linear inseparable problem and the onewith attitude angle is better than the other

As shown in Table 7 the assessment results of proposedsystem Our model with attitude angle features obviouslyperforms better than typical model e professionals andamateurs are 887 and 851 respectively for passing For theshooting action the accuracy of the model is 93 which alsosignifies that for the shooting action the skill gap between theprofessional and the amateur is bigger than the passing action

Table 2 Parameters setting of SVM

Penalty parameter (C) c Kernel1 10minus 5 Linear500 5times 10minus 5 Polynomial103 10minus 4 RBF5times 103 5times 10minus 4 Sigmoid104 10minus 3

5times 103 5times 10minus 3

Table 3 Soccer motion recognition accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 10minus4 09SVM C 1 c 5times10minus4 088KNN N_Neighbors 4 085Decision Tree Min_Samples Split 3 max depth 6 086

Journal of Healthcare Engineering 11

5 Conclusion and Discussion

An AIoT system to recognize different soccer (football)motions and assess the skill levels of soccer players wasproposed in this paper e proposed IoT system consists ofwearable devices a mobile device (eg a mobile phone ortablet) and a cloud-based data processing platform In thisproposed system aWWSD ofMEMSmotion sensors is usedto collect raw data from soccer players By using BTLEtechnology the data is transmitted to the cloud-based dataprocessing platform which analyzes the data and outputsthe results to a mobile device in real-time A SVM modelclassification algorithm with an ankle-based attitude anglemodel is proposed to recognize different motions and assessdifferent skill levels is intelligent system can be a newparadigm and emerging technology based on wearablesensors attaining the experience-driven to data-driventransition in the field of sports training and talent selectionand can be easily extended to analyze other foot-relatedsports motions and skill levels

In this paper we propose a system for recognizing soccermotions based on IoT devices which can recognize usersrsquosoccer motions and evaluate the quality of the motionshowever this system still needs to be improved in terms ofstorage performance and computing speed in the cloud system

e increasing popularity of IoT applications demands thecomputing power of IoTsystems and edge computing is one ofthe main methods to enhance the computing speed of IoTapplications [32] In future research we will try to combine thisrecognition system with edge computing as well as [33ndash36] inwhich authors used edge computing in combination with IoTdevices to increase the computational speed of the system andthus reduce its response time

Data Availability

e soccer motion dataset used to support the findings ofthis study are currently under embargo while the researchfindings are commercialized Requests for data 12 monthsafter publication of this article will be considered by thecorresponding author

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the Natural ScienceFoundation of China under grant 61702336 the Hubei

Table 4 e motion recognition result

Criterion Motion Classification AveragePassing ShootingAccuracy 0857 0885 0871Recall 096 092 094F1-score 0906 0902 0904

Table 5 Shooting level assessment accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 093SVM C 1 c 10minus4 091KNN N_Neighbors 4 087Decision tree Min_Samples Split 3 max depth 6 086

Table 6 Passing level classification accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 087SVM C 1 c 10minus4 086KNN N_Neighbors 4 084Decision Tree Min_Samples Split 3 max depth 6 079

Table 7 e assessment result for passing and shooting

CriterionSkill level of passing

AverageSkill level of shooting

AverageElites Amateurs Elites Amateurs

Accuracy 0887 0851 0869 0959 0902 0931Recall 094 08 087 094 092 093F1-score 0913 0825 0869 0949 0911 093

12 Journal of Healthcare Engineering

Provincial Natural Science Foundation of China under grant2020CFB629 the Fundamental Research Funds for theCentral Universities under grants CZT20027 and CSQ21018and the Research Start-Up Funds of South-Central Uni-versity for Nationalities under grant YZZ18006

References

[1] M Bhatia ldquoIot-inspired framework for athlete performanceassessment in smart sport industryrdquo IEEE Internet of ingsJournal vol 8 no 12 pp 9523ndash9530 2020

[2] Q Zhang Y Zhang C Li C Yan Y Duan and H WangldquoSport location-based user clustering with privacy-preserva-tion in wireless iot-driven healthcarerdquo IEEE Access vol 9pp 12906ndash12913 2021

[3] A Rajsp and I Fister ldquoA systematic literature review of in-telligent data analysis methods for smart sport trainingrdquoApplied Sciences vol 10 no 9 p 3013 2020

[4] J Haladjian D Schlabbers S Taheri M arr andB Bruegge ldquoSensor-based detection and classification ofsoccer goalkeeper training exercisesrdquo ACM Transactions onInternet Technology vol 1 no 2 pp 1ndash20 2020

[5] X Bu ldquoHumanmotion gesture recognition algorithm in videobased on convolutional neural features of training imagesrdquoIEEE Access vol 8 pp 160 025ndash160 039 2020

[6] Y Wang M Chen X Wang R H M Chan andW J Li ldquoIotfor next-generation racket sports trainingrdquo IEEE Internet ofings Journal vol 5 no 6 pp 4558ndash4566 2018

[7] Y Wang Y Zhao R H M Chan and W J Li ldquoVolleyballskill assessment using a single wearable micro inertial mea-surement unit at wristrdquo IEEE Access vol 6pp 13 758ndash813 765 2018

[8] N F Ghazali N Shahar N A Rahmad N A J SufriM A Asrsquoari and H F M Latif ldquoCommon sport activityrecognition using inertial sensorrdquo in Proceedings of the IEEE14th International Colloquium on Signal Processing Its Ap-plications (CSPA) pp 67ndash71 Malaysia March 2018

[9] M Ramanathan W-Y Yau and E K Teoh ldquoHuman actionrecognition with video data research and evaluation chal-lengesrdquo IEEE Transactions on Human-Machine Systemsvol 44 no 5 pp 650ndash663 2014

[10] J K Aggarwal and M S Ryoo ldquoHuman activity analysisrdquoACM Computing Surveys vol 43 no 3 pp 1ndash43 2011

[11] M Stikic D Larlus S Ebert and B Schiele ldquoWeakly su-pervised recognition of daily life activities with wearablesensorsrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 33 no 12 pp 2521ndash2537 2011

[12] J Wang Y Chen S Hao X Peng and L Hu ldquoDeep learningfor sensor-based activity recognition a surveyrdquo PatternRecognition Letters vol 119 pp 3ndash11 2019

[13] KWang J He and L Zhang ldquoAttention-based convolutionalneural network for weakly labeled human activitiesrsquo recog-nition with wearable sensorsrdquo IEEE Sensors Journal vol 19no 17 pp 7598ndash7604 2019

[14] C Chen R Jafari and N Kehtarnavaz ldquoImproving humanaction recognition using fusion of depth camera and inertialsensorsrdquo IEEE Transactions on Human-Machine Systemsvol 45 no 1 pp 51ndash61 2015

[15] K Kui Liu C Chen Chen R Jafari and N KehtarnavazldquoFusion of inertial and depth sensor data for robust handgesture recognitionrdquo IEEE Sensors Journal vol 14 no 6pp 1898ndash1903 2014

[16] A Mannini S S Intille M Rosenberger A M Sabatini andW Haskell ldquoActivity recognition using a single accelerometer

placed at the wrist or anklerdquo Medicine amp Science in Sports ampExercise vol 45 no 11 pp 2193ndash2203 2013

[17] A Mannini M Rosenberger W L Haskell A M Sabatiniand S S Intille ldquoActivity recognition in youth using singleaccelerometer placed at wrist or anklerdquoMedicine amp Science inSports amp Exercise vol 49 no 4 pp 801ndash812 2017

[18] M Shoaib S Bosch O Incel H Scholten and P HavingaldquoComplex human activity recognition using smartphone andwrist-worn motion sensorsrdquo Sensors vol 16 no 4 p 4262016

[19] J P Varkey D Pompili and T A Walls ldquoHuman motionrecognition using a wireless sensor-based wearable systemrdquoPersonal And Ubiquitous Computing vol 16 pp 897ndash9102016

[20] J S Akins N R Heebner M Lovalekar and T C SellldquoReliability and validity of instrumented soccer equipmentrdquoJournal of Applied Biomechanics vol 31 no 3 pp 195ndash2012015

[21] A Meamarbashi and S Hossaini ldquoApplication of novel in-ertial technique to compare the kinematics and kinetics of thelegs in the soccer instep kickrdquo Journal of Human Kineticsvol 23 no 1 pp 5ndash13 2010

[22] BMI160 Datasheet Document Bst-Bmi160-Ds000-07 BoschSensortec Reutlingen Germany 2015

[23] G Okeyo L Chen HWang and R Sterritt ldquoDynamic sensordata segmentation for real-time knowledge-driven activityrecognitionrdquo Pervasive and Mobile Computing vol 10pp 155ndash172 2014

[24] S Ioannidou and G Pantazis ldquoHelmert transformationproblem from euler angles method to quaternion algebrardquoISPRS International Journal of Geo-Information vol 9 no 9p 494 2020

[25] N Jouppi C Young N Patil and D Patterson ldquoMotivationfor and evaluation of the first tensor processing unitrdquo IEEEMicro vol 38 no 3 pp 10ndash19 2018

[26] H Ranocha M Sayyari L Dalcin M Parsani andD I Ketcheson ldquoRelaxation Rkutta methods fully discreteexplicit entropy-stable schemes for the compressible euler andNavierndashStokes equationsrdquo SIAM Journal on Scientific Com-puting vol 42 no 2 pp A612ndashA638 2020

[27] A M Andrew An Introduction to Support Vector Machinesand Other Kernel-Based Learning Methods by Nello Chris-tianini and John Shawe-Taylor Cambridge University PressCambridge UK 2000

[28] X Yang and Y L Tian ldquoEigen joints-based action recognitionusing Naıve-Bayes-nearest-neighborrdquo in Proceedings of theIEEE Computer Society Conference on Computer Vision andPattern Recognition Workshops pp 14ndash19 Providence RIUSA June 2012

[29] T V Gestel B Baesens P V Dijcke J A K Suykens andT Alderweireld ldquoLinear and nonlinear credit scoring bycombining logistic regression and support vector machinesrdquoSocial Science Electronic Publishing vol 1 no 4 2005

[30] A Jindal A Dua K Kaur M Singh N Kumar andS Mishra ldquoDecision tree and SVM-based data analytics fortheft detection in smart gridrdquo IEEE Transactions on IndustrialInformatics vol 12 no 3 pp 1005ndash1016 2016

[31] P anh Noi and M Kappas ldquoComparison of random forestk-nearest neighbor and support vector machine classifiers forland cover classification using sentinel-2 imageryrdquo Sensorsvol 18 no 1 p 18 2018

[32] H Li K Ota and M Dong ldquoLearning iot in edge deeplearning for the internet of things with edge computingrdquo IEEEnetwork vol 32 no 1 pp 96ndash101 2018

Journal of Healthcare Engineering 13

[33] W Rafique L Qi I Yaqoob M Imran R U Rasool andW Dou ldquoComplementing iot services through software de-fined networking and edge computing a comprehensivesurveyrdquo IEEE Communications Surveys amp Tutorials vol 22no 3 pp 1761ndash1804 2020

[34] L Yang H Yao J Wang C Jiang A Benslimane and Y LiuldquoMulti-UAV-enabled load-balance mobile-edge computingfor iot networksrdquo IEEE Internet ofings Journal vol 7 no 8pp 6898ndash6908 2020

[35] C Gong F Lin X Gong and Y Lu ldquoIntelligent cooperativeedge computing in internet of thingsrdquo IEEE Internet of ingsJournal vol 7 no 10 pp 9372ndash9382 2020

[36] B C Kavitha R Vallikannu and K S Sankaran ldquoDelay-aware concurrent data management method for iot collab-orative mobile edge computing environmentrdquo Microproces-sors and Microsystems vol 74 Article ID 103021 2020

14 Journal of Healthcare Engineering

where ax0 ay0 az0 represent the initial acceleration θ ψand φ represent the initial yaw roll and pitch respec-tively magx0 and magy0 are calculated by ax0 ay0 az0

specifically magx0 ax0ay0 + ax0az01 minus 2(a2y0 + a2

z0) andmagy0 ax0ay0 + ay0az01 minus 2(a2

x0 + a2z0) [25] e initial

four elements are calculated from the initial angle

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

WxWyWz

(a)

Ang

ular

vel

ocity

(degs

)

0

ndash1200

1200

0 100 200Times

AxAyAz

(b)

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

WxWyWz

(c)

Ang

ular

vel

ocity

(degs

)

0

ndash1200

1200

0 100 200Times

AxAyAz

(d)

Figure 4 Passing and shooting signal collected from an elite subject (a) Acceleration of a passing (b) Angular velocity of a passing(c) Acceleration of a shooting (d) Angular velocity of a shooting

BLE-based IoT sensor

(a)

Target location

Start position 1

Start position 2

(b)

Figure 3e sensor placement and experimental site setting for raw data collection (a)eWSD is attached to a subjectrsquos or playerrsquos rightankle (b) Positions of soccer players to perform passing and shooting

Journal of Healthcare Engineering 5

Acquire Preprocess Partition Modeling Clsssfier

Low-passfilter

segmentation Pendingdata

Labels

Trainingdataset

SVMAlgorithm

Recognition modelor

Assessment modelTest dataset

Angletrajectory

model

Featureextraction

PCA

Raw data

Labels

Figure 5 e process of soccer motion recognition and assessment system

Ang

ular

vel

ocity

(degs

)

0

ndash1200

1200

0 100 200Times

ProfessionalAmateur

(a)

ProfessionalAmateur

Ang

ular

vel

ocity

(degs

)0

ndash1200

1200

0 100 200Times

(b)

ProfessionalAmateur

Ang

ular

vel

ocity

(degs

)

0

ndash1200

1200

0 100 200Times

(c)

ProfessionalAmateur

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

(d)

ProfessionalAmateur

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

(e)

ProfessionalAmateur

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

(f )

Figure 6e elitesrsquo and amateursrsquo passing motion data comparison result (a) (b) and (c) are the contrast of acceleration on the 3-axes and(d) (e) and (f) are the contrast of angular velocity on 3-axes

6 Journal of Healthcare Engineering

p0

p1

p2

p3

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

sinψ2sin

θ2sin

φ2

+ cosψ2cos

θ2cos

φ2

sinψ2cos

θ2cos

φ2

minus cosψ2sin

θ2sin

φ2

cosψ2sin

θ2cos

φ2

+ sinψ2cos

θ2sin

φ2

cosψ2cos

θ2sin

φ2

minus sinψ2sin

θ2cos

φ2

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(2)

e RungendashKutta method [26] which mainly eliminatesthe complicated process of solving differential equationswhen the derivatives and initial value information of theequation are known is used to solve four elements

p0

p1

p2

p3

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

t+t

p0

p1

p2

p3

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

t

+t

2

minusωx middot p1 minus ωy middot p2 minus ωz middot p3

+ωx middot p0 minus ωz middot p2 minus ωy middot p3

+ωy middot p0 minus ωz middot p1 + ωx middot p3

minusωz middot p0 + ωy middot p1 minus ωx middot p2

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(3)

where t means the current time value and t represents theinterval of next data e following uses the quaternion to solvethe attitude anglee coordinate transformationmatrixD of theknownobject coordinate system to the Earth coordinate system is

D

1 0 00 1 00 0 1

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦ + 2 cos

θ2

0 minusn sinθ2

m sinθ2

n sinθ2

0 minusl sinθ2

minusm sinθ2

l sinθ2

0

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

+ 2

minus m2

+ n2

1113872 1113873sin2θ2

lm sin2θ2

nl sin2θ2

lm sin2θ2

minus l2

+ n2

1113872 1113873sin2θ2

mn sin2θ2

nl sin2θ2

mn sin2θ2

minus m2

+ l2

1113872 1113873sin2θ2

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(4)

m l and n mean the direction vector projected onto thegeographic coordinate along the direction of rotation vectore rotation of the object corresponds to the rotationaround the axis e trigonometry of a quaternion iscos θ2 + (l i

rarr+ m j

rarr+ n k

rarr)sin θ2 cos θ2 + u

rarr sin θ2 inwhich i

rarr jrarr

krarr

denote the unit direction vector in thegeographic graticule system

p0 cosθ2

p1 l sinθ2

p2 m sinθ2

p3 n sinθ2

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(5)

e constructed quaternion describes the fixed-pointrotation problem of the substance Correspondinglythe object system is formed by one-time equivalent ro-tation of the Earth system Equation (5) is substituted intoD1

D1

p20 + p

21 minus p

22 minus p

23 2 p1p2 minus p0p3( 1113857 2 p0p2 + p1p3( 1113857

2 p0p3 + p1p2( 1113857 p20 minus p

21 + p

22 minus p

23 2 p2p3 minus p0p1( 1113857

2 p1p3 minus p0p2( 1113857 2 p0p1 + p2p3( 1113857 p20 minus p

21 minus p

22 + p

23

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(6)

Let the unit vector in the coordinate system X be(ex1 ey1 ez1)

T and correspond to (ex2 ey2 ez2)T in Y and

the direction of the projection of the (ex2 ey2 ez2)T is Cn

b Suppose there is a vector R whose magnitude on X is(x y z)T and on Y is (x2 y2 z2)

T then coordinatetransformation is (x y z)T Cn

b(x2 y2 z2)T e attitude

angle transformation matrix is also converted into the co-ordinate transformation matrix form of object coordinatesystem to a geographic coordinate system and Cn

b is obtainedas follows

Z

Y

X

θ

Zprime

Yprime

Xprime

(a)

Z

YφZprime

Yprime

Xprime(X)

(b)

ZY

ψ

ZprimeYprime

X

Xprime

(c)

Figure 7 Attitude angle an exploded view of the process of the attitude angle rotation

Journal of Healthcare Engineering 7

Cnb Rot(zψ)Rot(y θ)Rot(xφ)

cos ψ minussin ψ 0

cos ψ cos ψ 0

0 0 1

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1 0 0

0 cos φ minussin φ

0 sin φ cos φ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

cos θ 0 sin θ

0 1 0

minussin θ 0 cos θ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

cos ψ cos θ minussin ψ cos φ + cos ψ sin θ sin φ sin ψ sin φ + cos ψ sin θ cos φ

sin ψ cos θ cos ψ cos φ + sin ψ sin θ sin φ minuscos ψ sin φ + sin ψ sin θ cos φ

minussin θ cos θ sin φ cos θ cos φ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(7)

Both D1 and Cnb are matrices transforming the object

coordinate system attitude into the geographic coordinatesystem that is

2 p1p3 minus p0p2( 1113857 minussin θ g1

2 p0p1 + p2p3( 1113857 sin φ cos θ g2m

p20 minus p

21 minus p

22 + p

23 cos θ cos φ g3

2 p0p3 + p1p2( 1113857 sin ψ cos θ g4

p20 + p

21 minus p

22 minus p

23 cos ψ cos θ g5

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎩

(8)

e equation for solving the attitude angle by thequaternion is

roll arctan2 p0p3 + p1p2( 1113857

1 minus 2 p20 + p

211113872 1113873

pitch arctan2 p1p3 minus p0p2( 1113857

4 p0p3 + p1p2( 11138572

+ 1 minus 2 p20 + p

211113872 11138731113872 1113873

2

yaw arctan2 p0p1 + p2p3( 1113857

1 minus 2 p22 + p

231113872 1113873

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(9)

e frequency of the sensor repetition rate is 100Hzen the angular velocity value is substituted into the so-lution attitude angle Anminusi Figure 8 shows the comparison ofattitude angle of elites and amateurs

By combining Anminusi with Snminusi as a part of Rnminusi we extractthe basic and morphology features fni of segments R(n)i asshown in Table 1

F(n) (f1 fmn) are merged into one largematrix F

of size n times m e updated data form is shown in Figure 9

42 Principle Component Analysis e cloud server benefitsfrom reducing memory requirements computing load andnecessary bandwidth during model implementation Torepresent identified variables in compact feature variableswe perform PCA processing e model with PCA canachieve higher accuracy than other nonlinear dimensionalityreduction methods [7] PCA is essentially a base transfor-mation that makes transformed data have the largest vari-ance which means the variance between one axis (spindle)and the point is minimized by the rotation of the coordinate

axis and the translation of the coordinate origin Suppose thematrix F with the dimension of n times m which means thatthere is total n samples in m-dimensions F can bedecomposed into UΣVT where U has the same size as F andthe orthogonal matrix V is m times m Σ is a diagonal matrix ofthe same size as V en YrUΣr F(ΣVT)(minus 1)Σr

We extracted a total of 34-dimensional features for avector characterizing where F [F1 F2 F34] AfterPCA the obtained new features can be expressed as Yr

[Yr1 Yr2 Yrm] and m represents the calculateddimension

Yrm F1aij + F2aij + middot middot middot + Fmaij (10)

where aij are eigenvalues of the covariance matrix Equation(10) can be simplified as follows

Yrm F1a1 + F2a2 + middot middot middot + Fmam (11)

43 Support Vector Machine SVM is widely applied in thedomain of machine learning and pattern recognition as atool to solve the classification problem e basic model ofSVM aims to maximize the distance of the closet sampleswhich is called support vector to the hyperplane on theeigenspace [27] In SVM a training data point is regarded asa $p$-dimensional vector the objective is to separate suchpoints using a $(p-1)$-dimensional hyperplane [19] at isthe training points are mapped to points in space to max-imize the width of the gap between the two and severalcategories by SVM Compared to other supervised algo-rithms such as logistics regression (LR) model or NaıveBayes algorithm [28] the SVM algorithm is more suitable todeal with high-dimensional and linear inseparable problemsby freely selecting the parametric model and only usesupport vector as the classification basis of hyperplane whichmeets our expectations as our problem is a small samplelinear inseparable problem

e training dataset is set as T (Xk yk)|X isin Rm yk isin1113864

0 1 nk1 Here are total n samples X represents an

m-dimensional matrix the value of the classification label yk

is 0 or 1 and the current sample is Xk erefore therepresentation of the optimization objective function is

minωb

12ω

2

st yk ω middot xk + b( 1113857 minus 1ge 0 k 1 2 N

(12)

8 Journal of Healthcare Engineering

Here ω is a vector on a hyperplane and the offset b of thesuperflat is along ω from the origin Problem (12) is a convexquadratic programming problem [29] According to theconvex optimization theory transform problem (12) into anunconstrained problem e optimization function can bedenoted as

L(ω b α) 12ω

2minus 1113944

N

k1αkyk ω middot xk + b( 1113857 + 1113944

N

k1αk (13)

where αk is a Lagrangian multiplierαk ge 0 (k 1 2 3 n) e original problem is expressedas

maxα

minωb

L(ω b α) (14)

αk is a Lagrangian multiplier whereαk ge 0 (k 1 2 3 n) Solving ω and b as a minimumproblem we can get the value of ω and b

ω 1113944N

k1αkykxk

1113944

N

k1αkyk 0

⎧⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎩

(15)

Substituting the obtained solution into the Lagrangianfunction for a minimize problem the following optimizationfunction can be obtained after substituting

stminα

12

1113944

N

k11113944

N

j1αiαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 0αk ge 0 k 1 2 N (16)

en introduce a slack variable ζk for (xk yk) for someabnormal sample points make the training set linearly

inseparable e penalty parameter Cge 0 e original issueis described as

Acc

eler

atio

n (g

)45

0

ndash450 100 200

Times

ProfessionalAmateur

(a)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(b)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(c)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(d)

Acc

eler

atio

n (g

)45

0

ndash450 100 200

Times

ProfessionalAmateur

(e)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(f )

Figure 8 Comparison of elitersquos and amateurrsquos attitude angular (a) is the contrast of pitch (b) is the contrast of roll (c) is the contrast ofyaw when passing (d) is the contrast of pitch (e) is the contrast of roll and (f ) is the contrast of yaw when shooting

Journal of Healthcare Engineering 9

Table 1 Basic and morphology features

Number Symbol Description1 Ray Root mean square (RMS) of y-axis acceleration2 Rwx RMS of x-axis angular velocity3 Rwy RMS of y-axis angular velocity4 Rwz RMS of z-axis angular velocity5 Dax Variance of x-axis acceleration6 Day Variance of y-axis acceleration7 Dwy Variance of y-axis angular velocity8 Maxxy Maximum of x-axis acceleration9 Maxay Maximum of y-axis acceleration10 Minay Minimum of y-axis acceleration11 Minwz Minimum of y-axis angular velocity12 Sax Skewness of x-axis acceleration13 Say Skewness of y-axis acceleration14 Saz Skewness of z-axis acceleration15 Swx Skewness of x-axis angular velocity16 Swy Skewness of y-axis angular velocity17 Swz Skewness of z-axis angular velocity18 IQRax Interquartile range of x-axis acceleration19 IQRay Interquartile range of y-axis acceleration20 IQRaz Interquartile range of z-axis acceleration21 IQRwx Interquartile range of x-axis angular velocity22 IQRwy Interquartile range of y-axis angular velocity23 IQRwz Interquartile range of z-axis angular velocity24 Stdax Standard deviation of x-axis acceleration25 Stdwy Standard deviation of y-axis angular velocity26 Mp Mean value of pitch27 Mr Mean value of roll28 My Mean value of yaw29 Rp Root mean square of pitch30 Rr Root mean square of roll31 Ry Root mean square of yaw32 Stdp Standard deviation of pitch33 Stdr Standard deviation of raw34 Stdy Standard deviation of yaw

Preprocess Dimension extended feature selection Matrices merge

Shooting matrices Fshoot

F1s

F2s

F11s

Passing matrices

f1ndash1

f2ndash1

f1ndash2

f2ndash2

f1ndashm1

F1p

F2s F11s

F1p

F2p

F11p

Fpass

f2ndash1f2ndash2

f2ndashm2

F2p

f11ndash1f11ndash2

f11ndashm11

F11p

f2ndashm2

f1ndash1f1ndash2

F1s

f1ndashm1

f11ndash1f11ndash2

f11ndashm11

Subject 1raw data

S1ndash1s S1ndash1p

S2ndash1s S2ndash1p

S11ndash1s S11ndash1p

R1ndash1s R1ndash1p

R2ndash1s R2ndash1p

R11ndash1s R11ndash1p

Subject 2raw data

Subject 11raw data

Angularmodel

6-dimetions

9-dimetions

Featureselect

9-dimetions

m-dimetions

Figure 9 With preprocessing we obtain effective data Snminusis for shooting and Snminusip for passing Add attitude angle for a suitable data formRnminusis for shooting and Rnminusip for passing Select key features Fis and Fip that contribute to maximizing classifier success rates Combine allsamples to obtain Fs and Fp

10 Journal of Healthcare Engineering

stminωbζ

12ω

2+ C 1113944

N

k1ζkyk ω middot xk + b( 1113857ge 1 minus ζk k 1 2 Nζk ge 0 k 1 2 N (17)

e ultimate function is obtained after conversion

stminα

12

1113944

N

k11113944

N

j1αkαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 00le αk leC k 1 2 N (18)

ere are two labels passing and shooting representedby 0 and 1 respectively in the action classification exper-iment We obtained 264 sets for passing and 250 sets forshooting e datasets are randomly divided into trainingand test sets 80 of them used for training and the rest usedas test data Randomly selected parameters are shown inTable 2e selectable parameter C ranges from 1 to 50000 crange is from 000001 to 005 and the kernel function in-cludes Linear RBF Sigmoid and Polynomial In order tosolve overfitting 3-fold cross is adopted Experimental re-sults show that C 1 c is 00001 and linear kernel functionis used in the optimal model

We use the proposed SVM-based classifier to recognizethe shooting and passing Decision Tree [30] and K-nearestneighbor algorithm (KNN) [31] are the competitors in ourexperiments e accuracy of different models for soccermotion recognition can be found in Table 3 Apparently ourproposed SVM-based algorithm has better performancethan other competitorse one with the attitude angle has abetter performance than the others

Table 4 shows the recognition and assessment results ofour proposed systeme recognition accuracies for passingand shooting are 857 and 885 which illustrated a cleardistinction between passing and shooting e accuracy ofclassifying different behaviors can be 871 on average isresult suggests our system works

44TalentSelectionSystem Similar to the above analysis thelabels are elites and amateurs in the talent selection ex-periment e selectable parameters are still randomly se-lected from Table 2 re-fold cross-validation is used againto prevent overfitting Finally the best performing classifiersare selected Experimental results show that C 1 and linearkernel function applies to all optimal models As for cdifferent from motion recognition the value is 000005

As shown in Table 5 the accuracies of shooting levelclassifiers are higher than others ere is a clear differencebetween different level participants when shooting eresults show that the SVMwith the attitude angle model alsohas satisfactory accuracy

e results in Table 6 are all inferior to others whichmeans there might be little difference between participantswhen passing As listed in Tables 3 5 and 6 comparing withKNN and Decision Tree SVM demonstrated superiority indealing with the linear inseparable problem and the onewith attitude angle is better than the other

As shown in Table 7 the assessment results of proposedsystem Our model with attitude angle features obviouslyperforms better than typical model e professionals andamateurs are 887 and 851 respectively for passing For theshooting action the accuracy of the model is 93 which alsosignifies that for the shooting action the skill gap between theprofessional and the amateur is bigger than the passing action

Table 2 Parameters setting of SVM

Penalty parameter (C) c Kernel1 10minus 5 Linear500 5times 10minus 5 Polynomial103 10minus 4 RBF5times 103 5times 10minus 4 Sigmoid104 10minus 3

5times 103 5times 10minus 3

Table 3 Soccer motion recognition accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 10minus4 09SVM C 1 c 5times10minus4 088KNN N_Neighbors 4 085Decision Tree Min_Samples Split 3 max depth 6 086

Journal of Healthcare Engineering 11

5 Conclusion and Discussion

An AIoT system to recognize different soccer (football)motions and assess the skill levels of soccer players wasproposed in this paper e proposed IoT system consists ofwearable devices a mobile device (eg a mobile phone ortablet) and a cloud-based data processing platform In thisproposed system aWWSD ofMEMSmotion sensors is usedto collect raw data from soccer players By using BTLEtechnology the data is transmitted to the cloud-based dataprocessing platform which analyzes the data and outputsthe results to a mobile device in real-time A SVM modelclassification algorithm with an ankle-based attitude anglemodel is proposed to recognize different motions and assessdifferent skill levels is intelligent system can be a newparadigm and emerging technology based on wearablesensors attaining the experience-driven to data-driventransition in the field of sports training and talent selectionand can be easily extended to analyze other foot-relatedsports motions and skill levels

In this paper we propose a system for recognizing soccermotions based on IoT devices which can recognize usersrsquosoccer motions and evaluate the quality of the motionshowever this system still needs to be improved in terms ofstorage performance and computing speed in the cloud system

e increasing popularity of IoT applications demands thecomputing power of IoTsystems and edge computing is one ofthe main methods to enhance the computing speed of IoTapplications [32] In future research we will try to combine thisrecognition system with edge computing as well as [33ndash36] inwhich authors used edge computing in combination with IoTdevices to increase the computational speed of the system andthus reduce its response time

Data Availability

e soccer motion dataset used to support the findings ofthis study are currently under embargo while the researchfindings are commercialized Requests for data 12 monthsafter publication of this article will be considered by thecorresponding author

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the Natural ScienceFoundation of China under grant 61702336 the Hubei

Table 4 e motion recognition result

Criterion Motion Classification AveragePassing ShootingAccuracy 0857 0885 0871Recall 096 092 094F1-score 0906 0902 0904

Table 5 Shooting level assessment accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 093SVM C 1 c 10minus4 091KNN N_Neighbors 4 087Decision tree Min_Samples Split 3 max depth 6 086

Table 6 Passing level classification accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 087SVM C 1 c 10minus4 086KNN N_Neighbors 4 084Decision Tree Min_Samples Split 3 max depth 6 079

Table 7 e assessment result for passing and shooting

CriterionSkill level of passing

AverageSkill level of shooting

AverageElites Amateurs Elites Amateurs

Accuracy 0887 0851 0869 0959 0902 0931Recall 094 08 087 094 092 093F1-score 0913 0825 0869 0949 0911 093

12 Journal of Healthcare Engineering

Provincial Natural Science Foundation of China under grant2020CFB629 the Fundamental Research Funds for theCentral Universities under grants CZT20027 and CSQ21018and the Research Start-Up Funds of South-Central Uni-versity for Nationalities under grant YZZ18006

References

[1] M Bhatia ldquoIot-inspired framework for athlete performanceassessment in smart sport industryrdquo IEEE Internet of ingsJournal vol 8 no 12 pp 9523ndash9530 2020

[2] Q Zhang Y Zhang C Li C Yan Y Duan and H WangldquoSport location-based user clustering with privacy-preserva-tion in wireless iot-driven healthcarerdquo IEEE Access vol 9pp 12906ndash12913 2021

[3] A Rajsp and I Fister ldquoA systematic literature review of in-telligent data analysis methods for smart sport trainingrdquoApplied Sciences vol 10 no 9 p 3013 2020

[4] J Haladjian D Schlabbers S Taheri M arr andB Bruegge ldquoSensor-based detection and classification ofsoccer goalkeeper training exercisesrdquo ACM Transactions onInternet Technology vol 1 no 2 pp 1ndash20 2020

[5] X Bu ldquoHumanmotion gesture recognition algorithm in videobased on convolutional neural features of training imagesrdquoIEEE Access vol 8 pp 160 025ndash160 039 2020

[6] Y Wang M Chen X Wang R H M Chan andW J Li ldquoIotfor next-generation racket sports trainingrdquo IEEE Internet ofings Journal vol 5 no 6 pp 4558ndash4566 2018

[7] Y Wang Y Zhao R H M Chan and W J Li ldquoVolleyballskill assessment using a single wearable micro inertial mea-surement unit at wristrdquo IEEE Access vol 6pp 13 758ndash813 765 2018

[8] N F Ghazali N Shahar N A Rahmad N A J SufriM A Asrsquoari and H F M Latif ldquoCommon sport activityrecognition using inertial sensorrdquo in Proceedings of the IEEE14th International Colloquium on Signal Processing Its Ap-plications (CSPA) pp 67ndash71 Malaysia March 2018

[9] M Ramanathan W-Y Yau and E K Teoh ldquoHuman actionrecognition with video data research and evaluation chal-lengesrdquo IEEE Transactions on Human-Machine Systemsvol 44 no 5 pp 650ndash663 2014

[10] J K Aggarwal and M S Ryoo ldquoHuman activity analysisrdquoACM Computing Surveys vol 43 no 3 pp 1ndash43 2011

[11] M Stikic D Larlus S Ebert and B Schiele ldquoWeakly su-pervised recognition of daily life activities with wearablesensorsrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 33 no 12 pp 2521ndash2537 2011

[12] J Wang Y Chen S Hao X Peng and L Hu ldquoDeep learningfor sensor-based activity recognition a surveyrdquo PatternRecognition Letters vol 119 pp 3ndash11 2019

[13] KWang J He and L Zhang ldquoAttention-based convolutionalneural network for weakly labeled human activitiesrsquo recog-nition with wearable sensorsrdquo IEEE Sensors Journal vol 19no 17 pp 7598ndash7604 2019

[14] C Chen R Jafari and N Kehtarnavaz ldquoImproving humanaction recognition using fusion of depth camera and inertialsensorsrdquo IEEE Transactions on Human-Machine Systemsvol 45 no 1 pp 51ndash61 2015

[15] K Kui Liu C Chen Chen R Jafari and N KehtarnavazldquoFusion of inertial and depth sensor data for robust handgesture recognitionrdquo IEEE Sensors Journal vol 14 no 6pp 1898ndash1903 2014

[16] A Mannini S S Intille M Rosenberger A M Sabatini andW Haskell ldquoActivity recognition using a single accelerometer

placed at the wrist or anklerdquo Medicine amp Science in Sports ampExercise vol 45 no 11 pp 2193ndash2203 2013

[17] A Mannini M Rosenberger W L Haskell A M Sabatiniand S S Intille ldquoActivity recognition in youth using singleaccelerometer placed at wrist or anklerdquoMedicine amp Science inSports amp Exercise vol 49 no 4 pp 801ndash812 2017

[18] M Shoaib S Bosch O Incel H Scholten and P HavingaldquoComplex human activity recognition using smartphone andwrist-worn motion sensorsrdquo Sensors vol 16 no 4 p 4262016

[19] J P Varkey D Pompili and T A Walls ldquoHuman motionrecognition using a wireless sensor-based wearable systemrdquoPersonal And Ubiquitous Computing vol 16 pp 897ndash9102016

[20] J S Akins N R Heebner M Lovalekar and T C SellldquoReliability and validity of instrumented soccer equipmentrdquoJournal of Applied Biomechanics vol 31 no 3 pp 195ndash2012015

[21] A Meamarbashi and S Hossaini ldquoApplication of novel in-ertial technique to compare the kinematics and kinetics of thelegs in the soccer instep kickrdquo Journal of Human Kineticsvol 23 no 1 pp 5ndash13 2010

[22] BMI160 Datasheet Document Bst-Bmi160-Ds000-07 BoschSensortec Reutlingen Germany 2015

[23] G Okeyo L Chen HWang and R Sterritt ldquoDynamic sensordata segmentation for real-time knowledge-driven activityrecognitionrdquo Pervasive and Mobile Computing vol 10pp 155ndash172 2014

[24] S Ioannidou and G Pantazis ldquoHelmert transformationproblem from euler angles method to quaternion algebrardquoISPRS International Journal of Geo-Information vol 9 no 9p 494 2020

[25] N Jouppi C Young N Patil and D Patterson ldquoMotivationfor and evaluation of the first tensor processing unitrdquo IEEEMicro vol 38 no 3 pp 10ndash19 2018

[26] H Ranocha M Sayyari L Dalcin M Parsani andD I Ketcheson ldquoRelaxation Rkutta methods fully discreteexplicit entropy-stable schemes for the compressible euler andNavierndashStokes equationsrdquo SIAM Journal on Scientific Com-puting vol 42 no 2 pp A612ndashA638 2020

[27] A M Andrew An Introduction to Support Vector Machinesand Other Kernel-Based Learning Methods by Nello Chris-tianini and John Shawe-Taylor Cambridge University PressCambridge UK 2000

[28] X Yang and Y L Tian ldquoEigen joints-based action recognitionusing Naıve-Bayes-nearest-neighborrdquo in Proceedings of theIEEE Computer Society Conference on Computer Vision andPattern Recognition Workshops pp 14ndash19 Providence RIUSA June 2012

[29] T V Gestel B Baesens P V Dijcke J A K Suykens andT Alderweireld ldquoLinear and nonlinear credit scoring bycombining logistic regression and support vector machinesrdquoSocial Science Electronic Publishing vol 1 no 4 2005

[30] A Jindal A Dua K Kaur M Singh N Kumar andS Mishra ldquoDecision tree and SVM-based data analytics fortheft detection in smart gridrdquo IEEE Transactions on IndustrialInformatics vol 12 no 3 pp 1005ndash1016 2016

[31] P anh Noi and M Kappas ldquoComparison of random forestk-nearest neighbor and support vector machine classifiers forland cover classification using sentinel-2 imageryrdquo Sensorsvol 18 no 1 p 18 2018

[32] H Li K Ota and M Dong ldquoLearning iot in edge deeplearning for the internet of things with edge computingrdquo IEEEnetwork vol 32 no 1 pp 96ndash101 2018

Journal of Healthcare Engineering 13

[33] W Rafique L Qi I Yaqoob M Imran R U Rasool andW Dou ldquoComplementing iot services through software de-fined networking and edge computing a comprehensivesurveyrdquo IEEE Communications Surveys amp Tutorials vol 22no 3 pp 1761ndash1804 2020

[34] L Yang H Yao J Wang C Jiang A Benslimane and Y LiuldquoMulti-UAV-enabled load-balance mobile-edge computingfor iot networksrdquo IEEE Internet ofings Journal vol 7 no 8pp 6898ndash6908 2020

[35] C Gong F Lin X Gong and Y Lu ldquoIntelligent cooperativeedge computing in internet of thingsrdquo IEEE Internet of ingsJournal vol 7 no 10 pp 9372ndash9382 2020

[36] B C Kavitha R Vallikannu and K S Sankaran ldquoDelay-aware concurrent data management method for iot collab-orative mobile edge computing environmentrdquo Microproces-sors and Microsystems vol 74 Article ID 103021 2020

14 Journal of Healthcare Engineering

Acquire Preprocess Partition Modeling Clsssfier

Low-passfilter

segmentation Pendingdata

Labels

Trainingdataset

SVMAlgorithm

Recognition modelor

Assessment modelTest dataset

Angletrajectory

model

Featureextraction

PCA

Raw data

Labels

Figure 5 e process of soccer motion recognition and assessment system

Ang

ular

vel

ocity

(degs

)

0

ndash1200

1200

0 100 200Times

ProfessionalAmateur

(a)

ProfessionalAmateur

Ang

ular

vel

ocity

(degs

)0

ndash1200

1200

0 100 200Times

(b)

ProfessionalAmateur

Ang

ular

vel

ocity

(degs

)

0

ndash1200

1200

0 100 200Times

(c)

ProfessionalAmateur

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

(d)

ProfessionalAmateur

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

(e)

ProfessionalAmateur

Acc

eler

atio

n (g

)

15

0

ndash10

0 100 200Times

(f )

Figure 6e elitesrsquo and amateursrsquo passing motion data comparison result (a) (b) and (c) are the contrast of acceleration on the 3-axes and(d) (e) and (f) are the contrast of angular velocity on 3-axes

6 Journal of Healthcare Engineering

p0

p1

p2

p3

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

sinψ2sin

θ2sin

φ2

+ cosψ2cos

θ2cos

φ2

sinψ2cos

θ2cos

φ2

minus cosψ2sin

θ2sin

φ2

cosψ2sin

θ2cos

φ2

+ sinψ2cos

θ2sin

φ2

cosψ2cos

θ2sin

φ2

minus sinψ2sin

θ2cos

φ2

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(2)

e RungendashKutta method [26] which mainly eliminatesthe complicated process of solving differential equationswhen the derivatives and initial value information of theequation are known is used to solve four elements

p0

p1

p2

p3

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

t+t

p0

p1

p2

p3

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

t

+t

2

minusωx middot p1 minus ωy middot p2 minus ωz middot p3

+ωx middot p0 minus ωz middot p2 minus ωy middot p3

+ωy middot p0 minus ωz middot p1 + ωx middot p3

minusωz middot p0 + ωy middot p1 minus ωx middot p2

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(3)

where t means the current time value and t represents theinterval of next data e following uses the quaternion to solvethe attitude anglee coordinate transformationmatrixD of theknownobject coordinate system to the Earth coordinate system is

D

1 0 00 1 00 0 1

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦ + 2 cos

θ2

0 minusn sinθ2

m sinθ2

n sinθ2

0 minusl sinθ2

minusm sinθ2

l sinθ2

0

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

+ 2

minus m2

+ n2

1113872 1113873sin2θ2

lm sin2θ2

nl sin2θ2

lm sin2θ2

minus l2

+ n2

1113872 1113873sin2θ2

mn sin2θ2

nl sin2θ2

mn sin2θ2

minus m2

+ l2

1113872 1113873sin2θ2

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(4)

m l and n mean the direction vector projected onto thegeographic coordinate along the direction of rotation vectore rotation of the object corresponds to the rotationaround the axis e trigonometry of a quaternion iscos θ2 + (l i

rarr+ m j

rarr+ n k

rarr)sin θ2 cos θ2 + u

rarr sin θ2 inwhich i

rarr jrarr

krarr

denote the unit direction vector in thegeographic graticule system

p0 cosθ2

p1 l sinθ2

p2 m sinθ2

p3 n sinθ2

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(5)

e constructed quaternion describes the fixed-pointrotation problem of the substance Correspondinglythe object system is formed by one-time equivalent ro-tation of the Earth system Equation (5) is substituted intoD1

D1

p20 + p

21 minus p

22 minus p

23 2 p1p2 minus p0p3( 1113857 2 p0p2 + p1p3( 1113857

2 p0p3 + p1p2( 1113857 p20 minus p

21 + p

22 minus p

23 2 p2p3 minus p0p1( 1113857

2 p1p3 minus p0p2( 1113857 2 p0p1 + p2p3( 1113857 p20 minus p

21 minus p

22 + p

23

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(6)

Let the unit vector in the coordinate system X be(ex1 ey1 ez1)

T and correspond to (ex2 ey2 ez2)T in Y and

the direction of the projection of the (ex2 ey2 ez2)T is Cn

b Suppose there is a vector R whose magnitude on X is(x y z)T and on Y is (x2 y2 z2)

T then coordinatetransformation is (x y z)T Cn

b(x2 y2 z2)T e attitude

angle transformation matrix is also converted into the co-ordinate transformation matrix form of object coordinatesystem to a geographic coordinate system and Cn

b is obtainedas follows

Z

Y

X

θ

Zprime

Yprime

Xprime

(a)

Z

YφZprime

Yprime

Xprime(X)

(b)

ZY

ψ

ZprimeYprime

X

Xprime

(c)

Figure 7 Attitude angle an exploded view of the process of the attitude angle rotation

Journal of Healthcare Engineering 7

Cnb Rot(zψ)Rot(y θ)Rot(xφ)

cos ψ minussin ψ 0

cos ψ cos ψ 0

0 0 1

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1 0 0

0 cos φ minussin φ

0 sin φ cos φ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

cos θ 0 sin θ

0 1 0

minussin θ 0 cos θ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

cos ψ cos θ minussin ψ cos φ + cos ψ sin θ sin φ sin ψ sin φ + cos ψ sin θ cos φ

sin ψ cos θ cos ψ cos φ + sin ψ sin θ sin φ minuscos ψ sin φ + sin ψ sin θ cos φ

minussin θ cos θ sin φ cos θ cos φ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(7)

Both D1 and Cnb are matrices transforming the object

coordinate system attitude into the geographic coordinatesystem that is

2 p1p3 minus p0p2( 1113857 minussin θ g1

2 p0p1 + p2p3( 1113857 sin φ cos θ g2m

p20 minus p

21 minus p

22 + p

23 cos θ cos φ g3

2 p0p3 + p1p2( 1113857 sin ψ cos θ g4

p20 + p

21 minus p

22 minus p

23 cos ψ cos θ g5

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎩

(8)

e equation for solving the attitude angle by thequaternion is

roll arctan2 p0p3 + p1p2( 1113857

1 minus 2 p20 + p

211113872 1113873

pitch arctan2 p1p3 minus p0p2( 1113857

4 p0p3 + p1p2( 11138572

+ 1 minus 2 p20 + p

211113872 11138731113872 1113873

2

yaw arctan2 p0p1 + p2p3( 1113857

1 minus 2 p22 + p

231113872 1113873

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(9)

e frequency of the sensor repetition rate is 100Hzen the angular velocity value is substituted into the so-lution attitude angle Anminusi Figure 8 shows the comparison ofattitude angle of elites and amateurs

By combining Anminusi with Snminusi as a part of Rnminusi we extractthe basic and morphology features fni of segments R(n)i asshown in Table 1

F(n) (f1 fmn) are merged into one largematrix F

of size n times m e updated data form is shown in Figure 9

42 Principle Component Analysis e cloud server benefitsfrom reducing memory requirements computing load andnecessary bandwidth during model implementation Torepresent identified variables in compact feature variableswe perform PCA processing e model with PCA canachieve higher accuracy than other nonlinear dimensionalityreduction methods [7] PCA is essentially a base transfor-mation that makes transformed data have the largest vari-ance which means the variance between one axis (spindle)and the point is minimized by the rotation of the coordinate

axis and the translation of the coordinate origin Suppose thematrix F with the dimension of n times m which means thatthere is total n samples in m-dimensions F can bedecomposed into UΣVT where U has the same size as F andthe orthogonal matrix V is m times m Σ is a diagonal matrix ofthe same size as V en YrUΣr F(ΣVT)(minus 1)Σr

We extracted a total of 34-dimensional features for avector characterizing where F [F1 F2 F34] AfterPCA the obtained new features can be expressed as Yr

[Yr1 Yr2 Yrm] and m represents the calculateddimension

Yrm F1aij + F2aij + middot middot middot + Fmaij (10)

where aij are eigenvalues of the covariance matrix Equation(10) can be simplified as follows

Yrm F1a1 + F2a2 + middot middot middot + Fmam (11)

43 Support Vector Machine SVM is widely applied in thedomain of machine learning and pattern recognition as atool to solve the classification problem e basic model ofSVM aims to maximize the distance of the closet sampleswhich is called support vector to the hyperplane on theeigenspace [27] In SVM a training data point is regarded asa $p$-dimensional vector the objective is to separate suchpoints using a $(p-1)$-dimensional hyperplane [19] at isthe training points are mapped to points in space to max-imize the width of the gap between the two and severalcategories by SVM Compared to other supervised algo-rithms such as logistics regression (LR) model or NaıveBayes algorithm [28] the SVM algorithm is more suitable todeal with high-dimensional and linear inseparable problemsby freely selecting the parametric model and only usesupport vector as the classification basis of hyperplane whichmeets our expectations as our problem is a small samplelinear inseparable problem

e training dataset is set as T (Xk yk)|X isin Rm yk isin1113864

0 1 nk1 Here are total n samples X represents an

m-dimensional matrix the value of the classification label yk

is 0 or 1 and the current sample is Xk erefore therepresentation of the optimization objective function is

minωb

12ω

2

st yk ω middot xk + b( 1113857 minus 1ge 0 k 1 2 N

(12)

8 Journal of Healthcare Engineering

Here ω is a vector on a hyperplane and the offset b of thesuperflat is along ω from the origin Problem (12) is a convexquadratic programming problem [29] According to theconvex optimization theory transform problem (12) into anunconstrained problem e optimization function can bedenoted as

L(ω b α) 12ω

2minus 1113944

N

k1αkyk ω middot xk + b( 1113857 + 1113944

N

k1αk (13)

where αk is a Lagrangian multiplierαk ge 0 (k 1 2 3 n) e original problem is expressedas

maxα

minωb

L(ω b α) (14)

αk is a Lagrangian multiplier whereαk ge 0 (k 1 2 3 n) Solving ω and b as a minimumproblem we can get the value of ω and b

ω 1113944N

k1αkykxk

1113944

N

k1αkyk 0

⎧⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎩

(15)

Substituting the obtained solution into the Lagrangianfunction for a minimize problem the following optimizationfunction can be obtained after substituting

stminα

12

1113944

N

k11113944

N

j1αiαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 0αk ge 0 k 1 2 N (16)

en introduce a slack variable ζk for (xk yk) for someabnormal sample points make the training set linearly

inseparable e penalty parameter Cge 0 e original issueis described as

Acc

eler

atio

n (g

)45

0

ndash450 100 200

Times

ProfessionalAmateur

(a)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(b)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(c)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(d)

Acc

eler

atio

n (g

)45

0

ndash450 100 200

Times

ProfessionalAmateur

(e)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(f )

Figure 8 Comparison of elitersquos and amateurrsquos attitude angular (a) is the contrast of pitch (b) is the contrast of roll (c) is the contrast ofyaw when passing (d) is the contrast of pitch (e) is the contrast of roll and (f ) is the contrast of yaw when shooting

Journal of Healthcare Engineering 9

Table 1 Basic and morphology features

Number Symbol Description1 Ray Root mean square (RMS) of y-axis acceleration2 Rwx RMS of x-axis angular velocity3 Rwy RMS of y-axis angular velocity4 Rwz RMS of z-axis angular velocity5 Dax Variance of x-axis acceleration6 Day Variance of y-axis acceleration7 Dwy Variance of y-axis angular velocity8 Maxxy Maximum of x-axis acceleration9 Maxay Maximum of y-axis acceleration10 Minay Minimum of y-axis acceleration11 Minwz Minimum of y-axis angular velocity12 Sax Skewness of x-axis acceleration13 Say Skewness of y-axis acceleration14 Saz Skewness of z-axis acceleration15 Swx Skewness of x-axis angular velocity16 Swy Skewness of y-axis angular velocity17 Swz Skewness of z-axis angular velocity18 IQRax Interquartile range of x-axis acceleration19 IQRay Interquartile range of y-axis acceleration20 IQRaz Interquartile range of z-axis acceleration21 IQRwx Interquartile range of x-axis angular velocity22 IQRwy Interquartile range of y-axis angular velocity23 IQRwz Interquartile range of z-axis angular velocity24 Stdax Standard deviation of x-axis acceleration25 Stdwy Standard deviation of y-axis angular velocity26 Mp Mean value of pitch27 Mr Mean value of roll28 My Mean value of yaw29 Rp Root mean square of pitch30 Rr Root mean square of roll31 Ry Root mean square of yaw32 Stdp Standard deviation of pitch33 Stdr Standard deviation of raw34 Stdy Standard deviation of yaw

Preprocess Dimension extended feature selection Matrices merge

Shooting matrices Fshoot

F1s

F2s

F11s

Passing matrices

f1ndash1

f2ndash1

f1ndash2

f2ndash2

f1ndashm1

F1p

F2s F11s

F1p

F2p

F11p

Fpass

f2ndash1f2ndash2

f2ndashm2

F2p

f11ndash1f11ndash2

f11ndashm11

F11p

f2ndashm2

f1ndash1f1ndash2

F1s

f1ndashm1

f11ndash1f11ndash2

f11ndashm11

Subject 1raw data

S1ndash1s S1ndash1p

S2ndash1s S2ndash1p

S11ndash1s S11ndash1p

R1ndash1s R1ndash1p

R2ndash1s R2ndash1p

R11ndash1s R11ndash1p

Subject 2raw data

Subject 11raw data

Angularmodel

6-dimetions

9-dimetions

Featureselect

9-dimetions

m-dimetions

Figure 9 With preprocessing we obtain effective data Snminusis for shooting and Snminusip for passing Add attitude angle for a suitable data formRnminusis for shooting and Rnminusip for passing Select key features Fis and Fip that contribute to maximizing classifier success rates Combine allsamples to obtain Fs and Fp

10 Journal of Healthcare Engineering

stminωbζ

12ω

2+ C 1113944

N

k1ζkyk ω middot xk + b( 1113857ge 1 minus ζk k 1 2 Nζk ge 0 k 1 2 N (17)

e ultimate function is obtained after conversion

stminα

12

1113944

N

k11113944

N

j1αkαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 00le αk leC k 1 2 N (18)

ere are two labels passing and shooting representedby 0 and 1 respectively in the action classification exper-iment We obtained 264 sets for passing and 250 sets forshooting e datasets are randomly divided into trainingand test sets 80 of them used for training and the rest usedas test data Randomly selected parameters are shown inTable 2e selectable parameter C ranges from 1 to 50000 crange is from 000001 to 005 and the kernel function in-cludes Linear RBF Sigmoid and Polynomial In order tosolve overfitting 3-fold cross is adopted Experimental re-sults show that C 1 c is 00001 and linear kernel functionis used in the optimal model

We use the proposed SVM-based classifier to recognizethe shooting and passing Decision Tree [30] and K-nearestneighbor algorithm (KNN) [31] are the competitors in ourexperiments e accuracy of different models for soccermotion recognition can be found in Table 3 Apparently ourproposed SVM-based algorithm has better performancethan other competitorse one with the attitude angle has abetter performance than the others

Table 4 shows the recognition and assessment results ofour proposed systeme recognition accuracies for passingand shooting are 857 and 885 which illustrated a cleardistinction between passing and shooting e accuracy ofclassifying different behaviors can be 871 on average isresult suggests our system works

44TalentSelectionSystem Similar to the above analysis thelabels are elites and amateurs in the talent selection ex-periment e selectable parameters are still randomly se-lected from Table 2 re-fold cross-validation is used againto prevent overfitting Finally the best performing classifiersare selected Experimental results show that C 1 and linearkernel function applies to all optimal models As for cdifferent from motion recognition the value is 000005

As shown in Table 5 the accuracies of shooting levelclassifiers are higher than others ere is a clear differencebetween different level participants when shooting eresults show that the SVMwith the attitude angle model alsohas satisfactory accuracy

e results in Table 6 are all inferior to others whichmeans there might be little difference between participantswhen passing As listed in Tables 3 5 and 6 comparing withKNN and Decision Tree SVM demonstrated superiority indealing with the linear inseparable problem and the onewith attitude angle is better than the other

As shown in Table 7 the assessment results of proposedsystem Our model with attitude angle features obviouslyperforms better than typical model e professionals andamateurs are 887 and 851 respectively for passing For theshooting action the accuracy of the model is 93 which alsosignifies that for the shooting action the skill gap between theprofessional and the amateur is bigger than the passing action

Table 2 Parameters setting of SVM

Penalty parameter (C) c Kernel1 10minus 5 Linear500 5times 10minus 5 Polynomial103 10minus 4 RBF5times 103 5times 10minus 4 Sigmoid104 10minus 3

5times 103 5times 10minus 3

Table 3 Soccer motion recognition accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 10minus4 09SVM C 1 c 5times10minus4 088KNN N_Neighbors 4 085Decision Tree Min_Samples Split 3 max depth 6 086

Journal of Healthcare Engineering 11

5 Conclusion and Discussion

An AIoT system to recognize different soccer (football)motions and assess the skill levels of soccer players wasproposed in this paper e proposed IoT system consists ofwearable devices a mobile device (eg a mobile phone ortablet) and a cloud-based data processing platform In thisproposed system aWWSD ofMEMSmotion sensors is usedto collect raw data from soccer players By using BTLEtechnology the data is transmitted to the cloud-based dataprocessing platform which analyzes the data and outputsthe results to a mobile device in real-time A SVM modelclassification algorithm with an ankle-based attitude anglemodel is proposed to recognize different motions and assessdifferent skill levels is intelligent system can be a newparadigm and emerging technology based on wearablesensors attaining the experience-driven to data-driventransition in the field of sports training and talent selectionand can be easily extended to analyze other foot-relatedsports motions and skill levels

In this paper we propose a system for recognizing soccermotions based on IoT devices which can recognize usersrsquosoccer motions and evaluate the quality of the motionshowever this system still needs to be improved in terms ofstorage performance and computing speed in the cloud system

e increasing popularity of IoT applications demands thecomputing power of IoTsystems and edge computing is one ofthe main methods to enhance the computing speed of IoTapplications [32] In future research we will try to combine thisrecognition system with edge computing as well as [33ndash36] inwhich authors used edge computing in combination with IoTdevices to increase the computational speed of the system andthus reduce its response time

Data Availability

e soccer motion dataset used to support the findings ofthis study are currently under embargo while the researchfindings are commercialized Requests for data 12 monthsafter publication of this article will be considered by thecorresponding author

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the Natural ScienceFoundation of China under grant 61702336 the Hubei

Table 4 e motion recognition result

Criterion Motion Classification AveragePassing ShootingAccuracy 0857 0885 0871Recall 096 092 094F1-score 0906 0902 0904

Table 5 Shooting level assessment accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 093SVM C 1 c 10minus4 091KNN N_Neighbors 4 087Decision tree Min_Samples Split 3 max depth 6 086

Table 6 Passing level classification accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 087SVM C 1 c 10minus4 086KNN N_Neighbors 4 084Decision Tree Min_Samples Split 3 max depth 6 079

Table 7 e assessment result for passing and shooting

CriterionSkill level of passing

AverageSkill level of shooting

AverageElites Amateurs Elites Amateurs

Accuracy 0887 0851 0869 0959 0902 0931Recall 094 08 087 094 092 093F1-score 0913 0825 0869 0949 0911 093

12 Journal of Healthcare Engineering

Provincial Natural Science Foundation of China under grant2020CFB629 the Fundamental Research Funds for theCentral Universities under grants CZT20027 and CSQ21018and the Research Start-Up Funds of South-Central Uni-versity for Nationalities under grant YZZ18006

References

[1] M Bhatia ldquoIot-inspired framework for athlete performanceassessment in smart sport industryrdquo IEEE Internet of ingsJournal vol 8 no 12 pp 9523ndash9530 2020

[2] Q Zhang Y Zhang C Li C Yan Y Duan and H WangldquoSport location-based user clustering with privacy-preserva-tion in wireless iot-driven healthcarerdquo IEEE Access vol 9pp 12906ndash12913 2021

[3] A Rajsp and I Fister ldquoA systematic literature review of in-telligent data analysis methods for smart sport trainingrdquoApplied Sciences vol 10 no 9 p 3013 2020

[4] J Haladjian D Schlabbers S Taheri M arr andB Bruegge ldquoSensor-based detection and classification ofsoccer goalkeeper training exercisesrdquo ACM Transactions onInternet Technology vol 1 no 2 pp 1ndash20 2020

[5] X Bu ldquoHumanmotion gesture recognition algorithm in videobased on convolutional neural features of training imagesrdquoIEEE Access vol 8 pp 160 025ndash160 039 2020

[6] Y Wang M Chen X Wang R H M Chan andW J Li ldquoIotfor next-generation racket sports trainingrdquo IEEE Internet ofings Journal vol 5 no 6 pp 4558ndash4566 2018

[7] Y Wang Y Zhao R H M Chan and W J Li ldquoVolleyballskill assessment using a single wearable micro inertial mea-surement unit at wristrdquo IEEE Access vol 6pp 13 758ndash813 765 2018

[8] N F Ghazali N Shahar N A Rahmad N A J SufriM A Asrsquoari and H F M Latif ldquoCommon sport activityrecognition using inertial sensorrdquo in Proceedings of the IEEE14th International Colloquium on Signal Processing Its Ap-plications (CSPA) pp 67ndash71 Malaysia March 2018

[9] M Ramanathan W-Y Yau and E K Teoh ldquoHuman actionrecognition with video data research and evaluation chal-lengesrdquo IEEE Transactions on Human-Machine Systemsvol 44 no 5 pp 650ndash663 2014

[10] J K Aggarwal and M S Ryoo ldquoHuman activity analysisrdquoACM Computing Surveys vol 43 no 3 pp 1ndash43 2011

[11] M Stikic D Larlus S Ebert and B Schiele ldquoWeakly su-pervised recognition of daily life activities with wearablesensorsrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 33 no 12 pp 2521ndash2537 2011

[12] J Wang Y Chen S Hao X Peng and L Hu ldquoDeep learningfor sensor-based activity recognition a surveyrdquo PatternRecognition Letters vol 119 pp 3ndash11 2019

[13] KWang J He and L Zhang ldquoAttention-based convolutionalneural network for weakly labeled human activitiesrsquo recog-nition with wearable sensorsrdquo IEEE Sensors Journal vol 19no 17 pp 7598ndash7604 2019

[14] C Chen R Jafari and N Kehtarnavaz ldquoImproving humanaction recognition using fusion of depth camera and inertialsensorsrdquo IEEE Transactions on Human-Machine Systemsvol 45 no 1 pp 51ndash61 2015

[15] K Kui Liu C Chen Chen R Jafari and N KehtarnavazldquoFusion of inertial and depth sensor data for robust handgesture recognitionrdquo IEEE Sensors Journal vol 14 no 6pp 1898ndash1903 2014

[16] A Mannini S S Intille M Rosenberger A M Sabatini andW Haskell ldquoActivity recognition using a single accelerometer

placed at the wrist or anklerdquo Medicine amp Science in Sports ampExercise vol 45 no 11 pp 2193ndash2203 2013

[17] A Mannini M Rosenberger W L Haskell A M Sabatiniand S S Intille ldquoActivity recognition in youth using singleaccelerometer placed at wrist or anklerdquoMedicine amp Science inSports amp Exercise vol 49 no 4 pp 801ndash812 2017

[18] M Shoaib S Bosch O Incel H Scholten and P HavingaldquoComplex human activity recognition using smartphone andwrist-worn motion sensorsrdquo Sensors vol 16 no 4 p 4262016

[19] J P Varkey D Pompili and T A Walls ldquoHuman motionrecognition using a wireless sensor-based wearable systemrdquoPersonal And Ubiquitous Computing vol 16 pp 897ndash9102016

[20] J S Akins N R Heebner M Lovalekar and T C SellldquoReliability and validity of instrumented soccer equipmentrdquoJournal of Applied Biomechanics vol 31 no 3 pp 195ndash2012015

[21] A Meamarbashi and S Hossaini ldquoApplication of novel in-ertial technique to compare the kinematics and kinetics of thelegs in the soccer instep kickrdquo Journal of Human Kineticsvol 23 no 1 pp 5ndash13 2010

[22] BMI160 Datasheet Document Bst-Bmi160-Ds000-07 BoschSensortec Reutlingen Germany 2015

[23] G Okeyo L Chen HWang and R Sterritt ldquoDynamic sensordata segmentation for real-time knowledge-driven activityrecognitionrdquo Pervasive and Mobile Computing vol 10pp 155ndash172 2014

[24] S Ioannidou and G Pantazis ldquoHelmert transformationproblem from euler angles method to quaternion algebrardquoISPRS International Journal of Geo-Information vol 9 no 9p 494 2020

[25] N Jouppi C Young N Patil and D Patterson ldquoMotivationfor and evaluation of the first tensor processing unitrdquo IEEEMicro vol 38 no 3 pp 10ndash19 2018

[26] H Ranocha M Sayyari L Dalcin M Parsani andD I Ketcheson ldquoRelaxation Rkutta methods fully discreteexplicit entropy-stable schemes for the compressible euler andNavierndashStokes equationsrdquo SIAM Journal on Scientific Com-puting vol 42 no 2 pp A612ndashA638 2020

[27] A M Andrew An Introduction to Support Vector Machinesand Other Kernel-Based Learning Methods by Nello Chris-tianini and John Shawe-Taylor Cambridge University PressCambridge UK 2000

[28] X Yang and Y L Tian ldquoEigen joints-based action recognitionusing Naıve-Bayes-nearest-neighborrdquo in Proceedings of theIEEE Computer Society Conference on Computer Vision andPattern Recognition Workshops pp 14ndash19 Providence RIUSA June 2012

[29] T V Gestel B Baesens P V Dijcke J A K Suykens andT Alderweireld ldquoLinear and nonlinear credit scoring bycombining logistic regression and support vector machinesrdquoSocial Science Electronic Publishing vol 1 no 4 2005

[30] A Jindal A Dua K Kaur M Singh N Kumar andS Mishra ldquoDecision tree and SVM-based data analytics fortheft detection in smart gridrdquo IEEE Transactions on IndustrialInformatics vol 12 no 3 pp 1005ndash1016 2016

[31] P anh Noi and M Kappas ldquoComparison of random forestk-nearest neighbor and support vector machine classifiers forland cover classification using sentinel-2 imageryrdquo Sensorsvol 18 no 1 p 18 2018

[32] H Li K Ota and M Dong ldquoLearning iot in edge deeplearning for the internet of things with edge computingrdquo IEEEnetwork vol 32 no 1 pp 96ndash101 2018

Journal of Healthcare Engineering 13

[33] W Rafique L Qi I Yaqoob M Imran R U Rasool andW Dou ldquoComplementing iot services through software de-fined networking and edge computing a comprehensivesurveyrdquo IEEE Communications Surveys amp Tutorials vol 22no 3 pp 1761ndash1804 2020

[34] L Yang H Yao J Wang C Jiang A Benslimane and Y LiuldquoMulti-UAV-enabled load-balance mobile-edge computingfor iot networksrdquo IEEE Internet ofings Journal vol 7 no 8pp 6898ndash6908 2020

[35] C Gong F Lin X Gong and Y Lu ldquoIntelligent cooperativeedge computing in internet of thingsrdquo IEEE Internet of ingsJournal vol 7 no 10 pp 9372ndash9382 2020

[36] B C Kavitha R Vallikannu and K S Sankaran ldquoDelay-aware concurrent data management method for iot collab-orative mobile edge computing environmentrdquo Microproces-sors and Microsystems vol 74 Article ID 103021 2020

14 Journal of Healthcare Engineering

p0

p1

p2

p3

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

sinψ2sin

θ2sin

φ2

+ cosψ2cos

θ2cos

φ2

sinψ2cos

θ2cos

φ2

minus cosψ2sin

θ2sin

φ2

cosψ2sin

θ2cos

φ2

+ sinψ2cos

θ2sin

φ2

cosψ2cos

θ2sin

φ2

minus sinψ2sin

θ2cos

φ2

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(2)

e RungendashKutta method [26] which mainly eliminatesthe complicated process of solving differential equationswhen the derivatives and initial value information of theequation are known is used to solve four elements

p0

p1

p2

p3

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

t+t

p0

p1

p2

p3

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

t

+t

2

minusωx middot p1 minus ωy middot p2 minus ωz middot p3

+ωx middot p0 minus ωz middot p2 minus ωy middot p3

+ωy middot p0 minus ωz middot p1 + ωx middot p3

minusωz middot p0 + ωy middot p1 minus ωx middot p2

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(3)

where t means the current time value and t represents theinterval of next data e following uses the quaternion to solvethe attitude anglee coordinate transformationmatrixD of theknownobject coordinate system to the Earth coordinate system is

D

1 0 00 1 00 0 1

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦ + 2 cos

θ2

0 minusn sinθ2

m sinθ2

n sinθ2

0 minusl sinθ2

minusm sinθ2

l sinθ2

0

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

+ 2

minus m2

+ n2

1113872 1113873sin2θ2

lm sin2θ2

nl sin2θ2

lm sin2θ2

minus l2

+ n2

1113872 1113873sin2θ2

mn sin2θ2

nl sin2θ2

mn sin2θ2

minus m2

+ l2

1113872 1113873sin2θ2

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(4)

m l and n mean the direction vector projected onto thegeographic coordinate along the direction of rotation vectore rotation of the object corresponds to the rotationaround the axis e trigonometry of a quaternion iscos θ2 + (l i

rarr+ m j

rarr+ n k

rarr)sin θ2 cos θ2 + u

rarr sin θ2 inwhich i

rarr jrarr

krarr

denote the unit direction vector in thegeographic graticule system

p0 cosθ2

p1 l sinθ2

p2 m sinθ2

p3 n sinθ2

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(5)

e constructed quaternion describes the fixed-pointrotation problem of the substance Correspondinglythe object system is formed by one-time equivalent ro-tation of the Earth system Equation (5) is substituted intoD1

D1

p20 + p

21 minus p

22 minus p

23 2 p1p2 minus p0p3( 1113857 2 p0p2 + p1p3( 1113857

2 p0p3 + p1p2( 1113857 p20 minus p

21 + p

22 minus p

23 2 p2p3 minus p0p1( 1113857

2 p1p3 minus p0p2( 1113857 2 p0p1 + p2p3( 1113857 p20 minus p

21 minus p

22 + p

23

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(6)

Let the unit vector in the coordinate system X be(ex1 ey1 ez1)

T and correspond to (ex2 ey2 ez2)T in Y and

the direction of the projection of the (ex2 ey2 ez2)T is Cn

b Suppose there is a vector R whose magnitude on X is(x y z)T and on Y is (x2 y2 z2)

T then coordinatetransformation is (x y z)T Cn

b(x2 y2 z2)T e attitude

angle transformation matrix is also converted into the co-ordinate transformation matrix form of object coordinatesystem to a geographic coordinate system and Cn

b is obtainedas follows

Z

Y

X

θ

Zprime

Yprime

Xprime

(a)

Z

YφZprime

Yprime

Xprime(X)

(b)

ZY

ψ

ZprimeYprime

X

Xprime

(c)

Figure 7 Attitude angle an exploded view of the process of the attitude angle rotation

Journal of Healthcare Engineering 7

Cnb Rot(zψ)Rot(y θ)Rot(xφ)

cos ψ minussin ψ 0

cos ψ cos ψ 0

0 0 1

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1 0 0

0 cos φ minussin φ

0 sin φ cos φ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

cos θ 0 sin θ

0 1 0

minussin θ 0 cos θ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

cos ψ cos θ minussin ψ cos φ + cos ψ sin θ sin φ sin ψ sin φ + cos ψ sin θ cos φ

sin ψ cos θ cos ψ cos φ + sin ψ sin θ sin φ minuscos ψ sin φ + sin ψ sin θ cos φ

minussin θ cos θ sin φ cos θ cos φ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(7)

Both D1 and Cnb are matrices transforming the object

coordinate system attitude into the geographic coordinatesystem that is

2 p1p3 minus p0p2( 1113857 minussin θ g1

2 p0p1 + p2p3( 1113857 sin φ cos θ g2m

p20 minus p

21 minus p

22 + p

23 cos θ cos φ g3

2 p0p3 + p1p2( 1113857 sin ψ cos θ g4

p20 + p

21 minus p

22 minus p

23 cos ψ cos θ g5

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎩

(8)

e equation for solving the attitude angle by thequaternion is

roll arctan2 p0p3 + p1p2( 1113857

1 minus 2 p20 + p

211113872 1113873

pitch arctan2 p1p3 minus p0p2( 1113857

4 p0p3 + p1p2( 11138572

+ 1 minus 2 p20 + p

211113872 11138731113872 1113873

2

yaw arctan2 p0p1 + p2p3( 1113857

1 minus 2 p22 + p

231113872 1113873

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(9)

e frequency of the sensor repetition rate is 100Hzen the angular velocity value is substituted into the so-lution attitude angle Anminusi Figure 8 shows the comparison ofattitude angle of elites and amateurs

By combining Anminusi with Snminusi as a part of Rnminusi we extractthe basic and morphology features fni of segments R(n)i asshown in Table 1

F(n) (f1 fmn) are merged into one largematrix F

of size n times m e updated data form is shown in Figure 9

42 Principle Component Analysis e cloud server benefitsfrom reducing memory requirements computing load andnecessary bandwidth during model implementation Torepresent identified variables in compact feature variableswe perform PCA processing e model with PCA canachieve higher accuracy than other nonlinear dimensionalityreduction methods [7] PCA is essentially a base transfor-mation that makes transformed data have the largest vari-ance which means the variance between one axis (spindle)and the point is minimized by the rotation of the coordinate

axis and the translation of the coordinate origin Suppose thematrix F with the dimension of n times m which means thatthere is total n samples in m-dimensions F can bedecomposed into UΣVT where U has the same size as F andthe orthogonal matrix V is m times m Σ is a diagonal matrix ofthe same size as V en YrUΣr F(ΣVT)(minus 1)Σr

We extracted a total of 34-dimensional features for avector characterizing where F [F1 F2 F34] AfterPCA the obtained new features can be expressed as Yr

[Yr1 Yr2 Yrm] and m represents the calculateddimension

Yrm F1aij + F2aij + middot middot middot + Fmaij (10)

where aij are eigenvalues of the covariance matrix Equation(10) can be simplified as follows

Yrm F1a1 + F2a2 + middot middot middot + Fmam (11)

43 Support Vector Machine SVM is widely applied in thedomain of machine learning and pattern recognition as atool to solve the classification problem e basic model ofSVM aims to maximize the distance of the closet sampleswhich is called support vector to the hyperplane on theeigenspace [27] In SVM a training data point is regarded asa $p$-dimensional vector the objective is to separate suchpoints using a $(p-1)$-dimensional hyperplane [19] at isthe training points are mapped to points in space to max-imize the width of the gap between the two and severalcategories by SVM Compared to other supervised algo-rithms such as logistics regression (LR) model or NaıveBayes algorithm [28] the SVM algorithm is more suitable todeal with high-dimensional and linear inseparable problemsby freely selecting the parametric model and only usesupport vector as the classification basis of hyperplane whichmeets our expectations as our problem is a small samplelinear inseparable problem

e training dataset is set as T (Xk yk)|X isin Rm yk isin1113864

0 1 nk1 Here are total n samples X represents an

m-dimensional matrix the value of the classification label yk

is 0 or 1 and the current sample is Xk erefore therepresentation of the optimization objective function is

minωb

12ω

2

st yk ω middot xk + b( 1113857 minus 1ge 0 k 1 2 N

(12)

8 Journal of Healthcare Engineering

Here ω is a vector on a hyperplane and the offset b of thesuperflat is along ω from the origin Problem (12) is a convexquadratic programming problem [29] According to theconvex optimization theory transform problem (12) into anunconstrained problem e optimization function can bedenoted as

L(ω b α) 12ω

2minus 1113944

N

k1αkyk ω middot xk + b( 1113857 + 1113944

N

k1αk (13)

where αk is a Lagrangian multiplierαk ge 0 (k 1 2 3 n) e original problem is expressedas

maxα

minωb

L(ω b α) (14)

αk is a Lagrangian multiplier whereαk ge 0 (k 1 2 3 n) Solving ω and b as a minimumproblem we can get the value of ω and b

ω 1113944N

k1αkykxk

1113944

N

k1αkyk 0

⎧⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎩

(15)

Substituting the obtained solution into the Lagrangianfunction for a minimize problem the following optimizationfunction can be obtained after substituting

stminα

12

1113944

N

k11113944

N

j1αiαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 0αk ge 0 k 1 2 N (16)

en introduce a slack variable ζk for (xk yk) for someabnormal sample points make the training set linearly

inseparable e penalty parameter Cge 0 e original issueis described as

Acc

eler

atio

n (g

)45

0

ndash450 100 200

Times

ProfessionalAmateur

(a)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(b)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(c)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(d)

Acc

eler

atio

n (g

)45

0

ndash450 100 200

Times

ProfessionalAmateur

(e)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(f )

Figure 8 Comparison of elitersquos and amateurrsquos attitude angular (a) is the contrast of pitch (b) is the contrast of roll (c) is the contrast ofyaw when passing (d) is the contrast of pitch (e) is the contrast of roll and (f ) is the contrast of yaw when shooting

Journal of Healthcare Engineering 9

Table 1 Basic and morphology features

Number Symbol Description1 Ray Root mean square (RMS) of y-axis acceleration2 Rwx RMS of x-axis angular velocity3 Rwy RMS of y-axis angular velocity4 Rwz RMS of z-axis angular velocity5 Dax Variance of x-axis acceleration6 Day Variance of y-axis acceleration7 Dwy Variance of y-axis angular velocity8 Maxxy Maximum of x-axis acceleration9 Maxay Maximum of y-axis acceleration10 Minay Minimum of y-axis acceleration11 Minwz Minimum of y-axis angular velocity12 Sax Skewness of x-axis acceleration13 Say Skewness of y-axis acceleration14 Saz Skewness of z-axis acceleration15 Swx Skewness of x-axis angular velocity16 Swy Skewness of y-axis angular velocity17 Swz Skewness of z-axis angular velocity18 IQRax Interquartile range of x-axis acceleration19 IQRay Interquartile range of y-axis acceleration20 IQRaz Interquartile range of z-axis acceleration21 IQRwx Interquartile range of x-axis angular velocity22 IQRwy Interquartile range of y-axis angular velocity23 IQRwz Interquartile range of z-axis angular velocity24 Stdax Standard deviation of x-axis acceleration25 Stdwy Standard deviation of y-axis angular velocity26 Mp Mean value of pitch27 Mr Mean value of roll28 My Mean value of yaw29 Rp Root mean square of pitch30 Rr Root mean square of roll31 Ry Root mean square of yaw32 Stdp Standard deviation of pitch33 Stdr Standard deviation of raw34 Stdy Standard deviation of yaw

Preprocess Dimension extended feature selection Matrices merge

Shooting matrices Fshoot

F1s

F2s

F11s

Passing matrices

f1ndash1

f2ndash1

f1ndash2

f2ndash2

f1ndashm1

F1p

F2s F11s

F1p

F2p

F11p

Fpass

f2ndash1f2ndash2

f2ndashm2

F2p

f11ndash1f11ndash2

f11ndashm11

F11p

f2ndashm2

f1ndash1f1ndash2

F1s

f1ndashm1

f11ndash1f11ndash2

f11ndashm11

Subject 1raw data

S1ndash1s S1ndash1p

S2ndash1s S2ndash1p

S11ndash1s S11ndash1p

R1ndash1s R1ndash1p

R2ndash1s R2ndash1p

R11ndash1s R11ndash1p

Subject 2raw data

Subject 11raw data

Angularmodel

6-dimetions

9-dimetions

Featureselect

9-dimetions

m-dimetions

Figure 9 With preprocessing we obtain effective data Snminusis for shooting and Snminusip for passing Add attitude angle for a suitable data formRnminusis for shooting and Rnminusip for passing Select key features Fis and Fip that contribute to maximizing classifier success rates Combine allsamples to obtain Fs and Fp

10 Journal of Healthcare Engineering

stminωbζ

12ω

2+ C 1113944

N

k1ζkyk ω middot xk + b( 1113857ge 1 minus ζk k 1 2 Nζk ge 0 k 1 2 N (17)

e ultimate function is obtained after conversion

stminα

12

1113944

N

k11113944

N

j1αkαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 00le αk leC k 1 2 N (18)

ere are two labels passing and shooting representedby 0 and 1 respectively in the action classification exper-iment We obtained 264 sets for passing and 250 sets forshooting e datasets are randomly divided into trainingand test sets 80 of them used for training and the rest usedas test data Randomly selected parameters are shown inTable 2e selectable parameter C ranges from 1 to 50000 crange is from 000001 to 005 and the kernel function in-cludes Linear RBF Sigmoid and Polynomial In order tosolve overfitting 3-fold cross is adopted Experimental re-sults show that C 1 c is 00001 and linear kernel functionis used in the optimal model

We use the proposed SVM-based classifier to recognizethe shooting and passing Decision Tree [30] and K-nearestneighbor algorithm (KNN) [31] are the competitors in ourexperiments e accuracy of different models for soccermotion recognition can be found in Table 3 Apparently ourproposed SVM-based algorithm has better performancethan other competitorse one with the attitude angle has abetter performance than the others

Table 4 shows the recognition and assessment results ofour proposed systeme recognition accuracies for passingand shooting are 857 and 885 which illustrated a cleardistinction between passing and shooting e accuracy ofclassifying different behaviors can be 871 on average isresult suggests our system works

44TalentSelectionSystem Similar to the above analysis thelabels are elites and amateurs in the talent selection ex-periment e selectable parameters are still randomly se-lected from Table 2 re-fold cross-validation is used againto prevent overfitting Finally the best performing classifiersare selected Experimental results show that C 1 and linearkernel function applies to all optimal models As for cdifferent from motion recognition the value is 000005

As shown in Table 5 the accuracies of shooting levelclassifiers are higher than others ere is a clear differencebetween different level participants when shooting eresults show that the SVMwith the attitude angle model alsohas satisfactory accuracy

e results in Table 6 are all inferior to others whichmeans there might be little difference between participantswhen passing As listed in Tables 3 5 and 6 comparing withKNN and Decision Tree SVM demonstrated superiority indealing with the linear inseparable problem and the onewith attitude angle is better than the other

As shown in Table 7 the assessment results of proposedsystem Our model with attitude angle features obviouslyperforms better than typical model e professionals andamateurs are 887 and 851 respectively for passing For theshooting action the accuracy of the model is 93 which alsosignifies that for the shooting action the skill gap between theprofessional and the amateur is bigger than the passing action

Table 2 Parameters setting of SVM

Penalty parameter (C) c Kernel1 10minus 5 Linear500 5times 10minus 5 Polynomial103 10minus 4 RBF5times 103 5times 10minus 4 Sigmoid104 10minus 3

5times 103 5times 10minus 3

Table 3 Soccer motion recognition accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 10minus4 09SVM C 1 c 5times10minus4 088KNN N_Neighbors 4 085Decision Tree Min_Samples Split 3 max depth 6 086

Journal of Healthcare Engineering 11

5 Conclusion and Discussion

An AIoT system to recognize different soccer (football)motions and assess the skill levels of soccer players wasproposed in this paper e proposed IoT system consists ofwearable devices a mobile device (eg a mobile phone ortablet) and a cloud-based data processing platform In thisproposed system aWWSD ofMEMSmotion sensors is usedto collect raw data from soccer players By using BTLEtechnology the data is transmitted to the cloud-based dataprocessing platform which analyzes the data and outputsthe results to a mobile device in real-time A SVM modelclassification algorithm with an ankle-based attitude anglemodel is proposed to recognize different motions and assessdifferent skill levels is intelligent system can be a newparadigm and emerging technology based on wearablesensors attaining the experience-driven to data-driventransition in the field of sports training and talent selectionand can be easily extended to analyze other foot-relatedsports motions and skill levels

In this paper we propose a system for recognizing soccermotions based on IoT devices which can recognize usersrsquosoccer motions and evaluate the quality of the motionshowever this system still needs to be improved in terms ofstorage performance and computing speed in the cloud system

e increasing popularity of IoT applications demands thecomputing power of IoTsystems and edge computing is one ofthe main methods to enhance the computing speed of IoTapplications [32] In future research we will try to combine thisrecognition system with edge computing as well as [33ndash36] inwhich authors used edge computing in combination with IoTdevices to increase the computational speed of the system andthus reduce its response time

Data Availability

e soccer motion dataset used to support the findings ofthis study are currently under embargo while the researchfindings are commercialized Requests for data 12 monthsafter publication of this article will be considered by thecorresponding author

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the Natural ScienceFoundation of China under grant 61702336 the Hubei

Table 4 e motion recognition result

Criterion Motion Classification AveragePassing ShootingAccuracy 0857 0885 0871Recall 096 092 094F1-score 0906 0902 0904

Table 5 Shooting level assessment accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 093SVM C 1 c 10minus4 091KNN N_Neighbors 4 087Decision tree Min_Samples Split 3 max depth 6 086

Table 6 Passing level classification accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 087SVM C 1 c 10minus4 086KNN N_Neighbors 4 084Decision Tree Min_Samples Split 3 max depth 6 079

Table 7 e assessment result for passing and shooting

CriterionSkill level of passing

AverageSkill level of shooting

AverageElites Amateurs Elites Amateurs

Accuracy 0887 0851 0869 0959 0902 0931Recall 094 08 087 094 092 093F1-score 0913 0825 0869 0949 0911 093

12 Journal of Healthcare Engineering

Provincial Natural Science Foundation of China under grant2020CFB629 the Fundamental Research Funds for theCentral Universities under grants CZT20027 and CSQ21018and the Research Start-Up Funds of South-Central Uni-versity for Nationalities under grant YZZ18006

References

[1] M Bhatia ldquoIot-inspired framework for athlete performanceassessment in smart sport industryrdquo IEEE Internet of ingsJournal vol 8 no 12 pp 9523ndash9530 2020

[2] Q Zhang Y Zhang C Li C Yan Y Duan and H WangldquoSport location-based user clustering with privacy-preserva-tion in wireless iot-driven healthcarerdquo IEEE Access vol 9pp 12906ndash12913 2021

[3] A Rajsp and I Fister ldquoA systematic literature review of in-telligent data analysis methods for smart sport trainingrdquoApplied Sciences vol 10 no 9 p 3013 2020

[4] J Haladjian D Schlabbers S Taheri M arr andB Bruegge ldquoSensor-based detection and classification ofsoccer goalkeeper training exercisesrdquo ACM Transactions onInternet Technology vol 1 no 2 pp 1ndash20 2020

[5] X Bu ldquoHumanmotion gesture recognition algorithm in videobased on convolutional neural features of training imagesrdquoIEEE Access vol 8 pp 160 025ndash160 039 2020

[6] Y Wang M Chen X Wang R H M Chan andW J Li ldquoIotfor next-generation racket sports trainingrdquo IEEE Internet ofings Journal vol 5 no 6 pp 4558ndash4566 2018

[7] Y Wang Y Zhao R H M Chan and W J Li ldquoVolleyballskill assessment using a single wearable micro inertial mea-surement unit at wristrdquo IEEE Access vol 6pp 13 758ndash813 765 2018

[8] N F Ghazali N Shahar N A Rahmad N A J SufriM A Asrsquoari and H F M Latif ldquoCommon sport activityrecognition using inertial sensorrdquo in Proceedings of the IEEE14th International Colloquium on Signal Processing Its Ap-plications (CSPA) pp 67ndash71 Malaysia March 2018

[9] M Ramanathan W-Y Yau and E K Teoh ldquoHuman actionrecognition with video data research and evaluation chal-lengesrdquo IEEE Transactions on Human-Machine Systemsvol 44 no 5 pp 650ndash663 2014

[10] J K Aggarwal and M S Ryoo ldquoHuman activity analysisrdquoACM Computing Surveys vol 43 no 3 pp 1ndash43 2011

[11] M Stikic D Larlus S Ebert and B Schiele ldquoWeakly su-pervised recognition of daily life activities with wearablesensorsrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 33 no 12 pp 2521ndash2537 2011

[12] J Wang Y Chen S Hao X Peng and L Hu ldquoDeep learningfor sensor-based activity recognition a surveyrdquo PatternRecognition Letters vol 119 pp 3ndash11 2019

[13] KWang J He and L Zhang ldquoAttention-based convolutionalneural network for weakly labeled human activitiesrsquo recog-nition with wearable sensorsrdquo IEEE Sensors Journal vol 19no 17 pp 7598ndash7604 2019

[14] C Chen R Jafari and N Kehtarnavaz ldquoImproving humanaction recognition using fusion of depth camera and inertialsensorsrdquo IEEE Transactions on Human-Machine Systemsvol 45 no 1 pp 51ndash61 2015

[15] K Kui Liu C Chen Chen R Jafari and N KehtarnavazldquoFusion of inertial and depth sensor data for robust handgesture recognitionrdquo IEEE Sensors Journal vol 14 no 6pp 1898ndash1903 2014

[16] A Mannini S S Intille M Rosenberger A M Sabatini andW Haskell ldquoActivity recognition using a single accelerometer

placed at the wrist or anklerdquo Medicine amp Science in Sports ampExercise vol 45 no 11 pp 2193ndash2203 2013

[17] A Mannini M Rosenberger W L Haskell A M Sabatiniand S S Intille ldquoActivity recognition in youth using singleaccelerometer placed at wrist or anklerdquoMedicine amp Science inSports amp Exercise vol 49 no 4 pp 801ndash812 2017

[18] M Shoaib S Bosch O Incel H Scholten and P HavingaldquoComplex human activity recognition using smartphone andwrist-worn motion sensorsrdquo Sensors vol 16 no 4 p 4262016

[19] J P Varkey D Pompili and T A Walls ldquoHuman motionrecognition using a wireless sensor-based wearable systemrdquoPersonal And Ubiquitous Computing vol 16 pp 897ndash9102016

[20] J S Akins N R Heebner M Lovalekar and T C SellldquoReliability and validity of instrumented soccer equipmentrdquoJournal of Applied Biomechanics vol 31 no 3 pp 195ndash2012015

[21] A Meamarbashi and S Hossaini ldquoApplication of novel in-ertial technique to compare the kinematics and kinetics of thelegs in the soccer instep kickrdquo Journal of Human Kineticsvol 23 no 1 pp 5ndash13 2010

[22] BMI160 Datasheet Document Bst-Bmi160-Ds000-07 BoschSensortec Reutlingen Germany 2015

[23] G Okeyo L Chen HWang and R Sterritt ldquoDynamic sensordata segmentation for real-time knowledge-driven activityrecognitionrdquo Pervasive and Mobile Computing vol 10pp 155ndash172 2014

[24] S Ioannidou and G Pantazis ldquoHelmert transformationproblem from euler angles method to quaternion algebrardquoISPRS International Journal of Geo-Information vol 9 no 9p 494 2020

[25] N Jouppi C Young N Patil and D Patterson ldquoMotivationfor and evaluation of the first tensor processing unitrdquo IEEEMicro vol 38 no 3 pp 10ndash19 2018

[26] H Ranocha M Sayyari L Dalcin M Parsani andD I Ketcheson ldquoRelaxation Rkutta methods fully discreteexplicit entropy-stable schemes for the compressible euler andNavierndashStokes equationsrdquo SIAM Journal on Scientific Com-puting vol 42 no 2 pp A612ndashA638 2020

[27] A M Andrew An Introduction to Support Vector Machinesand Other Kernel-Based Learning Methods by Nello Chris-tianini and John Shawe-Taylor Cambridge University PressCambridge UK 2000

[28] X Yang and Y L Tian ldquoEigen joints-based action recognitionusing Naıve-Bayes-nearest-neighborrdquo in Proceedings of theIEEE Computer Society Conference on Computer Vision andPattern Recognition Workshops pp 14ndash19 Providence RIUSA June 2012

[29] T V Gestel B Baesens P V Dijcke J A K Suykens andT Alderweireld ldquoLinear and nonlinear credit scoring bycombining logistic regression and support vector machinesrdquoSocial Science Electronic Publishing vol 1 no 4 2005

[30] A Jindal A Dua K Kaur M Singh N Kumar andS Mishra ldquoDecision tree and SVM-based data analytics fortheft detection in smart gridrdquo IEEE Transactions on IndustrialInformatics vol 12 no 3 pp 1005ndash1016 2016

[31] P anh Noi and M Kappas ldquoComparison of random forestk-nearest neighbor and support vector machine classifiers forland cover classification using sentinel-2 imageryrdquo Sensorsvol 18 no 1 p 18 2018

[32] H Li K Ota and M Dong ldquoLearning iot in edge deeplearning for the internet of things with edge computingrdquo IEEEnetwork vol 32 no 1 pp 96ndash101 2018

Journal of Healthcare Engineering 13

[33] W Rafique L Qi I Yaqoob M Imran R U Rasool andW Dou ldquoComplementing iot services through software de-fined networking and edge computing a comprehensivesurveyrdquo IEEE Communications Surveys amp Tutorials vol 22no 3 pp 1761ndash1804 2020

[34] L Yang H Yao J Wang C Jiang A Benslimane and Y LiuldquoMulti-UAV-enabled load-balance mobile-edge computingfor iot networksrdquo IEEE Internet ofings Journal vol 7 no 8pp 6898ndash6908 2020

[35] C Gong F Lin X Gong and Y Lu ldquoIntelligent cooperativeedge computing in internet of thingsrdquo IEEE Internet of ingsJournal vol 7 no 10 pp 9372ndash9382 2020

[36] B C Kavitha R Vallikannu and K S Sankaran ldquoDelay-aware concurrent data management method for iot collab-orative mobile edge computing environmentrdquo Microproces-sors and Microsystems vol 74 Article ID 103021 2020

14 Journal of Healthcare Engineering

Cnb Rot(zψ)Rot(y θ)Rot(xφ)

cos ψ minussin ψ 0

cos ψ cos ψ 0

0 0 1

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1 0 0

0 cos φ minussin φ

0 sin φ cos φ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

cos θ 0 sin θ

0 1 0

minussin θ 0 cos θ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

cos ψ cos θ minussin ψ cos φ + cos ψ sin θ sin φ sin ψ sin φ + cos ψ sin θ cos φ

sin ψ cos θ cos ψ cos φ + sin ψ sin θ sin φ minuscos ψ sin φ + sin ψ sin θ cos φ

minussin θ cos θ sin φ cos θ cos φ

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(7)

Both D1 and Cnb are matrices transforming the object

coordinate system attitude into the geographic coordinatesystem that is

2 p1p3 minus p0p2( 1113857 minussin θ g1

2 p0p1 + p2p3( 1113857 sin φ cos θ g2m

p20 minus p

21 minus p

22 + p

23 cos θ cos φ g3

2 p0p3 + p1p2( 1113857 sin ψ cos θ g4

p20 + p

21 minus p

22 minus p

23 cos ψ cos θ g5

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎩

(8)

e equation for solving the attitude angle by thequaternion is

roll arctan2 p0p3 + p1p2( 1113857

1 minus 2 p20 + p

211113872 1113873

pitch arctan2 p1p3 minus p0p2( 1113857

4 p0p3 + p1p2( 11138572

+ 1 minus 2 p20 + p

211113872 11138731113872 1113873

2

yaw arctan2 p0p1 + p2p3( 1113857

1 minus 2 p22 + p

231113872 1113873

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(9)

e frequency of the sensor repetition rate is 100Hzen the angular velocity value is substituted into the so-lution attitude angle Anminusi Figure 8 shows the comparison ofattitude angle of elites and amateurs

By combining Anminusi with Snminusi as a part of Rnminusi we extractthe basic and morphology features fni of segments R(n)i asshown in Table 1

F(n) (f1 fmn) are merged into one largematrix F

of size n times m e updated data form is shown in Figure 9

42 Principle Component Analysis e cloud server benefitsfrom reducing memory requirements computing load andnecessary bandwidth during model implementation Torepresent identified variables in compact feature variableswe perform PCA processing e model with PCA canachieve higher accuracy than other nonlinear dimensionalityreduction methods [7] PCA is essentially a base transfor-mation that makes transformed data have the largest vari-ance which means the variance between one axis (spindle)and the point is minimized by the rotation of the coordinate

axis and the translation of the coordinate origin Suppose thematrix F with the dimension of n times m which means thatthere is total n samples in m-dimensions F can bedecomposed into UΣVT where U has the same size as F andthe orthogonal matrix V is m times m Σ is a diagonal matrix ofthe same size as V en YrUΣr F(ΣVT)(minus 1)Σr

We extracted a total of 34-dimensional features for avector characterizing where F [F1 F2 F34] AfterPCA the obtained new features can be expressed as Yr

[Yr1 Yr2 Yrm] and m represents the calculateddimension

Yrm F1aij + F2aij + middot middot middot + Fmaij (10)

where aij are eigenvalues of the covariance matrix Equation(10) can be simplified as follows

Yrm F1a1 + F2a2 + middot middot middot + Fmam (11)

43 Support Vector Machine SVM is widely applied in thedomain of machine learning and pattern recognition as atool to solve the classification problem e basic model ofSVM aims to maximize the distance of the closet sampleswhich is called support vector to the hyperplane on theeigenspace [27] In SVM a training data point is regarded asa $p$-dimensional vector the objective is to separate suchpoints using a $(p-1)$-dimensional hyperplane [19] at isthe training points are mapped to points in space to max-imize the width of the gap between the two and severalcategories by SVM Compared to other supervised algo-rithms such as logistics regression (LR) model or NaıveBayes algorithm [28] the SVM algorithm is more suitable todeal with high-dimensional and linear inseparable problemsby freely selecting the parametric model and only usesupport vector as the classification basis of hyperplane whichmeets our expectations as our problem is a small samplelinear inseparable problem

e training dataset is set as T (Xk yk)|X isin Rm yk isin1113864

0 1 nk1 Here are total n samples X represents an

m-dimensional matrix the value of the classification label yk

is 0 or 1 and the current sample is Xk erefore therepresentation of the optimization objective function is

minωb

12ω

2

st yk ω middot xk + b( 1113857 minus 1ge 0 k 1 2 N

(12)

8 Journal of Healthcare Engineering

Here ω is a vector on a hyperplane and the offset b of thesuperflat is along ω from the origin Problem (12) is a convexquadratic programming problem [29] According to theconvex optimization theory transform problem (12) into anunconstrained problem e optimization function can bedenoted as

L(ω b α) 12ω

2minus 1113944

N

k1αkyk ω middot xk + b( 1113857 + 1113944

N

k1αk (13)

where αk is a Lagrangian multiplierαk ge 0 (k 1 2 3 n) e original problem is expressedas

maxα

minωb

L(ω b α) (14)

αk is a Lagrangian multiplier whereαk ge 0 (k 1 2 3 n) Solving ω and b as a minimumproblem we can get the value of ω and b

ω 1113944N

k1αkykxk

1113944

N

k1αkyk 0

⎧⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎩

(15)

Substituting the obtained solution into the Lagrangianfunction for a minimize problem the following optimizationfunction can be obtained after substituting

stminα

12

1113944

N

k11113944

N

j1αiαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 0αk ge 0 k 1 2 N (16)

en introduce a slack variable ζk for (xk yk) for someabnormal sample points make the training set linearly

inseparable e penalty parameter Cge 0 e original issueis described as

Acc

eler

atio

n (g

)45

0

ndash450 100 200

Times

ProfessionalAmateur

(a)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(b)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(c)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(d)

Acc

eler

atio

n (g

)45

0

ndash450 100 200

Times

ProfessionalAmateur

(e)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(f )

Figure 8 Comparison of elitersquos and amateurrsquos attitude angular (a) is the contrast of pitch (b) is the contrast of roll (c) is the contrast ofyaw when passing (d) is the contrast of pitch (e) is the contrast of roll and (f ) is the contrast of yaw when shooting

Journal of Healthcare Engineering 9

Table 1 Basic and morphology features

Number Symbol Description1 Ray Root mean square (RMS) of y-axis acceleration2 Rwx RMS of x-axis angular velocity3 Rwy RMS of y-axis angular velocity4 Rwz RMS of z-axis angular velocity5 Dax Variance of x-axis acceleration6 Day Variance of y-axis acceleration7 Dwy Variance of y-axis angular velocity8 Maxxy Maximum of x-axis acceleration9 Maxay Maximum of y-axis acceleration10 Minay Minimum of y-axis acceleration11 Minwz Minimum of y-axis angular velocity12 Sax Skewness of x-axis acceleration13 Say Skewness of y-axis acceleration14 Saz Skewness of z-axis acceleration15 Swx Skewness of x-axis angular velocity16 Swy Skewness of y-axis angular velocity17 Swz Skewness of z-axis angular velocity18 IQRax Interquartile range of x-axis acceleration19 IQRay Interquartile range of y-axis acceleration20 IQRaz Interquartile range of z-axis acceleration21 IQRwx Interquartile range of x-axis angular velocity22 IQRwy Interquartile range of y-axis angular velocity23 IQRwz Interquartile range of z-axis angular velocity24 Stdax Standard deviation of x-axis acceleration25 Stdwy Standard deviation of y-axis angular velocity26 Mp Mean value of pitch27 Mr Mean value of roll28 My Mean value of yaw29 Rp Root mean square of pitch30 Rr Root mean square of roll31 Ry Root mean square of yaw32 Stdp Standard deviation of pitch33 Stdr Standard deviation of raw34 Stdy Standard deviation of yaw

Preprocess Dimension extended feature selection Matrices merge

Shooting matrices Fshoot

F1s

F2s

F11s

Passing matrices

f1ndash1

f2ndash1

f1ndash2

f2ndash2

f1ndashm1

F1p

F2s F11s

F1p

F2p

F11p

Fpass

f2ndash1f2ndash2

f2ndashm2

F2p

f11ndash1f11ndash2

f11ndashm11

F11p

f2ndashm2

f1ndash1f1ndash2

F1s

f1ndashm1

f11ndash1f11ndash2

f11ndashm11

Subject 1raw data

S1ndash1s S1ndash1p

S2ndash1s S2ndash1p

S11ndash1s S11ndash1p

R1ndash1s R1ndash1p

R2ndash1s R2ndash1p

R11ndash1s R11ndash1p

Subject 2raw data

Subject 11raw data

Angularmodel

6-dimetions

9-dimetions

Featureselect

9-dimetions

m-dimetions

Figure 9 With preprocessing we obtain effective data Snminusis for shooting and Snminusip for passing Add attitude angle for a suitable data formRnminusis for shooting and Rnminusip for passing Select key features Fis and Fip that contribute to maximizing classifier success rates Combine allsamples to obtain Fs and Fp

10 Journal of Healthcare Engineering

stminωbζ

12ω

2+ C 1113944

N

k1ζkyk ω middot xk + b( 1113857ge 1 minus ζk k 1 2 Nζk ge 0 k 1 2 N (17)

e ultimate function is obtained after conversion

stminα

12

1113944

N

k11113944

N

j1αkαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 00le αk leC k 1 2 N (18)

ere are two labels passing and shooting representedby 0 and 1 respectively in the action classification exper-iment We obtained 264 sets for passing and 250 sets forshooting e datasets are randomly divided into trainingand test sets 80 of them used for training and the rest usedas test data Randomly selected parameters are shown inTable 2e selectable parameter C ranges from 1 to 50000 crange is from 000001 to 005 and the kernel function in-cludes Linear RBF Sigmoid and Polynomial In order tosolve overfitting 3-fold cross is adopted Experimental re-sults show that C 1 c is 00001 and linear kernel functionis used in the optimal model

We use the proposed SVM-based classifier to recognizethe shooting and passing Decision Tree [30] and K-nearestneighbor algorithm (KNN) [31] are the competitors in ourexperiments e accuracy of different models for soccermotion recognition can be found in Table 3 Apparently ourproposed SVM-based algorithm has better performancethan other competitorse one with the attitude angle has abetter performance than the others

Table 4 shows the recognition and assessment results ofour proposed systeme recognition accuracies for passingand shooting are 857 and 885 which illustrated a cleardistinction between passing and shooting e accuracy ofclassifying different behaviors can be 871 on average isresult suggests our system works

44TalentSelectionSystem Similar to the above analysis thelabels are elites and amateurs in the talent selection ex-periment e selectable parameters are still randomly se-lected from Table 2 re-fold cross-validation is used againto prevent overfitting Finally the best performing classifiersare selected Experimental results show that C 1 and linearkernel function applies to all optimal models As for cdifferent from motion recognition the value is 000005

As shown in Table 5 the accuracies of shooting levelclassifiers are higher than others ere is a clear differencebetween different level participants when shooting eresults show that the SVMwith the attitude angle model alsohas satisfactory accuracy

e results in Table 6 are all inferior to others whichmeans there might be little difference between participantswhen passing As listed in Tables 3 5 and 6 comparing withKNN and Decision Tree SVM demonstrated superiority indealing with the linear inseparable problem and the onewith attitude angle is better than the other

As shown in Table 7 the assessment results of proposedsystem Our model with attitude angle features obviouslyperforms better than typical model e professionals andamateurs are 887 and 851 respectively for passing For theshooting action the accuracy of the model is 93 which alsosignifies that for the shooting action the skill gap between theprofessional and the amateur is bigger than the passing action

Table 2 Parameters setting of SVM

Penalty parameter (C) c Kernel1 10minus 5 Linear500 5times 10minus 5 Polynomial103 10minus 4 RBF5times 103 5times 10minus 4 Sigmoid104 10minus 3

5times 103 5times 10minus 3

Table 3 Soccer motion recognition accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 10minus4 09SVM C 1 c 5times10minus4 088KNN N_Neighbors 4 085Decision Tree Min_Samples Split 3 max depth 6 086

Journal of Healthcare Engineering 11

5 Conclusion and Discussion

An AIoT system to recognize different soccer (football)motions and assess the skill levels of soccer players wasproposed in this paper e proposed IoT system consists ofwearable devices a mobile device (eg a mobile phone ortablet) and a cloud-based data processing platform In thisproposed system aWWSD ofMEMSmotion sensors is usedto collect raw data from soccer players By using BTLEtechnology the data is transmitted to the cloud-based dataprocessing platform which analyzes the data and outputsthe results to a mobile device in real-time A SVM modelclassification algorithm with an ankle-based attitude anglemodel is proposed to recognize different motions and assessdifferent skill levels is intelligent system can be a newparadigm and emerging technology based on wearablesensors attaining the experience-driven to data-driventransition in the field of sports training and talent selectionand can be easily extended to analyze other foot-relatedsports motions and skill levels

In this paper we propose a system for recognizing soccermotions based on IoT devices which can recognize usersrsquosoccer motions and evaluate the quality of the motionshowever this system still needs to be improved in terms ofstorage performance and computing speed in the cloud system

e increasing popularity of IoT applications demands thecomputing power of IoTsystems and edge computing is one ofthe main methods to enhance the computing speed of IoTapplications [32] In future research we will try to combine thisrecognition system with edge computing as well as [33ndash36] inwhich authors used edge computing in combination with IoTdevices to increase the computational speed of the system andthus reduce its response time

Data Availability

e soccer motion dataset used to support the findings ofthis study are currently under embargo while the researchfindings are commercialized Requests for data 12 monthsafter publication of this article will be considered by thecorresponding author

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the Natural ScienceFoundation of China under grant 61702336 the Hubei

Table 4 e motion recognition result

Criterion Motion Classification AveragePassing ShootingAccuracy 0857 0885 0871Recall 096 092 094F1-score 0906 0902 0904

Table 5 Shooting level assessment accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 093SVM C 1 c 10minus4 091KNN N_Neighbors 4 087Decision tree Min_Samples Split 3 max depth 6 086

Table 6 Passing level classification accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 087SVM C 1 c 10minus4 086KNN N_Neighbors 4 084Decision Tree Min_Samples Split 3 max depth 6 079

Table 7 e assessment result for passing and shooting

CriterionSkill level of passing

AverageSkill level of shooting

AverageElites Amateurs Elites Amateurs

Accuracy 0887 0851 0869 0959 0902 0931Recall 094 08 087 094 092 093F1-score 0913 0825 0869 0949 0911 093

12 Journal of Healthcare Engineering

Provincial Natural Science Foundation of China under grant2020CFB629 the Fundamental Research Funds for theCentral Universities under grants CZT20027 and CSQ21018and the Research Start-Up Funds of South-Central Uni-versity for Nationalities under grant YZZ18006

References

[1] M Bhatia ldquoIot-inspired framework for athlete performanceassessment in smart sport industryrdquo IEEE Internet of ingsJournal vol 8 no 12 pp 9523ndash9530 2020

[2] Q Zhang Y Zhang C Li C Yan Y Duan and H WangldquoSport location-based user clustering with privacy-preserva-tion in wireless iot-driven healthcarerdquo IEEE Access vol 9pp 12906ndash12913 2021

[3] A Rajsp and I Fister ldquoA systematic literature review of in-telligent data analysis methods for smart sport trainingrdquoApplied Sciences vol 10 no 9 p 3013 2020

[4] J Haladjian D Schlabbers S Taheri M arr andB Bruegge ldquoSensor-based detection and classification ofsoccer goalkeeper training exercisesrdquo ACM Transactions onInternet Technology vol 1 no 2 pp 1ndash20 2020

[5] X Bu ldquoHumanmotion gesture recognition algorithm in videobased on convolutional neural features of training imagesrdquoIEEE Access vol 8 pp 160 025ndash160 039 2020

[6] Y Wang M Chen X Wang R H M Chan andW J Li ldquoIotfor next-generation racket sports trainingrdquo IEEE Internet ofings Journal vol 5 no 6 pp 4558ndash4566 2018

[7] Y Wang Y Zhao R H M Chan and W J Li ldquoVolleyballskill assessment using a single wearable micro inertial mea-surement unit at wristrdquo IEEE Access vol 6pp 13 758ndash813 765 2018

[8] N F Ghazali N Shahar N A Rahmad N A J SufriM A Asrsquoari and H F M Latif ldquoCommon sport activityrecognition using inertial sensorrdquo in Proceedings of the IEEE14th International Colloquium on Signal Processing Its Ap-plications (CSPA) pp 67ndash71 Malaysia March 2018

[9] M Ramanathan W-Y Yau and E K Teoh ldquoHuman actionrecognition with video data research and evaluation chal-lengesrdquo IEEE Transactions on Human-Machine Systemsvol 44 no 5 pp 650ndash663 2014

[10] J K Aggarwal and M S Ryoo ldquoHuman activity analysisrdquoACM Computing Surveys vol 43 no 3 pp 1ndash43 2011

[11] M Stikic D Larlus S Ebert and B Schiele ldquoWeakly su-pervised recognition of daily life activities with wearablesensorsrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 33 no 12 pp 2521ndash2537 2011

[12] J Wang Y Chen S Hao X Peng and L Hu ldquoDeep learningfor sensor-based activity recognition a surveyrdquo PatternRecognition Letters vol 119 pp 3ndash11 2019

[13] KWang J He and L Zhang ldquoAttention-based convolutionalneural network for weakly labeled human activitiesrsquo recog-nition with wearable sensorsrdquo IEEE Sensors Journal vol 19no 17 pp 7598ndash7604 2019

[14] C Chen R Jafari and N Kehtarnavaz ldquoImproving humanaction recognition using fusion of depth camera and inertialsensorsrdquo IEEE Transactions on Human-Machine Systemsvol 45 no 1 pp 51ndash61 2015

[15] K Kui Liu C Chen Chen R Jafari and N KehtarnavazldquoFusion of inertial and depth sensor data for robust handgesture recognitionrdquo IEEE Sensors Journal vol 14 no 6pp 1898ndash1903 2014

[16] A Mannini S S Intille M Rosenberger A M Sabatini andW Haskell ldquoActivity recognition using a single accelerometer

placed at the wrist or anklerdquo Medicine amp Science in Sports ampExercise vol 45 no 11 pp 2193ndash2203 2013

[17] A Mannini M Rosenberger W L Haskell A M Sabatiniand S S Intille ldquoActivity recognition in youth using singleaccelerometer placed at wrist or anklerdquoMedicine amp Science inSports amp Exercise vol 49 no 4 pp 801ndash812 2017

[18] M Shoaib S Bosch O Incel H Scholten and P HavingaldquoComplex human activity recognition using smartphone andwrist-worn motion sensorsrdquo Sensors vol 16 no 4 p 4262016

[19] J P Varkey D Pompili and T A Walls ldquoHuman motionrecognition using a wireless sensor-based wearable systemrdquoPersonal And Ubiquitous Computing vol 16 pp 897ndash9102016

[20] J S Akins N R Heebner M Lovalekar and T C SellldquoReliability and validity of instrumented soccer equipmentrdquoJournal of Applied Biomechanics vol 31 no 3 pp 195ndash2012015

[21] A Meamarbashi and S Hossaini ldquoApplication of novel in-ertial technique to compare the kinematics and kinetics of thelegs in the soccer instep kickrdquo Journal of Human Kineticsvol 23 no 1 pp 5ndash13 2010

[22] BMI160 Datasheet Document Bst-Bmi160-Ds000-07 BoschSensortec Reutlingen Germany 2015

[23] G Okeyo L Chen HWang and R Sterritt ldquoDynamic sensordata segmentation for real-time knowledge-driven activityrecognitionrdquo Pervasive and Mobile Computing vol 10pp 155ndash172 2014

[24] S Ioannidou and G Pantazis ldquoHelmert transformationproblem from euler angles method to quaternion algebrardquoISPRS International Journal of Geo-Information vol 9 no 9p 494 2020

[25] N Jouppi C Young N Patil and D Patterson ldquoMotivationfor and evaluation of the first tensor processing unitrdquo IEEEMicro vol 38 no 3 pp 10ndash19 2018

[26] H Ranocha M Sayyari L Dalcin M Parsani andD I Ketcheson ldquoRelaxation Rkutta methods fully discreteexplicit entropy-stable schemes for the compressible euler andNavierndashStokes equationsrdquo SIAM Journal on Scientific Com-puting vol 42 no 2 pp A612ndashA638 2020

[27] A M Andrew An Introduction to Support Vector Machinesand Other Kernel-Based Learning Methods by Nello Chris-tianini and John Shawe-Taylor Cambridge University PressCambridge UK 2000

[28] X Yang and Y L Tian ldquoEigen joints-based action recognitionusing Naıve-Bayes-nearest-neighborrdquo in Proceedings of theIEEE Computer Society Conference on Computer Vision andPattern Recognition Workshops pp 14ndash19 Providence RIUSA June 2012

[29] T V Gestel B Baesens P V Dijcke J A K Suykens andT Alderweireld ldquoLinear and nonlinear credit scoring bycombining logistic regression and support vector machinesrdquoSocial Science Electronic Publishing vol 1 no 4 2005

[30] A Jindal A Dua K Kaur M Singh N Kumar andS Mishra ldquoDecision tree and SVM-based data analytics fortheft detection in smart gridrdquo IEEE Transactions on IndustrialInformatics vol 12 no 3 pp 1005ndash1016 2016

[31] P anh Noi and M Kappas ldquoComparison of random forestk-nearest neighbor and support vector machine classifiers forland cover classification using sentinel-2 imageryrdquo Sensorsvol 18 no 1 p 18 2018

[32] H Li K Ota and M Dong ldquoLearning iot in edge deeplearning for the internet of things with edge computingrdquo IEEEnetwork vol 32 no 1 pp 96ndash101 2018

Journal of Healthcare Engineering 13

[33] W Rafique L Qi I Yaqoob M Imran R U Rasool andW Dou ldquoComplementing iot services through software de-fined networking and edge computing a comprehensivesurveyrdquo IEEE Communications Surveys amp Tutorials vol 22no 3 pp 1761ndash1804 2020

[34] L Yang H Yao J Wang C Jiang A Benslimane and Y LiuldquoMulti-UAV-enabled load-balance mobile-edge computingfor iot networksrdquo IEEE Internet ofings Journal vol 7 no 8pp 6898ndash6908 2020

[35] C Gong F Lin X Gong and Y Lu ldquoIntelligent cooperativeedge computing in internet of thingsrdquo IEEE Internet of ingsJournal vol 7 no 10 pp 9372ndash9382 2020

[36] B C Kavitha R Vallikannu and K S Sankaran ldquoDelay-aware concurrent data management method for iot collab-orative mobile edge computing environmentrdquo Microproces-sors and Microsystems vol 74 Article ID 103021 2020

14 Journal of Healthcare Engineering

Here ω is a vector on a hyperplane and the offset b of thesuperflat is along ω from the origin Problem (12) is a convexquadratic programming problem [29] According to theconvex optimization theory transform problem (12) into anunconstrained problem e optimization function can bedenoted as

L(ω b α) 12ω

2minus 1113944

N

k1αkyk ω middot xk + b( 1113857 + 1113944

N

k1αk (13)

where αk is a Lagrangian multiplierαk ge 0 (k 1 2 3 n) e original problem is expressedas

maxα

minωb

L(ω b α) (14)

αk is a Lagrangian multiplier whereαk ge 0 (k 1 2 3 n) Solving ω and b as a minimumproblem we can get the value of ω and b

ω 1113944N

k1αkykxk

1113944

N

k1αkyk 0

⎧⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎩

(15)

Substituting the obtained solution into the Lagrangianfunction for a minimize problem the following optimizationfunction can be obtained after substituting

stminα

12

1113944

N

k11113944

N

j1αiαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 0αk ge 0 k 1 2 N (16)

en introduce a slack variable ζk for (xk yk) for someabnormal sample points make the training set linearly

inseparable e penalty parameter Cge 0 e original issueis described as

Acc

eler

atio

n (g

)45

0

ndash450 100 200

Times

ProfessionalAmateur

(a)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(b)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(c)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(d)

Acc

eler

atio

n (g

)45

0

ndash450 100 200

Times

ProfessionalAmateur

(e)

Acc

eler

atio

n (g

)

45

0

ndash450 100 200

Times

ProfessionalAmateur

(f )

Figure 8 Comparison of elitersquos and amateurrsquos attitude angular (a) is the contrast of pitch (b) is the contrast of roll (c) is the contrast ofyaw when passing (d) is the contrast of pitch (e) is the contrast of roll and (f ) is the contrast of yaw when shooting

Journal of Healthcare Engineering 9

Table 1 Basic and morphology features

Number Symbol Description1 Ray Root mean square (RMS) of y-axis acceleration2 Rwx RMS of x-axis angular velocity3 Rwy RMS of y-axis angular velocity4 Rwz RMS of z-axis angular velocity5 Dax Variance of x-axis acceleration6 Day Variance of y-axis acceleration7 Dwy Variance of y-axis angular velocity8 Maxxy Maximum of x-axis acceleration9 Maxay Maximum of y-axis acceleration10 Minay Minimum of y-axis acceleration11 Minwz Minimum of y-axis angular velocity12 Sax Skewness of x-axis acceleration13 Say Skewness of y-axis acceleration14 Saz Skewness of z-axis acceleration15 Swx Skewness of x-axis angular velocity16 Swy Skewness of y-axis angular velocity17 Swz Skewness of z-axis angular velocity18 IQRax Interquartile range of x-axis acceleration19 IQRay Interquartile range of y-axis acceleration20 IQRaz Interquartile range of z-axis acceleration21 IQRwx Interquartile range of x-axis angular velocity22 IQRwy Interquartile range of y-axis angular velocity23 IQRwz Interquartile range of z-axis angular velocity24 Stdax Standard deviation of x-axis acceleration25 Stdwy Standard deviation of y-axis angular velocity26 Mp Mean value of pitch27 Mr Mean value of roll28 My Mean value of yaw29 Rp Root mean square of pitch30 Rr Root mean square of roll31 Ry Root mean square of yaw32 Stdp Standard deviation of pitch33 Stdr Standard deviation of raw34 Stdy Standard deviation of yaw

Preprocess Dimension extended feature selection Matrices merge

Shooting matrices Fshoot

F1s

F2s

F11s

Passing matrices

f1ndash1

f2ndash1

f1ndash2

f2ndash2

f1ndashm1

F1p

F2s F11s

F1p

F2p

F11p

Fpass

f2ndash1f2ndash2

f2ndashm2

F2p

f11ndash1f11ndash2

f11ndashm11

F11p

f2ndashm2

f1ndash1f1ndash2

F1s

f1ndashm1

f11ndash1f11ndash2

f11ndashm11

Subject 1raw data

S1ndash1s S1ndash1p

S2ndash1s S2ndash1p

S11ndash1s S11ndash1p

R1ndash1s R1ndash1p

R2ndash1s R2ndash1p

R11ndash1s R11ndash1p

Subject 2raw data

Subject 11raw data

Angularmodel

6-dimetions

9-dimetions

Featureselect

9-dimetions

m-dimetions

Figure 9 With preprocessing we obtain effective data Snminusis for shooting and Snminusip for passing Add attitude angle for a suitable data formRnminusis for shooting and Rnminusip for passing Select key features Fis and Fip that contribute to maximizing classifier success rates Combine allsamples to obtain Fs and Fp

10 Journal of Healthcare Engineering

stminωbζ

12ω

2+ C 1113944

N

k1ζkyk ω middot xk + b( 1113857ge 1 minus ζk k 1 2 Nζk ge 0 k 1 2 N (17)

e ultimate function is obtained after conversion

stminα

12

1113944

N

k11113944

N

j1αkαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 00le αk leC k 1 2 N (18)

ere are two labels passing and shooting representedby 0 and 1 respectively in the action classification exper-iment We obtained 264 sets for passing and 250 sets forshooting e datasets are randomly divided into trainingand test sets 80 of them used for training and the rest usedas test data Randomly selected parameters are shown inTable 2e selectable parameter C ranges from 1 to 50000 crange is from 000001 to 005 and the kernel function in-cludes Linear RBF Sigmoid and Polynomial In order tosolve overfitting 3-fold cross is adopted Experimental re-sults show that C 1 c is 00001 and linear kernel functionis used in the optimal model

We use the proposed SVM-based classifier to recognizethe shooting and passing Decision Tree [30] and K-nearestneighbor algorithm (KNN) [31] are the competitors in ourexperiments e accuracy of different models for soccermotion recognition can be found in Table 3 Apparently ourproposed SVM-based algorithm has better performancethan other competitorse one with the attitude angle has abetter performance than the others

Table 4 shows the recognition and assessment results ofour proposed systeme recognition accuracies for passingand shooting are 857 and 885 which illustrated a cleardistinction between passing and shooting e accuracy ofclassifying different behaviors can be 871 on average isresult suggests our system works

44TalentSelectionSystem Similar to the above analysis thelabels are elites and amateurs in the talent selection ex-periment e selectable parameters are still randomly se-lected from Table 2 re-fold cross-validation is used againto prevent overfitting Finally the best performing classifiersare selected Experimental results show that C 1 and linearkernel function applies to all optimal models As for cdifferent from motion recognition the value is 000005

As shown in Table 5 the accuracies of shooting levelclassifiers are higher than others ere is a clear differencebetween different level participants when shooting eresults show that the SVMwith the attitude angle model alsohas satisfactory accuracy

e results in Table 6 are all inferior to others whichmeans there might be little difference between participantswhen passing As listed in Tables 3 5 and 6 comparing withKNN and Decision Tree SVM demonstrated superiority indealing with the linear inseparable problem and the onewith attitude angle is better than the other

As shown in Table 7 the assessment results of proposedsystem Our model with attitude angle features obviouslyperforms better than typical model e professionals andamateurs are 887 and 851 respectively for passing For theshooting action the accuracy of the model is 93 which alsosignifies that for the shooting action the skill gap between theprofessional and the amateur is bigger than the passing action

Table 2 Parameters setting of SVM

Penalty parameter (C) c Kernel1 10minus 5 Linear500 5times 10minus 5 Polynomial103 10minus 4 RBF5times 103 5times 10minus 4 Sigmoid104 10minus 3

5times 103 5times 10minus 3

Table 3 Soccer motion recognition accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 10minus4 09SVM C 1 c 5times10minus4 088KNN N_Neighbors 4 085Decision Tree Min_Samples Split 3 max depth 6 086

Journal of Healthcare Engineering 11

5 Conclusion and Discussion

An AIoT system to recognize different soccer (football)motions and assess the skill levels of soccer players wasproposed in this paper e proposed IoT system consists ofwearable devices a mobile device (eg a mobile phone ortablet) and a cloud-based data processing platform In thisproposed system aWWSD ofMEMSmotion sensors is usedto collect raw data from soccer players By using BTLEtechnology the data is transmitted to the cloud-based dataprocessing platform which analyzes the data and outputsthe results to a mobile device in real-time A SVM modelclassification algorithm with an ankle-based attitude anglemodel is proposed to recognize different motions and assessdifferent skill levels is intelligent system can be a newparadigm and emerging technology based on wearablesensors attaining the experience-driven to data-driventransition in the field of sports training and talent selectionand can be easily extended to analyze other foot-relatedsports motions and skill levels

In this paper we propose a system for recognizing soccermotions based on IoT devices which can recognize usersrsquosoccer motions and evaluate the quality of the motionshowever this system still needs to be improved in terms ofstorage performance and computing speed in the cloud system

e increasing popularity of IoT applications demands thecomputing power of IoTsystems and edge computing is one ofthe main methods to enhance the computing speed of IoTapplications [32] In future research we will try to combine thisrecognition system with edge computing as well as [33ndash36] inwhich authors used edge computing in combination with IoTdevices to increase the computational speed of the system andthus reduce its response time

Data Availability

e soccer motion dataset used to support the findings ofthis study are currently under embargo while the researchfindings are commercialized Requests for data 12 monthsafter publication of this article will be considered by thecorresponding author

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the Natural ScienceFoundation of China under grant 61702336 the Hubei

Table 4 e motion recognition result

Criterion Motion Classification AveragePassing ShootingAccuracy 0857 0885 0871Recall 096 092 094F1-score 0906 0902 0904

Table 5 Shooting level assessment accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 093SVM C 1 c 10minus4 091KNN N_Neighbors 4 087Decision tree Min_Samples Split 3 max depth 6 086

Table 6 Passing level classification accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 087SVM C 1 c 10minus4 086KNN N_Neighbors 4 084Decision Tree Min_Samples Split 3 max depth 6 079

Table 7 e assessment result for passing and shooting

CriterionSkill level of passing

AverageSkill level of shooting

AverageElites Amateurs Elites Amateurs

Accuracy 0887 0851 0869 0959 0902 0931Recall 094 08 087 094 092 093F1-score 0913 0825 0869 0949 0911 093

12 Journal of Healthcare Engineering

Provincial Natural Science Foundation of China under grant2020CFB629 the Fundamental Research Funds for theCentral Universities under grants CZT20027 and CSQ21018and the Research Start-Up Funds of South-Central Uni-versity for Nationalities under grant YZZ18006

References

[1] M Bhatia ldquoIot-inspired framework for athlete performanceassessment in smart sport industryrdquo IEEE Internet of ingsJournal vol 8 no 12 pp 9523ndash9530 2020

[2] Q Zhang Y Zhang C Li C Yan Y Duan and H WangldquoSport location-based user clustering with privacy-preserva-tion in wireless iot-driven healthcarerdquo IEEE Access vol 9pp 12906ndash12913 2021

[3] A Rajsp and I Fister ldquoA systematic literature review of in-telligent data analysis methods for smart sport trainingrdquoApplied Sciences vol 10 no 9 p 3013 2020

[4] J Haladjian D Schlabbers S Taheri M arr andB Bruegge ldquoSensor-based detection and classification ofsoccer goalkeeper training exercisesrdquo ACM Transactions onInternet Technology vol 1 no 2 pp 1ndash20 2020

[5] X Bu ldquoHumanmotion gesture recognition algorithm in videobased on convolutional neural features of training imagesrdquoIEEE Access vol 8 pp 160 025ndash160 039 2020

[6] Y Wang M Chen X Wang R H M Chan andW J Li ldquoIotfor next-generation racket sports trainingrdquo IEEE Internet ofings Journal vol 5 no 6 pp 4558ndash4566 2018

[7] Y Wang Y Zhao R H M Chan and W J Li ldquoVolleyballskill assessment using a single wearable micro inertial mea-surement unit at wristrdquo IEEE Access vol 6pp 13 758ndash813 765 2018

[8] N F Ghazali N Shahar N A Rahmad N A J SufriM A Asrsquoari and H F M Latif ldquoCommon sport activityrecognition using inertial sensorrdquo in Proceedings of the IEEE14th International Colloquium on Signal Processing Its Ap-plications (CSPA) pp 67ndash71 Malaysia March 2018

[9] M Ramanathan W-Y Yau and E K Teoh ldquoHuman actionrecognition with video data research and evaluation chal-lengesrdquo IEEE Transactions on Human-Machine Systemsvol 44 no 5 pp 650ndash663 2014

[10] J K Aggarwal and M S Ryoo ldquoHuman activity analysisrdquoACM Computing Surveys vol 43 no 3 pp 1ndash43 2011

[11] M Stikic D Larlus S Ebert and B Schiele ldquoWeakly su-pervised recognition of daily life activities with wearablesensorsrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 33 no 12 pp 2521ndash2537 2011

[12] J Wang Y Chen S Hao X Peng and L Hu ldquoDeep learningfor sensor-based activity recognition a surveyrdquo PatternRecognition Letters vol 119 pp 3ndash11 2019

[13] KWang J He and L Zhang ldquoAttention-based convolutionalneural network for weakly labeled human activitiesrsquo recog-nition with wearable sensorsrdquo IEEE Sensors Journal vol 19no 17 pp 7598ndash7604 2019

[14] C Chen R Jafari and N Kehtarnavaz ldquoImproving humanaction recognition using fusion of depth camera and inertialsensorsrdquo IEEE Transactions on Human-Machine Systemsvol 45 no 1 pp 51ndash61 2015

[15] K Kui Liu C Chen Chen R Jafari and N KehtarnavazldquoFusion of inertial and depth sensor data for robust handgesture recognitionrdquo IEEE Sensors Journal vol 14 no 6pp 1898ndash1903 2014

[16] A Mannini S S Intille M Rosenberger A M Sabatini andW Haskell ldquoActivity recognition using a single accelerometer

placed at the wrist or anklerdquo Medicine amp Science in Sports ampExercise vol 45 no 11 pp 2193ndash2203 2013

[17] A Mannini M Rosenberger W L Haskell A M Sabatiniand S S Intille ldquoActivity recognition in youth using singleaccelerometer placed at wrist or anklerdquoMedicine amp Science inSports amp Exercise vol 49 no 4 pp 801ndash812 2017

[18] M Shoaib S Bosch O Incel H Scholten and P HavingaldquoComplex human activity recognition using smartphone andwrist-worn motion sensorsrdquo Sensors vol 16 no 4 p 4262016

[19] J P Varkey D Pompili and T A Walls ldquoHuman motionrecognition using a wireless sensor-based wearable systemrdquoPersonal And Ubiquitous Computing vol 16 pp 897ndash9102016

[20] J S Akins N R Heebner M Lovalekar and T C SellldquoReliability and validity of instrumented soccer equipmentrdquoJournal of Applied Biomechanics vol 31 no 3 pp 195ndash2012015

[21] A Meamarbashi and S Hossaini ldquoApplication of novel in-ertial technique to compare the kinematics and kinetics of thelegs in the soccer instep kickrdquo Journal of Human Kineticsvol 23 no 1 pp 5ndash13 2010

[22] BMI160 Datasheet Document Bst-Bmi160-Ds000-07 BoschSensortec Reutlingen Germany 2015

[23] G Okeyo L Chen HWang and R Sterritt ldquoDynamic sensordata segmentation for real-time knowledge-driven activityrecognitionrdquo Pervasive and Mobile Computing vol 10pp 155ndash172 2014

[24] S Ioannidou and G Pantazis ldquoHelmert transformationproblem from euler angles method to quaternion algebrardquoISPRS International Journal of Geo-Information vol 9 no 9p 494 2020

[25] N Jouppi C Young N Patil and D Patterson ldquoMotivationfor and evaluation of the first tensor processing unitrdquo IEEEMicro vol 38 no 3 pp 10ndash19 2018

[26] H Ranocha M Sayyari L Dalcin M Parsani andD I Ketcheson ldquoRelaxation Rkutta methods fully discreteexplicit entropy-stable schemes for the compressible euler andNavierndashStokes equationsrdquo SIAM Journal on Scientific Com-puting vol 42 no 2 pp A612ndashA638 2020

[27] A M Andrew An Introduction to Support Vector Machinesand Other Kernel-Based Learning Methods by Nello Chris-tianini and John Shawe-Taylor Cambridge University PressCambridge UK 2000

[28] X Yang and Y L Tian ldquoEigen joints-based action recognitionusing Naıve-Bayes-nearest-neighborrdquo in Proceedings of theIEEE Computer Society Conference on Computer Vision andPattern Recognition Workshops pp 14ndash19 Providence RIUSA June 2012

[29] T V Gestel B Baesens P V Dijcke J A K Suykens andT Alderweireld ldquoLinear and nonlinear credit scoring bycombining logistic regression and support vector machinesrdquoSocial Science Electronic Publishing vol 1 no 4 2005

[30] A Jindal A Dua K Kaur M Singh N Kumar andS Mishra ldquoDecision tree and SVM-based data analytics fortheft detection in smart gridrdquo IEEE Transactions on IndustrialInformatics vol 12 no 3 pp 1005ndash1016 2016

[31] P anh Noi and M Kappas ldquoComparison of random forestk-nearest neighbor and support vector machine classifiers forland cover classification using sentinel-2 imageryrdquo Sensorsvol 18 no 1 p 18 2018

[32] H Li K Ota and M Dong ldquoLearning iot in edge deeplearning for the internet of things with edge computingrdquo IEEEnetwork vol 32 no 1 pp 96ndash101 2018

Journal of Healthcare Engineering 13

[33] W Rafique L Qi I Yaqoob M Imran R U Rasool andW Dou ldquoComplementing iot services through software de-fined networking and edge computing a comprehensivesurveyrdquo IEEE Communications Surveys amp Tutorials vol 22no 3 pp 1761ndash1804 2020

[34] L Yang H Yao J Wang C Jiang A Benslimane and Y LiuldquoMulti-UAV-enabled load-balance mobile-edge computingfor iot networksrdquo IEEE Internet ofings Journal vol 7 no 8pp 6898ndash6908 2020

[35] C Gong F Lin X Gong and Y Lu ldquoIntelligent cooperativeedge computing in internet of thingsrdquo IEEE Internet of ingsJournal vol 7 no 10 pp 9372ndash9382 2020

[36] B C Kavitha R Vallikannu and K S Sankaran ldquoDelay-aware concurrent data management method for iot collab-orative mobile edge computing environmentrdquo Microproces-sors and Microsystems vol 74 Article ID 103021 2020

14 Journal of Healthcare Engineering

Table 1 Basic and morphology features

Number Symbol Description1 Ray Root mean square (RMS) of y-axis acceleration2 Rwx RMS of x-axis angular velocity3 Rwy RMS of y-axis angular velocity4 Rwz RMS of z-axis angular velocity5 Dax Variance of x-axis acceleration6 Day Variance of y-axis acceleration7 Dwy Variance of y-axis angular velocity8 Maxxy Maximum of x-axis acceleration9 Maxay Maximum of y-axis acceleration10 Minay Minimum of y-axis acceleration11 Minwz Minimum of y-axis angular velocity12 Sax Skewness of x-axis acceleration13 Say Skewness of y-axis acceleration14 Saz Skewness of z-axis acceleration15 Swx Skewness of x-axis angular velocity16 Swy Skewness of y-axis angular velocity17 Swz Skewness of z-axis angular velocity18 IQRax Interquartile range of x-axis acceleration19 IQRay Interquartile range of y-axis acceleration20 IQRaz Interquartile range of z-axis acceleration21 IQRwx Interquartile range of x-axis angular velocity22 IQRwy Interquartile range of y-axis angular velocity23 IQRwz Interquartile range of z-axis angular velocity24 Stdax Standard deviation of x-axis acceleration25 Stdwy Standard deviation of y-axis angular velocity26 Mp Mean value of pitch27 Mr Mean value of roll28 My Mean value of yaw29 Rp Root mean square of pitch30 Rr Root mean square of roll31 Ry Root mean square of yaw32 Stdp Standard deviation of pitch33 Stdr Standard deviation of raw34 Stdy Standard deviation of yaw

Preprocess Dimension extended feature selection Matrices merge

Shooting matrices Fshoot

F1s

F2s

F11s

Passing matrices

f1ndash1

f2ndash1

f1ndash2

f2ndash2

f1ndashm1

F1p

F2s F11s

F1p

F2p

F11p

Fpass

f2ndash1f2ndash2

f2ndashm2

F2p

f11ndash1f11ndash2

f11ndashm11

F11p

f2ndashm2

f1ndash1f1ndash2

F1s

f1ndashm1

f11ndash1f11ndash2

f11ndashm11

Subject 1raw data

S1ndash1s S1ndash1p

S2ndash1s S2ndash1p

S11ndash1s S11ndash1p

R1ndash1s R1ndash1p

R2ndash1s R2ndash1p

R11ndash1s R11ndash1p

Subject 2raw data

Subject 11raw data

Angularmodel

6-dimetions

9-dimetions

Featureselect

9-dimetions

m-dimetions

Figure 9 With preprocessing we obtain effective data Snminusis for shooting and Snminusip for passing Add attitude angle for a suitable data formRnminusis for shooting and Rnminusip for passing Select key features Fis and Fip that contribute to maximizing classifier success rates Combine allsamples to obtain Fs and Fp

10 Journal of Healthcare Engineering

stminωbζ

12ω

2+ C 1113944

N

k1ζkyk ω middot xk + b( 1113857ge 1 minus ζk k 1 2 Nζk ge 0 k 1 2 N (17)

e ultimate function is obtained after conversion

stminα

12

1113944

N

k11113944

N

j1αkαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 00le αk leC k 1 2 N (18)

ere are two labels passing and shooting representedby 0 and 1 respectively in the action classification exper-iment We obtained 264 sets for passing and 250 sets forshooting e datasets are randomly divided into trainingand test sets 80 of them used for training and the rest usedas test data Randomly selected parameters are shown inTable 2e selectable parameter C ranges from 1 to 50000 crange is from 000001 to 005 and the kernel function in-cludes Linear RBF Sigmoid and Polynomial In order tosolve overfitting 3-fold cross is adopted Experimental re-sults show that C 1 c is 00001 and linear kernel functionis used in the optimal model

We use the proposed SVM-based classifier to recognizethe shooting and passing Decision Tree [30] and K-nearestneighbor algorithm (KNN) [31] are the competitors in ourexperiments e accuracy of different models for soccermotion recognition can be found in Table 3 Apparently ourproposed SVM-based algorithm has better performancethan other competitorse one with the attitude angle has abetter performance than the others

Table 4 shows the recognition and assessment results ofour proposed systeme recognition accuracies for passingand shooting are 857 and 885 which illustrated a cleardistinction between passing and shooting e accuracy ofclassifying different behaviors can be 871 on average isresult suggests our system works

44TalentSelectionSystem Similar to the above analysis thelabels are elites and amateurs in the talent selection ex-periment e selectable parameters are still randomly se-lected from Table 2 re-fold cross-validation is used againto prevent overfitting Finally the best performing classifiersare selected Experimental results show that C 1 and linearkernel function applies to all optimal models As for cdifferent from motion recognition the value is 000005

As shown in Table 5 the accuracies of shooting levelclassifiers are higher than others ere is a clear differencebetween different level participants when shooting eresults show that the SVMwith the attitude angle model alsohas satisfactory accuracy

e results in Table 6 are all inferior to others whichmeans there might be little difference between participantswhen passing As listed in Tables 3 5 and 6 comparing withKNN and Decision Tree SVM demonstrated superiority indealing with the linear inseparable problem and the onewith attitude angle is better than the other

As shown in Table 7 the assessment results of proposedsystem Our model with attitude angle features obviouslyperforms better than typical model e professionals andamateurs are 887 and 851 respectively for passing For theshooting action the accuracy of the model is 93 which alsosignifies that for the shooting action the skill gap between theprofessional and the amateur is bigger than the passing action

Table 2 Parameters setting of SVM

Penalty parameter (C) c Kernel1 10minus 5 Linear500 5times 10minus 5 Polynomial103 10minus 4 RBF5times 103 5times 10minus 4 Sigmoid104 10minus 3

5times 103 5times 10minus 3

Table 3 Soccer motion recognition accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 10minus4 09SVM C 1 c 5times10minus4 088KNN N_Neighbors 4 085Decision Tree Min_Samples Split 3 max depth 6 086

Journal of Healthcare Engineering 11

5 Conclusion and Discussion

An AIoT system to recognize different soccer (football)motions and assess the skill levels of soccer players wasproposed in this paper e proposed IoT system consists ofwearable devices a mobile device (eg a mobile phone ortablet) and a cloud-based data processing platform In thisproposed system aWWSD ofMEMSmotion sensors is usedto collect raw data from soccer players By using BTLEtechnology the data is transmitted to the cloud-based dataprocessing platform which analyzes the data and outputsthe results to a mobile device in real-time A SVM modelclassification algorithm with an ankle-based attitude anglemodel is proposed to recognize different motions and assessdifferent skill levels is intelligent system can be a newparadigm and emerging technology based on wearablesensors attaining the experience-driven to data-driventransition in the field of sports training and talent selectionand can be easily extended to analyze other foot-relatedsports motions and skill levels

In this paper we propose a system for recognizing soccermotions based on IoT devices which can recognize usersrsquosoccer motions and evaluate the quality of the motionshowever this system still needs to be improved in terms ofstorage performance and computing speed in the cloud system

e increasing popularity of IoT applications demands thecomputing power of IoTsystems and edge computing is one ofthe main methods to enhance the computing speed of IoTapplications [32] In future research we will try to combine thisrecognition system with edge computing as well as [33ndash36] inwhich authors used edge computing in combination with IoTdevices to increase the computational speed of the system andthus reduce its response time

Data Availability

e soccer motion dataset used to support the findings ofthis study are currently under embargo while the researchfindings are commercialized Requests for data 12 monthsafter publication of this article will be considered by thecorresponding author

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the Natural ScienceFoundation of China under grant 61702336 the Hubei

Table 4 e motion recognition result

Criterion Motion Classification AveragePassing ShootingAccuracy 0857 0885 0871Recall 096 092 094F1-score 0906 0902 0904

Table 5 Shooting level assessment accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 093SVM C 1 c 10minus4 091KNN N_Neighbors 4 087Decision tree Min_Samples Split 3 max depth 6 086

Table 6 Passing level classification accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 087SVM C 1 c 10minus4 086KNN N_Neighbors 4 084Decision Tree Min_Samples Split 3 max depth 6 079

Table 7 e assessment result for passing and shooting

CriterionSkill level of passing

AverageSkill level of shooting

AverageElites Amateurs Elites Amateurs

Accuracy 0887 0851 0869 0959 0902 0931Recall 094 08 087 094 092 093F1-score 0913 0825 0869 0949 0911 093

12 Journal of Healthcare Engineering

Provincial Natural Science Foundation of China under grant2020CFB629 the Fundamental Research Funds for theCentral Universities under grants CZT20027 and CSQ21018and the Research Start-Up Funds of South-Central Uni-versity for Nationalities under grant YZZ18006

References

[1] M Bhatia ldquoIot-inspired framework for athlete performanceassessment in smart sport industryrdquo IEEE Internet of ingsJournal vol 8 no 12 pp 9523ndash9530 2020

[2] Q Zhang Y Zhang C Li C Yan Y Duan and H WangldquoSport location-based user clustering with privacy-preserva-tion in wireless iot-driven healthcarerdquo IEEE Access vol 9pp 12906ndash12913 2021

[3] A Rajsp and I Fister ldquoA systematic literature review of in-telligent data analysis methods for smart sport trainingrdquoApplied Sciences vol 10 no 9 p 3013 2020

[4] J Haladjian D Schlabbers S Taheri M arr andB Bruegge ldquoSensor-based detection and classification ofsoccer goalkeeper training exercisesrdquo ACM Transactions onInternet Technology vol 1 no 2 pp 1ndash20 2020

[5] X Bu ldquoHumanmotion gesture recognition algorithm in videobased on convolutional neural features of training imagesrdquoIEEE Access vol 8 pp 160 025ndash160 039 2020

[6] Y Wang M Chen X Wang R H M Chan andW J Li ldquoIotfor next-generation racket sports trainingrdquo IEEE Internet ofings Journal vol 5 no 6 pp 4558ndash4566 2018

[7] Y Wang Y Zhao R H M Chan and W J Li ldquoVolleyballskill assessment using a single wearable micro inertial mea-surement unit at wristrdquo IEEE Access vol 6pp 13 758ndash813 765 2018

[8] N F Ghazali N Shahar N A Rahmad N A J SufriM A Asrsquoari and H F M Latif ldquoCommon sport activityrecognition using inertial sensorrdquo in Proceedings of the IEEE14th International Colloquium on Signal Processing Its Ap-plications (CSPA) pp 67ndash71 Malaysia March 2018

[9] M Ramanathan W-Y Yau and E K Teoh ldquoHuman actionrecognition with video data research and evaluation chal-lengesrdquo IEEE Transactions on Human-Machine Systemsvol 44 no 5 pp 650ndash663 2014

[10] J K Aggarwal and M S Ryoo ldquoHuman activity analysisrdquoACM Computing Surveys vol 43 no 3 pp 1ndash43 2011

[11] M Stikic D Larlus S Ebert and B Schiele ldquoWeakly su-pervised recognition of daily life activities with wearablesensorsrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 33 no 12 pp 2521ndash2537 2011

[12] J Wang Y Chen S Hao X Peng and L Hu ldquoDeep learningfor sensor-based activity recognition a surveyrdquo PatternRecognition Letters vol 119 pp 3ndash11 2019

[13] KWang J He and L Zhang ldquoAttention-based convolutionalneural network for weakly labeled human activitiesrsquo recog-nition with wearable sensorsrdquo IEEE Sensors Journal vol 19no 17 pp 7598ndash7604 2019

[14] C Chen R Jafari and N Kehtarnavaz ldquoImproving humanaction recognition using fusion of depth camera and inertialsensorsrdquo IEEE Transactions on Human-Machine Systemsvol 45 no 1 pp 51ndash61 2015

[15] K Kui Liu C Chen Chen R Jafari and N KehtarnavazldquoFusion of inertial and depth sensor data for robust handgesture recognitionrdquo IEEE Sensors Journal vol 14 no 6pp 1898ndash1903 2014

[16] A Mannini S S Intille M Rosenberger A M Sabatini andW Haskell ldquoActivity recognition using a single accelerometer

placed at the wrist or anklerdquo Medicine amp Science in Sports ampExercise vol 45 no 11 pp 2193ndash2203 2013

[17] A Mannini M Rosenberger W L Haskell A M Sabatiniand S S Intille ldquoActivity recognition in youth using singleaccelerometer placed at wrist or anklerdquoMedicine amp Science inSports amp Exercise vol 49 no 4 pp 801ndash812 2017

[18] M Shoaib S Bosch O Incel H Scholten and P HavingaldquoComplex human activity recognition using smartphone andwrist-worn motion sensorsrdquo Sensors vol 16 no 4 p 4262016

[19] J P Varkey D Pompili and T A Walls ldquoHuman motionrecognition using a wireless sensor-based wearable systemrdquoPersonal And Ubiquitous Computing vol 16 pp 897ndash9102016

[20] J S Akins N R Heebner M Lovalekar and T C SellldquoReliability and validity of instrumented soccer equipmentrdquoJournal of Applied Biomechanics vol 31 no 3 pp 195ndash2012015

[21] A Meamarbashi and S Hossaini ldquoApplication of novel in-ertial technique to compare the kinematics and kinetics of thelegs in the soccer instep kickrdquo Journal of Human Kineticsvol 23 no 1 pp 5ndash13 2010

[22] BMI160 Datasheet Document Bst-Bmi160-Ds000-07 BoschSensortec Reutlingen Germany 2015

[23] G Okeyo L Chen HWang and R Sterritt ldquoDynamic sensordata segmentation for real-time knowledge-driven activityrecognitionrdquo Pervasive and Mobile Computing vol 10pp 155ndash172 2014

[24] S Ioannidou and G Pantazis ldquoHelmert transformationproblem from euler angles method to quaternion algebrardquoISPRS International Journal of Geo-Information vol 9 no 9p 494 2020

[25] N Jouppi C Young N Patil and D Patterson ldquoMotivationfor and evaluation of the first tensor processing unitrdquo IEEEMicro vol 38 no 3 pp 10ndash19 2018

[26] H Ranocha M Sayyari L Dalcin M Parsani andD I Ketcheson ldquoRelaxation Rkutta methods fully discreteexplicit entropy-stable schemes for the compressible euler andNavierndashStokes equationsrdquo SIAM Journal on Scientific Com-puting vol 42 no 2 pp A612ndashA638 2020

[27] A M Andrew An Introduction to Support Vector Machinesand Other Kernel-Based Learning Methods by Nello Chris-tianini and John Shawe-Taylor Cambridge University PressCambridge UK 2000

[28] X Yang and Y L Tian ldquoEigen joints-based action recognitionusing Naıve-Bayes-nearest-neighborrdquo in Proceedings of theIEEE Computer Society Conference on Computer Vision andPattern Recognition Workshops pp 14ndash19 Providence RIUSA June 2012

[29] T V Gestel B Baesens P V Dijcke J A K Suykens andT Alderweireld ldquoLinear and nonlinear credit scoring bycombining logistic regression and support vector machinesrdquoSocial Science Electronic Publishing vol 1 no 4 2005

[30] A Jindal A Dua K Kaur M Singh N Kumar andS Mishra ldquoDecision tree and SVM-based data analytics fortheft detection in smart gridrdquo IEEE Transactions on IndustrialInformatics vol 12 no 3 pp 1005ndash1016 2016

[31] P anh Noi and M Kappas ldquoComparison of random forestk-nearest neighbor and support vector machine classifiers forland cover classification using sentinel-2 imageryrdquo Sensorsvol 18 no 1 p 18 2018

[32] H Li K Ota and M Dong ldquoLearning iot in edge deeplearning for the internet of things with edge computingrdquo IEEEnetwork vol 32 no 1 pp 96ndash101 2018

Journal of Healthcare Engineering 13

[33] W Rafique L Qi I Yaqoob M Imran R U Rasool andW Dou ldquoComplementing iot services through software de-fined networking and edge computing a comprehensivesurveyrdquo IEEE Communications Surveys amp Tutorials vol 22no 3 pp 1761ndash1804 2020

[34] L Yang H Yao J Wang C Jiang A Benslimane and Y LiuldquoMulti-UAV-enabled load-balance mobile-edge computingfor iot networksrdquo IEEE Internet ofings Journal vol 7 no 8pp 6898ndash6908 2020

[35] C Gong F Lin X Gong and Y Lu ldquoIntelligent cooperativeedge computing in internet of thingsrdquo IEEE Internet of ingsJournal vol 7 no 10 pp 9372ndash9382 2020

[36] B C Kavitha R Vallikannu and K S Sankaran ldquoDelay-aware concurrent data management method for iot collab-orative mobile edge computing environmentrdquo Microproces-sors and Microsystems vol 74 Article ID 103021 2020

14 Journal of Healthcare Engineering

stminωbζ

12ω

2+ C 1113944

N

k1ζkyk ω middot xk + b( 1113857ge 1 minus ζk k 1 2 Nζk ge 0 k 1 2 N (17)

e ultimate function is obtained after conversion

stminα

12

1113944

N

k11113944

N

j1αkαjykyj xk middot xj1113872 1113873 minus 1113944

N

k1αk 1113944

N

k1αkyk 00le αk leC k 1 2 N (18)

ere are two labels passing and shooting representedby 0 and 1 respectively in the action classification exper-iment We obtained 264 sets for passing and 250 sets forshooting e datasets are randomly divided into trainingand test sets 80 of them used for training and the rest usedas test data Randomly selected parameters are shown inTable 2e selectable parameter C ranges from 1 to 50000 crange is from 000001 to 005 and the kernel function in-cludes Linear RBF Sigmoid and Polynomial In order tosolve overfitting 3-fold cross is adopted Experimental re-sults show that C 1 c is 00001 and linear kernel functionis used in the optimal model

We use the proposed SVM-based classifier to recognizethe shooting and passing Decision Tree [30] and K-nearestneighbor algorithm (KNN) [31] are the competitors in ourexperiments e accuracy of different models for soccermotion recognition can be found in Table 3 Apparently ourproposed SVM-based algorithm has better performancethan other competitorse one with the attitude angle has abetter performance than the others

Table 4 shows the recognition and assessment results ofour proposed systeme recognition accuracies for passingand shooting are 857 and 885 which illustrated a cleardistinction between passing and shooting e accuracy ofclassifying different behaviors can be 871 on average isresult suggests our system works

44TalentSelectionSystem Similar to the above analysis thelabels are elites and amateurs in the talent selection ex-periment e selectable parameters are still randomly se-lected from Table 2 re-fold cross-validation is used againto prevent overfitting Finally the best performing classifiersare selected Experimental results show that C 1 and linearkernel function applies to all optimal models As for cdifferent from motion recognition the value is 000005

As shown in Table 5 the accuracies of shooting levelclassifiers are higher than others ere is a clear differencebetween different level participants when shooting eresults show that the SVMwith the attitude angle model alsohas satisfactory accuracy

e results in Table 6 are all inferior to others whichmeans there might be little difference between participantswhen passing As listed in Tables 3 5 and 6 comparing withKNN and Decision Tree SVM demonstrated superiority indealing with the linear inseparable problem and the onewith attitude angle is better than the other

As shown in Table 7 the assessment results of proposedsystem Our model with attitude angle features obviouslyperforms better than typical model e professionals andamateurs are 887 and 851 respectively for passing For theshooting action the accuracy of the model is 93 which alsosignifies that for the shooting action the skill gap between theprofessional and the amateur is bigger than the passing action

Table 2 Parameters setting of SVM

Penalty parameter (C) c Kernel1 10minus 5 Linear500 5times 10minus 5 Polynomial103 10minus 4 RBF5times 103 5times 10minus 4 Sigmoid104 10minus 3

5times 103 5times 10minus 3

Table 3 Soccer motion recognition accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 10minus4 09SVM C 1 c 5times10minus4 088KNN N_Neighbors 4 085Decision Tree Min_Samples Split 3 max depth 6 086

Journal of Healthcare Engineering 11

5 Conclusion and Discussion

An AIoT system to recognize different soccer (football)motions and assess the skill levels of soccer players wasproposed in this paper e proposed IoT system consists ofwearable devices a mobile device (eg a mobile phone ortablet) and a cloud-based data processing platform In thisproposed system aWWSD ofMEMSmotion sensors is usedto collect raw data from soccer players By using BTLEtechnology the data is transmitted to the cloud-based dataprocessing platform which analyzes the data and outputsthe results to a mobile device in real-time A SVM modelclassification algorithm with an ankle-based attitude anglemodel is proposed to recognize different motions and assessdifferent skill levels is intelligent system can be a newparadigm and emerging technology based on wearablesensors attaining the experience-driven to data-driventransition in the field of sports training and talent selectionand can be easily extended to analyze other foot-relatedsports motions and skill levels

In this paper we propose a system for recognizing soccermotions based on IoT devices which can recognize usersrsquosoccer motions and evaluate the quality of the motionshowever this system still needs to be improved in terms ofstorage performance and computing speed in the cloud system

e increasing popularity of IoT applications demands thecomputing power of IoTsystems and edge computing is one ofthe main methods to enhance the computing speed of IoTapplications [32] In future research we will try to combine thisrecognition system with edge computing as well as [33ndash36] inwhich authors used edge computing in combination with IoTdevices to increase the computational speed of the system andthus reduce its response time

Data Availability

e soccer motion dataset used to support the findings ofthis study are currently under embargo while the researchfindings are commercialized Requests for data 12 monthsafter publication of this article will be considered by thecorresponding author

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the Natural ScienceFoundation of China under grant 61702336 the Hubei

Table 4 e motion recognition result

Criterion Motion Classification AveragePassing ShootingAccuracy 0857 0885 0871Recall 096 092 094F1-score 0906 0902 0904

Table 5 Shooting level assessment accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 093SVM C 1 c 10minus4 091KNN N_Neighbors 4 087Decision tree Min_Samples Split 3 max depth 6 086

Table 6 Passing level classification accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 087SVM C 1 c 10minus4 086KNN N_Neighbors 4 084Decision Tree Min_Samples Split 3 max depth 6 079

Table 7 e assessment result for passing and shooting

CriterionSkill level of passing

AverageSkill level of shooting

AverageElites Amateurs Elites Amateurs

Accuracy 0887 0851 0869 0959 0902 0931Recall 094 08 087 094 092 093F1-score 0913 0825 0869 0949 0911 093

12 Journal of Healthcare Engineering

Provincial Natural Science Foundation of China under grant2020CFB629 the Fundamental Research Funds for theCentral Universities under grants CZT20027 and CSQ21018and the Research Start-Up Funds of South-Central Uni-versity for Nationalities under grant YZZ18006

References

[1] M Bhatia ldquoIot-inspired framework for athlete performanceassessment in smart sport industryrdquo IEEE Internet of ingsJournal vol 8 no 12 pp 9523ndash9530 2020

[2] Q Zhang Y Zhang C Li C Yan Y Duan and H WangldquoSport location-based user clustering with privacy-preserva-tion in wireless iot-driven healthcarerdquo IEEE Access vol 9pp 12906ndash12913 2021

[3] A Rajsp and I Fister ldquoA systematic literature review of in-telligent data analysis methods for smart sport trainingrdquoApplied Sciences vol 10 no 9 p 3013 2020

[4] J Haladjian D Schlabbers S Taheri M arr andB Bruegge ldquoSensor-based detection and classification ofsoccer goalkeeper training exercisesrdquo ACM Transactions onInternet Technology vol 1 no 2 pp 1ndash20 2020

[5] X Bu ldquoHumanmotion gesture recognition algorithm in videobased on convolutional neural features of training imagesrdquoIEEE Access vol 8 pp 160 025ndash160 039 2020

[6] Y Wang M Chen X Wang R H M Chan andW J Li ldquoIotfor next-generation racket sports trainingrdquo IEEE Internet ofings Journal vol 5 no 6 pp 4558ndash4566 2018

[7] Y Wang Y Zhao R H M Chan and W J Li ldquoVolleyballskill assessment using a single wearable micro inertial mea-surement unit at wristrdquo IEEE Access vol 6pp 13 758ndash813 765 2018

[8] N F Ghazali N Shahar N A Rahmad N A J SufriM A Asrsquoari and H F M Latif ldquoCommon sport activityrecognition using inertial sensorrdquo in Proceedings of the IEEE14th International Colloquium on Signal Processing Its Ap-plications (CSPA) pp 67ndash71 Malaysia March 2018

[9] M Ramanathan W-Y Yau and E K Teoh ldquoHuman actionrecognition with video data research and evaluation chal-lengesrdquo IEEE Transactions on Human-Machine Systemsvol 44 no 5 pp 650ndash663 2014

[10] J K Aggarwal and M S Ryoo ldquoHuman activity analysisrdquoACM Computing Surveys vol 43 no 3 pp 1ndash43 2011

[11] M Stikic D Larlus S Ebert and B Schiele ldquoWeakly su-pervised recognition of daily life activities with wearablesensorsrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 33 no 12 pp 2521ndash2537 2011

[12] J Wang Y Chen S Hao X Peng and L Hu ldquoDeep learningfor sensor-based activity recognition a surveyrdquo PatternRecognition Letters vol 119 pp 3ndash11 2019

[13] KWang J He and L Zhang ldquoAttention-based convolutionalneural network for weakly labeled human activitiesrsquo recog-nition with wearable sensorsrdquo IEEE Sensors Journal vol 19no 17 pp 7598ndash7604 2019

[14] C Chen R Jafari and N Kehtarnavaz ldquoImproving humanaction recognition using fusion of depth camera and inertialsensorsrdquo IEEE Transactions on Human-Machine Systemsvol 45 no 1 pp 51ndash61 2015

[15] K Kui Liu C Chen Chen R Jafari and N KehtarnavazldquoFusion of inertial and depth sensor data for robust handgesture recognitionrdquo IEEE Sensors Journal vol 14 no 6pp 1898ndash1903 2014

[16] A Mannini S S Intille M Rosenberger A M Sabatini andW Haskell ldquoActivity recognition using a single accelerometer

placed at the wrist or anklerdquo Medicine amp Science in Sports ampExercise vol 45 no 11 pp 2193ndash2203 2013

[17] A Mannini M Rosenberger W L Haskell A M Sabatiniand S S Intille ldquoActivity recognition in youth using singleaccelerometer placed at wrist or anklerdquoMedicine amp Science inSports amp Exercise vol 49 no 4 pp 801ndash812 2017

[18] M Shoaib S Bosch O Incel H Scholten and P HavingaldquoComplex human activity recognition using smartphone andwrist-worn motion sensorsrdquo Sensors vol 16 no 4 p 4262016

[19] J P Varkey D Pompili and T A Walls ldquoHuman motionrecognition using a wireless sensor-based wearable systemrdquoPersonal And Ubiquitous Computing vol 16 pp 897ndash9102016

[20] J S Akins N R Heebner M Lovalekar and T C SellldquoReliability and validity of instrumented soccer equipmentrdquoJournal of Applied Biomechanics vol 31 no 3 pp 195ndash2012015

[21] A Meamarbashi and S Hossaini ldquoApplication of novel in-ertial technique to compare the kinematics and kinetics of thelegs in the soccer instep kickrdquo Journal of Human Kineticsvol 23 no 1 pp 5ndash13 2010

[22] BMI160 Datasheet Document Bst-Bmi160-Ds000-07 BoschSensortec Reutlingen Germany 2015

[23] G Okeyo L Chen HWang and R Sterritt ldquoDynamic sensordata segmentation for real-time knowledge-driven activityrecognitionrdquo Pervasive and Mobile Computing vol 10pp 155ndash172 2014

[24] S Ioannidou and G Pantazis ldquoHelmert transformationproblem from euler angles method to quaternion algebrardquoISPRS International Journal of Geo-Information vol 9 no 9p 494 2020

[25] N Jouppi C Young N Patil and D Patterson ldquoMotivationfor and evaluation of the first tensor processing unitrdquo IEEEMicro vol 38 no 3 pp 10ndash19 2018

[26] H Ranocha M Sayyari L Dalcin M Parsani andD I Ketcheson ldquoRelaxation Rkutta methods fully discreteexplicit entropy-stable schemes for the compressible euler andNavierndashStokes equationsrdquo SIAM Journal on Scientific Com-puting vol 42 no 2 pp A612ndashA638 2020

[27] A M Andrew An Introduction to Support Vector Machinesand Other Kernel-Based Learning Methods by Nello Chris-tianini and John Shawe-Taylor Cambridge University PressCambridge UK 2000

[28] X Yang and Y L Tian ldquoEigen joints-based action recognitionusing Naıve-Bayes-nearest-neighborrdquo in Proceedings of theIEEE Computer Society Conference on Computer Vision andPattern Recognition Workshops pp 14ndash19 Providence RIUSA June 2012

[29] T V Gestel B Baesens P V Dijcke J A K Suykens andT Alderweireld ldquoLinear and nonlinear credit scoring bycombining logistic regression and support vector machinesrdquoSocial Science Electronic Publishing vol 1 no 4 2005

[30] A Jindal A Dua K Kaur M Singh N Kumar andS Mishra ldquoDecision tree and SVM-based data analytics fortheft detection in smart gridrdquo IEEE Transactions on IndustrialInformatics vol 12 no 3 pp 1005ndash1016 2016

[31] P anh Noi and M Kappas ldquoComparison of random forestk-nearest neighbor and support vector machine classifiers forland cover classification using sentinel-2 imageryrdquo Sensorsvol 18 no 1 p 18 2018

[32] H Li K Ota and M Dong ldquoLearning iot in edge deeplearning for the internet of things with edge computingrdquo IEEEnetwork vol 32 no 1 pp 96ndash101 2018

Journal of Healthcare Engineering 13

[33] W Rafique L Qi I Yaqoob M Imran R U Rasool andW Dou ldquoComplementing iot services through software de-fined networking and edge computing a comprehensivesurveyrdquo IEEE Communications Surveys amp Tutorials vol 22no 3 pp 1761ndash1804 2020

[34] L Yang H Yao J Wang C Jiang A Benslimane and Y LiuldquoMulti-UAV-enabled load-balance mobile-edge computingfor iot networksrdquo IEEE Internet ofings Journal vol 7 no 8pp 6898ndash6908 2020

[35] C Gong F Lin X Gong and Y Lu ldquoIntelligent cooperativeedge computing in internet of thingsrdquo IEEE Internet of ingsJournal vol 7 no 10 pp 9372ndash9382 2020

[36] B C Kavitha R Vallikannu and K S Sankaran ldquoDelay-aware concurrent data management method for iot collab-orative mobile edge computing environmentrdquo Microproces-sors and Microsystems vol 74 Article ID 103021 2020

14 Journal of Healthcare Engineering

5 Conclusion and Discussion

An AIoT system to recognize different soccer (football)motions and assess the skill levels of soccer players wasproposed in this paper e proposed IoT system consists ofwearable devices a mobile device (eg a mobile phone ortablet) and a cloud-based data processing platform In thisproposed system aWWSD ofMEMSmotion sensors is usedto collect raw data from soccer players By using BTLEtechnology the data is transmitted to the cloud-based dataprocessing platform which analyzes the data and outputsthe results to a mobile device in real-time A SVM modelclassification algorithm with an ankle-based attitude anglemodel is proposed to recognize different motions and assessdifferent skill levels is intelligent system can be a newparadigm and emerging technology based on wearablesensors attaining the experience-driven to data-driventransition in the field of sports training and talent selectionand can be easily extended to analyze other foot-relatedsports motions and skill levels

In this paper we propose a system for recognizing soccermotions based on IoT devices which can recognize usersrsquosoccer motions and evaluate the quality of the motionshowever this system still needs to be improved in terms ofstorage performance and computing speed in the cloud system

e increasing popularity of IoT applications demands thecomputing power of IoTsystems and edge computing is one ofthe main methods to enhance the computing speed of IoTapplications [32] In future research we will try to combine thisrecognition system with edge computing as well as [33ndash36] inwhich authors used edge computing in combination with IoTdevices to increase the computational speed of the system andthus reduce its response time

Data Availability

e soccer motion dataset used to support the findings ofthis study are currently under embargo while the researchfindings are commercialized Requests for data 12 monthsafter publication of this article will be considered by thecorresponding author

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the Natural ScienceFoundation of China under grant 61702336 the Hubei

Table 4 e motion recognition result

Criterion Motion Classification AveragePassing ShootingAccuracy 0857 0885 0871Recall 096 092 094F1-score 0906 0902 0904

Table 5 Shooting level assessment accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 093SVM C 1 c 10minus4 091KNN N_Neighbors 4 087Decision tree Min_Samples Split 3 max depth 6 086

Table 6 Passing level classification accuracies comparison

Algorithms Parameter AccuracySVM+ angle trajectory model C 1 c 5times10minus5 087SVM C 1 c 10minus4 086KNN N_Neighbors 4 084Decision Tree Min_Samples Split 3 max depth 6 079

Table 7 e assessment result for passing and shooting

CriterionSkill level of passing

AverageSkill level of shooting

AverageElites Amateurs Elites Amateurs

Accuracy 0887 0851 0869 0959 0902 0931Recall 094 08 087 094 092 093F1-score 0913 0825 0869 0949 0911 093

12 Journal of Healthcare Engineering

Provincial Natural Science Foundation of China under grant2020CFB629 the Fundamental Research Funds for theCentral Universities under grants CZT20027 and CSQ21018and the Research Start-Up Funds of South-Central Uni-versity for Nationalities under grant YZZ18006

References

[1] M Bhatia ldquoIot-inspired framework for athlete performanceassessment in smart sport industryrdquo IEEE Internet of ingsJournal vol 8 no 12 pp 9523ndash9530 2020

[2] Q Zhang Y Zhang C Li C Yan Y Duan and H WangldquoSport location-based user clustering with privacy-preserva-tion in wireless iot-driven healthcarerdquo IEEE Access vol 9pp 12906ndash12913 2021

[3] A Rajsp and I Fister ldquoA systematic literature review of in-telligent data analysis methods for smart sport trainingrdquoApplied Sciences vol 10 no 9 p 3013 2020

[4] J Haladjian D Schlabbers S Taheri M arr andB Bruegge ldquoSensor-based detection and classification ofsoccer goalkeeper training exercisesrdquo ACM Transactions onInternet Technology vol 1 no 2 pp 1ndash20 2020

[5] X Bu ldquoHumanmotion gesture recognition algorithm in videobased on convolutional neural features of training imagesrdquoIEEE Access vol 8 pp 160 025ndash160 039 2020

[6] Y Wang M Chen X Wang R H M Chan andW J Li ldquoIotfor next-generation racket sports trainingrdquo IEEE Internet ofings Journal vol 5 no 6 pp 4558ndash4566 2018

[7] Y Wang Y Zhao R H M Chan and W J Li ldquoVolleyballskill assessment using a single wearable micro inertial mea-surement unit at wristrdquo IEEE Access vol 6pp 13 758ndash813 765 2018

[8] N F Ghazali N Shahar N A Rahmad N A J SufriM A Asrsquoari and H F M Latif ldquoCommon sport activityrecognition using inertial sensorrdquo in Proceedings of the IEEE14th International Colloquium on Signal Processing Its Ap-plications (CSPA) pp 67ndash71 Malaysia March 2018

[9] M Ramanathan W-Y Yau and E K Teoh ldquoHuman actionrecognition with video data research and evaluation chal-lengesrdquo IEEE Transactions on Human-Machine Systemsvol 44 no 5 pp 650ndash663 2014

[10] J K Aggarwal and M S Ryoo ldquoHuman activity analysisrdquoACM Computing Surveys vol 43 no 3 pp 1ndash43 2011

[11] M Stikic D Larlus S Ebert and B Schiele ldquoWeakly su-pervised recognition of daily life activities with wearablesensorsrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 33 no 12 pp 2521ndash2537 2011

[12] J Wang Y Chen S Hao X Peng and L Hu ldquoDeep learningfor sensor-based activity recognition a surveyrdquo PatternRecognition Letters vol 119 pp 3ndash11 2019

[13] KWang J He and L Zhang ldquoAttention-based convolutionalneural network for weakly labeled human activitiesrsquo recog-nition with wearable sensorsrdquo IEEE Sensors Journal vol 19no 17 pp 7598ndash7604 2019

[14] C Chen R Jafari and N Kehtarnavaz ldquoImproving humanaction recognition using fusion of depth camera and inertialsensorsrdquo IEEE Transactions on Human-Machine Systemsvol 45 no 1 pp 51ndash61 2015

[15] K Kui Liu C Chen Chen R Jafari and N KehtarnavazldquoFusion of inertial and depth sensor data for robust handgesture recognitionrdquo IEEE Sensors Journal vol 14 no 6pp 1898ndash1903 2014

[16] A Mannini S S Intille M Rosenberger A M Sabatini andW Haskell ldquoActivity recognition using a single accelerometer

placed at the wrist or anklerdquo Medicine amp Science in Sports ampExercise vol 45 no 11 pp 2193ndash2203 2013

[17] A Mannini M Rosenberger W L Haskell A M Sabatiniand S S Intille ldquoActivity recognition in youth using singleaccelerometer placed at wrist or anklerdquoMedicine amp Science inSports amp Exercise vol 49 no 4 pp 801ndash812 2017

[18] M Shoaib S Bosch O Incel H Scholten and P HavingaldquoComplex human activity recognition using smartphone andwrist-worn motion sensorsrdquo Sensors vol 16 no 4 p 4262016

[19] J P Varkey D Pompili and T A Walls ldquoHuman motionrecognition using a wireless sensor-based wearable systemrdquoPersonal And Ubiquitous Computing vol 16 pp 897ndash9102016

[20] J S Akins N R Heebner M Lovalekar and T C SellldquoReliability and validity of instrumented soccer equipmentrdquoJournal of Applied Biomechanics vol 31 no 3 pp 195ndash2012015

[21] A Meamarbashi and S Hossaini ldquoApplication of novel in-ertial technique to compare the kinematics and kinetics of thelegs in the soccer instep kickrdquo Journal of Human Kineticsvol 23 no 1 pp 5ndash13 2010

[22] BMI160 Datasheet Document Bst-Bmi160-Ds000-07 BoschSensortec Reutlingen Germany 2015

[23] G Okeyo L Chen HWang and R Sterritt ldquoDynamic sensordata segmentation for real-time knowledge-driven activityrecognitionrdquo Pervasive and Mobile Computing vol 10pp 155ndash172 2014

[24] S Ioannidou and G Pantazis ldquoHelmert transformationproblem from euler angles method to quaternion algebrardquoISPRS International Journal of Geo-Information vol 9 no 9p 494 2020

[25] N Jouppi C Young N Patil and D Patterson ldquoMotivationfor and evaluation of the first tensor processing unitrdquo IEEEMicro vol 38 no 3 pp 10ndash19 2018

[26] H Ranocha M Sayyari L Dalcin M Parsani andD I Ketcheson ldquoRelaxation Rkutta methods fully discreteexplicit entropy-stable schemes for the compressible euler andNavierndashStokes equationsrdquo SIAM Journal on Scientific Com-puting vol 42 no 2 pp A612ndashA638 2020

[27] A M Andrew An Introduction to Support Vector Machinesand Other Kernel-Based Learning Methods by Nello Chris-tianini and John Shawe-Taylor Cambridge University PressCambridge UK 2000

[28] X Yang and Y L Tian ldquoEigen joints-based action recognitionusing Naıve-Bayes-nearest-neighborrdquo in Proceedings of theIEEE Computer Society Conference on Computer Vision andPattern Recognition Workshops pp 14ndash19 Providence RIUSA June 2012

[29] T V Gestel B Baesens P V Dijcke J A K Suykens andT Alderweireld ldquoLinear and nonlinear credit scoring bycombining logistic regression and support vector machinesrdquoSocial Science Electronic Publishing vol 1 no 4 2005

[30] A Jindal A Dua K Kaur M Singh N Kumar andS Mishra ldquoDecision tree and SVM-based data analytics fortheft detection in smart gridrdquo IEEE Transactions on IndustrialInformatics vol 12 no 3 pp 1005ndash1016 2016

[31] P anh Noi and M Kappas ldquoComparison of random forestk-nearest neighbor and support vector machine classifiers forland cover classification using sentinel-2 imageryrdquo Sensorsvol 18 no 1 p 18 2018

[32] H Li K Ota and M Dong ldquoLearning iot in edge deeplearning for the internet of things with edge computingrdquo IEEEnetwork vol 32 no 1 pp 96ndash101 2018

Journal of Healthcare Engineering 13

[33] W Rafique L Qi I Yaqoob M Imran R U Rasool andW Dou ldquoComplementing iot services through software de-fined networking and edge computing a comprehensivesurveyrdquo IEEE Communications Surveys amp Tutorials vol 22no 3 pp 1761ndash1804 2020

[34] L Yang H Yao J Wang C Jiang A Benslimane and Y LiuldquoMulti-UAV-enabled load-balance mobile-edge computingfor iot networksrdquo IEEE Internet ofings Journal vol 7 no 8pp 6898ndash6908 2020

[35] C Gong F Lin X Gong and Y Lu ldquoIntelligent cooperativeedge computing in internet of thingsrdquo IEEE Internet of ingsJournal vol 7 no 10 pp 9372ndash9382 2020

[36] B C Kavitha R Vallikannu and K S Sankaran ldquoDelay-aware concurrent data management method for iot collab-orative mobile edge computing environmentrdquo Microproces-sors and Microsystems vol 74 Article ID 103021 2020

14 Journal of Healthcare Engineering

Provincial Natural Science Foundation of China under grant2020CFB629 the Fundamental Research Funds for theCentral Universities under grants CZT20027 and CSQ21018and the Research Start-Up Funds of South-Central Uni-versity for Nationalities under grant YZZ18006

References

[1] M Bhatia ldquoIot-inspired framework for athlete performanceassessment in smart sport industryrdquo IEEE Internet of ingsJournal vol 8 no 12 pp 9523ndash9530 2020

[2] Q Zhang Y Zhang C Li C Yan Y Duan and H WangldquoSport location-based user clustering with privacy-preserva-tion in wireless iot-driven healthcarerdquo IEEE Access vol 9pp 12906ndash12913 2021

[3] A Rajsp and I Fister ldquoA systematic literature review of in-telligent data analysis methods for smart sport trainingrdquoApplied Sciences vol 10 no 9 p 3013 2020

[4] J Haladjian D Schlabbers S Taheri M arr andB Bruegge ldquoSensor-based detection and classification ofsoccer goalkeeper training exercisesrdquo ACM Transactions onInternet Technology vol 1 no 2 pp 1ndash20 2020

[5] X Bu ldquoHumanmotion gesture recognition algorithm in videobased on convolutional neural features of training imagesrdquoIEEE Access vol 8 pp 160 025ndash160 039 2020

[6] Y Wang M Chen X Wang R H M Chan andW J Li ldquoIotfor next-generation racket sports trainingrdquo IEEE Internet ofings Journal vol 5 no 6 pp 4558ndash4566 2018

[7] Y Wang Y Zhao R H M Chan and W J Li ldquoVolleyballskill assessment using a single wearable micro inertial mea-surement unit at wristrdquo IEEE Access vol 6pp 13 758ndash813 765 2018

[8] N F Ghazali N Shahar N A Rahmad N A J SufriM A Asrsquoari and H F M Latif ldquoCommon sport activityrecognition using inertial sensorrdquo in Proceedings of the IEEE14th International Colloquium on Signal Processing Its Ap-plications (CSPA) pp 67ndash71 Malaysia March 2018

[9] M Ramanathan W-Y Yau and E K Teoh ldquoHuman actionrecognition with video data research and evaluation chal-lengesrdquo IEEE Transactions on Human-Machine Systemsvol 44 no 5 pp 650ndash663 2014

[10] J K Aggarwal and M S Ryoo ldquoHuman activity analysisrdquoACM Computing Surveys vol 43 no 3 pp 1ndash43 2011

[11] M Stikic D Larlus S Ebert and B Schiele ldquoWeakly su-pervised recognition of daily life activities with wearablesensorsrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 33 no 12 pp 2521ndash2537 2011

[12] J Wang Y Chen S Hao X Peng and L Hu ldquoDeep learningfor sensor-based activity recognition a surveyrdquo PatternRecognition Letters vol 119 pp 3ndash11 2019

[13] KWang J He and L Zhang ldquoAttention-based convolutionalneural network for weakly labeled human activitiesrsquo recog-nition with wearable sensorsrdquo IEEE Sensors Journal vol 19no 17 pp 7598ndash7604 2019

[14] C Chen R Jafari and N Kehtarnavaz ldquoImproving humanaction recognition using fusion of depth camera and inertialsensorsrdquo IEEE Transactions on Human-Machine Systemsvol 45 no 1 pp 51ndash61 2015

[15] K Kui Liu C Chen Chen R Jafari and N KehtarnavazldquoFusion of inertial and depth sensor data for robust handgesture recognitionrdquo IEEE Sensors Journal vol 14 no 6pp 1898ndash1903 2014

[16] A Mannini S S Intille M Rosenberger A M Sabatini andW Haskell ldquoActivity recognition using a single accelerometer

placed at the wrist or anklerdquo Medicine amp Science in Sports ampExercise vol 45 no 11 pp 2193ndash2203 2013

[17] A Mannini M Rosenberger W L Haskell A M Sabatiniand S S Intille ldquoActivity recognition in youth using singleaccelerometer placed at wrist or anklerdquoMedicine amp Science inSports amp Exercise vol 49 no 4 pp 801ndash812 2017

[18] M Shoaib S Bosch O Incel H Scholten and P HavingaldquoComplex human activity recognition using smartphone andwrist-worn motion sensorsrdquo Sensors vol 16 no 4 p 4262016

[19] J P Varkey D Pompili and T A Walls ldquoHuman motionrecognition using a wireless sensor-based wearable systemrdquoPersonal And Ubiquitous Computing vol 16 pp 897ndash9102016

[20] J S Akins N R Heebner M Lovalekar and T C SellldquoReliability and validity of instrumented soccer equipmentrdquoJournal of Applied Biomechanics vol 31 no 3 pp 195ndash2012015

[21] A Meamarbashi and S Hossaini ldquoApplication of novel in-ertial technique to compare the kinematics and kinetics of thelegs in the soccer instep kickrdquo Journal of Human Kineticsvol 23 no 1 pp 5ndash13 2010

[22] BMI160 Datasheet Document Bst-Bmi160-Ds000-07 BoschSensortec Reutlingen Germany 2015

[23] G Okeyo L Chen HWang and R Sterritt ldquoDynamic sensordata segmentation for real-time knowledge-driven activityrecognitionrdquo Pervasive and Mobile Computing vol 10pp 155ndash172 2014

[24] S Ioannidou and G Pantazis ldquoHelmert transformationproblem from euler angles method to quaternion algebrardquoISPRS International Journal of Geo-Information vol 9 no 9p 494 2020

[25] N Jouppi C Young N Patil and D Patterson ldquoMotivationfor and evaluation of the first tensor processing unitrdquo IEEEMicro vol 38 no 3 pp 10ndash19 2018

[26] H Ranocha M Sayyari L Dalcin M Parsani andD I Ketcheson ldquoRelaxation Rkutta methods fully discreteexplicit entropy-stable schemes for the compressible euler andNavierndashStokes equationsrdquo SIAM Journal on Scientific Com-puting vol 42 no 2 pp A612ndashA638 2020

[27] A M Andrew An Introduction to Support Vector Machinesand Other Kernel-Based Learning Methods by Nello Chris-tianini and John Shawe-Taylor Cambridge University PressCambridge UK 2000

[28] X Yang and Y L Tian ldquoEigen joints-based action recognitionusing Naıve-Bayes-nearest-neighborrdquo in Proceedings of theIEEE Computer Society Conference on Computer Vision andPattern Recognition Workshops pp 14ndash19 Providence RIUSA June 2012

[29] T V Gestel B Baesens P V Dijcke J A K Suykens andT Alderweireld ldquoLinear and nonlinear credit scoring bycombining logistic regression and support vector machinesrdquoSocial Science Electronic Publishing vol 1 no 4 2005

[30] A Jindal A Dua K Kaur M Singh N Kumar andS Mishra ldquoDecision tree and SVM-based data analytics fortheft detection in smart gridrdquo IEEE Transactions on IndustrialInformatics vol 12 no 3 pp 1005ndash1016 2016

[31] P anh Noi and M Kappas ldquoComparison of random forestk-nearest neighbor and support vector machine classifiers forland cover classification using sentinel-2 imageryrdquo Sensorsvol 18 no 1 p 18 2018

[32] H Li K Ota and M Dong ldquoLearning iot in edge deeplearning for the internet of things with edge computingrdquo IEEEnetwork vol 32 no 1 pp 96ndash101 2018

Journal of Healthcare Engineering 13

[33] W Rafique L Qi I Yaqoob M Imran R U Rasool andW Dou ldquoComplementing iot services through software de-fined networking and edge computing a comprehensivesurveyrdquo IEEE Communications Surveys amp Tutorials vol 22no 3 pp 1761ndash1804 2020

[34] L Yang H Yao J Wang C Jiang A Benslimane and Y LiuldquoMulti-UAV-enabled load-balance mobile-edge computingfor iot networksrdquo IEEE Internet ofings Journal vol 7 no 8pp 6898ndash6908 2020

[35] C Gong F Lin X Gong and Y Lu ldquoIntelligent cooperativeedge computing in internet of thingsrdquo IEEE Internet of ingsJournal vol 7 no 10 pp 9372ndash9382 2020

[36] B C Kavitha R Vallikannu and K S Sankaran ldquoDelay-aware concurrent data management method for iot collab-orative mobile edge computing environmentrdquo Microproces-sors and Microsystems vol 74 Article ID 103021 2020

14 Journal of Healthcare Engineering

[33] W Rafique L Qi I Yaqoob M Imran R U Rasool andW Dou ldquoComplementing iot services through software de-fined networking and edge computing a comprehensivesurveyrdquo IEEE Communications Surveys amp Tutorials vol 22no 3 pp 1761ndash1804 2020

[34] L Yang H Yao J Wang C Jiang A Benslimane and Y LiuldquoMulti-UAV-enabled load-balance mobile-edge computingfor iot networksrdquo IEEE Internet ofings Journal vol 7 no 8pp 6898ndash6908 2020

[35] C Gong F Lin X Gong and Y Lu ldquoIntelligent cooperativeedge computing in internet of thingsrdquo IEEE Internet of ingsJournal vol 7 no 10 pp 9372ndash9382 2020

[36] B C Kavitha R Vallikannu and K S Sankaran ldquoDelay-aware concurrent data management method for iot collab-orative mobile edge computing environmentrdquo Microproces-sors and Microsystems vol 74 Article ID 103021 2020

14 Journal of Healthcare Engineering