VisualInformationFeaturesandMachineLearningforWushu … · 2021. 8. 5. · Received 3 June 2021;...

Research ArticleVisual Information Features and Machine Learning for WushuArts Tracking

Jing Li1 and Guangren Zhou 2

1Sports Center Xirsquoan Jiaotong University Shaanxi 710049 Xirsquoan China2School of Comprehensive Foundation Wannan Medical College Wuhu 241002 Anhui China

Correspondence should be addressed to Guangren Zhou zhouguangren2021163com

Received 3 June 2021 Revised 4 July 2021 Accepted 19 July 2021 Published 5 August 2021

Academic Editor Fazlullah Khan

Copyright copy 2021 Jing Li and Guangren Zhou+is is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

Martial arts tracking is an important research topic in computer vision and artificial intelligence It has extensive and vitalapplications in video monitoring interactive animation and 3D simulation motion capture and advanced human-computerinteraction However due to the change of martial artsrsquo body posture clothing variability and light mixing the appearancechanges significantly As a result accurate posture tracking becomes a complicated problem A solution to this complicatedproblem is studied in this paper+e proposed solution improves the accuracy of martial arts tracking by the image representationmethod of martial arts tracking +is method is based on the second-generation strip wave transform and applies it to the videomartial arts tracking based on the machine learning method

1 Introduction

+e extraction of visual image information features is acrucial problem of computer vision and intelligent imageprocessing It is also an essential technology which hasreceived extensive attention in the past 20 years It mainlyrefers to using a computer algorithm to extract the repre-sentative image information in the image to determinewhether a point is a unique factor for identification +estandard image features can be divided into local featuresand global features from the representation range of theirfeatures +e content of image features can be divided intothe corner edge contour histogram region etc Dis-tinguishing particular features is the most fundamental basisof computer vision and image information understandingIts basic meaning is generating a dimension vector that canreflect the essential characteristics of the recognized patternaccording to the input system information +erefore theselection of features has become the most fundamental basisfor computer judgment +e most crucial feature of featureextraction is ldquorepeatabilityrdquo the features extracted from

different images of the same scene should be the same inwhich the computer can repeatedly select For examplefeatures make it possible to find similar martial arts struc-tures in multiple motion images +ese can be used as themodel input for further selection and processing In essencethe martial arts tracking method based on machine learningis to extract appropriate features and add appropriate ma-chine learning algorithms +e quality of feature extractioncan directly affect the classifierrsquos performance and the finaldetection result [1 2]

+e machine learning method can be divided intoclassification and regression according to the data categoryand discrete degree Classification can be seen as finding alabel that belongs to a particular class in a discrete categoryfor given data Generally it can be described as follows atraining sample set is known which is the feature set of asample and usually expressed in the form of a vector Eachelement of the vector is a description of an inevitable featureof the sample and is a label of the samplersquos categoryWhat wewant to establish is a classification rule For any unknownsample we can apply this rule to its eigenvector to determine

HindawiJournal of Healthcare EngineeringVolume 2021 Article ID 6713062 6 pageshttpsdoiorg10115520216713062

the sample category Regression is to make accurate pre-dictions or estimate the labelrsquos continuous real value cor-responding to the data which is the specific real value [3]

For continuous and comprehensive trend estimation inthe real number space the corresponding label in thisproblem is the continuous space such as the posture co-ordinate value of three-dimensional joint points in eachimage of martial arts +e machine learning method ofWushu motion tracking is to estimate the value of the 3Dposture corresponding to the image space +e standardmodels of discriminant tracking methods include para-metric methods and nonparametric methods such as NNKKr local GP shared Golem and skill +e data association isshown in Figure 1 which is the low-dimensional popular orimplicit variable space of data features [4]

However generally speaking many machine learningmethods are derived from some basic ideas+emost typicaland representative is the Gaussian process [5]

2 Machine Learning Methods in Wushu Arts

A Gaussian regression process is a set of random variablesand any numbers of its finite subsets are subject to the jointGaussian distribution+e Gaussian process has been widelystudied in martial arts tracking It poses a recovery in recentyears because of its output probability distribution functioncontinuity and other characteristics However we do notknow the spatial distribution characteristics of a specificsequence of martial arts postures Any traditional datadistribution (martial arts postures are no exception) is in-finitely close to a Gaussian normal distribution under themassive statistical results Learning the Gaussian processmeans learning the superparameters of the method insteadof learning the weights of the primary functions oftencontained in traditional machine learning methods A pa-rameter edge process can eliminate the correspondingweight to reduce the extra parameters In other words thesuperparameters are learned through a maximum likelihoodfunction Here the parameters are not the standard meanand variance matrix of traditional data statistics but themean and variance functions In other words a Gaussianprocess is entirely determined by its mean function andcovariance function As long as the mean function m(x) andcovariance function k(x xprime) are determined the Gaussianprocess is entirely determined In regression estimation akernel function is selected to assume the prior distribution ofthe data A posterior part of the function is obtained bycombining the primary and training data used to estimatethe new data [6ndash8]

+e traditional machine learning methods are studied onthe premise that the number of samples is enough +eperformance of the proposed methods is theoreticallyguaranteed when the total number of samples tends toinfinity when the number of samples is enough However inmost practical cases the number of samples is limited+erefore it is difficult to achieve the desired resultsHowever the Gaussian process finds a method to predictmultiple repetitions through limited data+e predicted dataspace is nearly infinite +erefore GPR can be used to

represent the nonlinear input-output mapping such asmartial arts tracking [9] +e machine learning problem canbe expressed as there is a specific dependence between theknown variable y and the input x In other words there is anunknown joint probability distribution F +e machinelearning problem estimates the maximum posterior prob-ability according to n different samples GPR represents aBayesian function with Gaussian distribution [10] Howevergenerally speaking many machine learning methods arederived from some basic ideas +e most typical and rep-resentative is the Gaussian process

p(f|X) N(0 K) (1)

where f [f1andand fn] is a series of data spaces with functionvalue and f1 f(x1) is the feature vector of the originalimage X [X1 XN]T is a covariance matrix and K isa covariance function Kij k(xi xj) In the actual opera-tion process we have assumed that the data are Gaussiandistribution Hence a radial basis function (RBF) or aGaussian kernel function can be represented by the dataassociation For example consider the functionKR(xi xj) exp(minuscx

xi minus xj

2) + λxδij where cx 0 is the

kernel width parameter λx 0 is the noise difference cx 0is the Kronecker trigonometric function if the value is 1other cases are 0 Since the prior of the joint distribution isGaussian distribution the posterior prediction of the newdata is based on the output value of the observed samples+e mean and variance are as follows

m y(d)

1113872 1113873 Y(d)

Kminus1x K

sx

σ2 y(d)

1113872 1113873 Kx(x x) minus KxX( 1113857

TK

minus1X K

xX

(2)

From the perspective of the graph structure theGaussian process can be recognized as a potential structuralassociation between any observed data pair+e square nodeis the observed vector and the circular node is the unob-served vector Each sample pair obeys Gaussian distributionand the data are also interrelated which affects the esti-mation of other variable functions [11ndash13]

Besides any machine learning process similar to theGaussian process involves the problem of data generaliza-tion or generalization ability +is is how we use the existingdata to fit the data distribution of unknown results as muchas possible and infer the existing observations into a morecomprehensive problem space [14] +e generalizationability of the data is one of the basic requirements to testwhether a machine learning algorithm has real wide avail-ability However it is impossible to know precisely whetherthe test data are consistent with the sample space of thetraining data According to the law of large numbers and thegeneral situation of the whole data most problems can besimplified as a Gaussian distribution or a linear superpo-sition of multiple Gaussian distributions +e posteriordistribution also conforms to the Gaussian theory It can bebelieved that the Gaussian process reflects the complexcorrelation of the sample data to a certain extent [15] +ecore of this correlation is the kernel function and covariancewhich are mainly regulated by the parameters of the kernel

2 Journal of Healthcare Engineering

function and some superparameters Some parameters canbe eliminated based on marginal parameters but a super-parameter itself needs further verification +erefore inGPR the selection of parameters is a crucial problem +eparameters of different features need further cross-valida-tion to avoid overfitting and underfitting [16ndash20]

3 Wushu Tracking Mechanism

+e subject of Wushu tracking comes from the urgent re-search needs of computer vision in recent 20 years As anessential branch of computer science and artificial intelli-gence computer vision aims to use various electronic im-aging systems to replace the human eye to obtain visualperception +e computer replaces the human brain to re-alize the processing and understanding of visual information[21ndash23] In short it is to make the computer have humanvisual recognition A complete vision system usually in-volves the following contents acquisition processing rep-resentation storage and transmission First the computerequipment based on control sensing collects the originaldata+en the acquired visual data are further characterizedor compressed by the computer analyzed and processed+en the data are stored and transmitted through thenetwork to realize a series of functions of human biologicalvision Finally the computer forms a clear and meaningfuldescription of the collected image content to perceive theobjective world visually Visual information processing is thecrucial and challenging point in the field of computer vision[24ndash27]

31 Martial Arts Tracking Using the Second-Generation BandWave Transform +e traditional representation methodbased on the image edge only describes the geometriccharacteristics of the image through the edge It is not onlynot strict but also tricky to describe the image well whichhinders further effective feature representation and ad-vanced computer vision processing +erefore Mallat in-troduced geometric flow to describe the geometriccharacteristics of images Based on the first-generationbandelet they proposed the second-generation strip wavetransform Based on the geometric flow of image charac-teristics a new image feature extraction method in martialarts tracking based on the second-generation strip wavetransform is proposed in this paper+is method extracts thetop feature of geometric flow in the region direction rep-resenting the main texture direction and change in theregion direction Because the representation of this method

is sparse and scale-invariant it can be used for illuminationchange +e main image also has good robustness (based onthe change of the gray image level rather than brightness) Itcan distinguish the difference under the apparent defor-mation of the image In this method the statistical featuresof the bandelet in the bandelet transform are used as imagefeatures +e Gaussian and double Gaussian processes arecombined to perform regression and track martial arts in theimage [28 29]

32 Geometric Flow Feature Extraction Method of the StripWave +e visual information of the image is the precon-dition of Wushu tracking Using the geometric flow featureof the strip wave to extract the image features of martial artscan accurately express the movement posture and thegeneral texture distribution in the image As the most criticalpart of bandelet geometric flow feature extraction our al-gorithm requires a full analysis of feature extraction early+e proposed algorithm uses the second-generation stripwave algorithm experiment to ensure that the specific fea-ture extraction method [30ndash33] can extract the most suitablepattern and texture information of the characters in theimage +e generated image [34] descriptor is unique andselective

33 Optimal Parameter Selection When the bandelettransform is used for image compression the primarypurpose is to reduce the number of nonzero coefficients asfar as possible +e parameters of the conventional bandelettransform are different from those of the martial artstracking which need to be determined by the parameterselection experiment In terms of parameter selection theexperiment adopts the same method as 2 comparing theROC curve after training the classifier with differenttransformation parameters to determine these parametersHere we choose the ROC curve of different detection ratesfor each possible false positive rate +e higher the ROCcurve tends to the left vertex angle the better the corre-sponding parameters are

4 Results and Discussion

In this section we described the results of the proposedscheme and explained them in detail

41 Two-Dimensional Wavelet Transforms We choose oneto five layers of-dimensional wavelet transform and do notcarry out two-dimensional wavelet transform a total of six

y

x x

x x x xy y y y yz

z z z zx zy

Regression MoE SLVM GPLVM Shared KIE Latent GMR

Figure 1 Machine learning mode

Journal of Healthcare Engineering 3

cases of experiments Among them j minus max 2j minus min 2and T 15 and the experimental results areconsistent When the wavelet level is 1-2 the ROC curve canpresent good detection results +erefore we choose anexcellent wavelet transform+e obtained image features areused for tracking with good results as shown in Figure 2

+e results obtained in this paper are consistent withthose in the literature and the best result is obtained byusing only one layer of the two-dimensional wavelettransform +e main reasons are as follows the more thedecomposition level the lower the representation ability ofthe feature of the higher layerrsquos low-frequency approxi-mation coefficient is compared to the better the high-fre-quency detail coefficient Furthermore using only one layerof the two-dimensional wavelet transform is also conduciveto selecting the scale range of features in the process oftracking the predictor regression mapping maintaining aunified quantization interval and avoiding the instabilitycaused by too extensive variation range of kernel parameters[35 36]

42 7e Scale of the Minimum Binary Partition and theMaximum Scale of Quadtree Upward In theory the smallerthe minimum partition is the larger j minus min is and the morereasonable the quadtree j minus max is Larger j minus max andsmaller j minus min will bring more time complexity to theprocess of feature extraction which is not conducive to thelearning of a vast database +e tracking error can be sta-bilized in a lower range and the time required for featureextraction is significantly reduced to demonstrate that theaverage joint error of each frame of three-dimensionalequine or human posture data on theoretical knot data ismm Obviously the lower the error the more accurate thetracking In the double Gaussian system with a neighborpruning algorithm the number of k-nearest neighbors is100 As a result a video sequence from the Humanivadatabase is selected for testing When the 4times 4 bandeletdescriptor parameters are selected the average joint error ofeach frame of the Wushu 3D pose data verified on thewalking data is mm

It should be noted that the results of this group of ex-periments are consistent and generalizable Suppose thesame feature extraction method is used on similar motiondata +e average effect of j-max 2 and j-min 2 will bebetter than that of other transform extraction features and a2-scale subdivision size is adopted It can be seen that only4times 4 size blocks are used to extract features from thebandelet transform A two-layer upward quadtree optimi-zation merging strategy is adopted It has the best repre-sentation ability and relatively low time consumption At thesame time we further use the features of large and smallblocks Although the tracking effect will be slightly affectedit can significantly reduce the dimension of the descriptor

43 Quantization 7reshold T +e purpose of determiningthe quantization threshold T is to control the quantizationrange in the process of image signal quantization +e valuewhose coefficient is less than t is set to zero thus omitting

redundant information In the image coding t is used tocontrol the compression ratio +e larger the value is thehigher the compression ratio is and the more pronouncedthe image distortion is On the contrary the selection of thequantization value affects the coefficient value more sig-nificantly than in a one-dimensional wavelet transform in acertain direction while searching for the optimal direction ofgeometric flow+erefore selecting too large or too small t isnot conducive to finding the optimal direction of geometricflow According to different application fields the processingof the T value is also different It is still necessary to find thebest t value through specific experiments When Level 1 j-max 2 j-min 2 and T15 are taken good results areobtained +e small range variation of this value has nonoticeable effect on the actual results It can be seen from theexisting literature and preliminary experiment 3 that theselection of T has little influence on the training error rateand test accuracy rate which the diversity of photos shouldproduce for the accurate extraction of martial arts imagefeatures

44 Block Size For the influence of subblock size selection onthe image signal large or small subblock partition will have adeviation effect on the actual image feature extraction results+ere is an optimal subblock size and the subblock seg-mentation is too small or too large We select 4times 4 (or 8times 8)subblock size for feature extraction and parameter selection inthe actual experiment +is choice is mainly based on the sizeof the image and the dimension of the description features

441 Strip Wave Feature Extraction Using AlgorithmOptimization +e implementation of the bandelet trans-form in the second generation of the bandelet transforminvolves a tedious sorting operationWe need to improve thealgorithm further and reduce the sorting complexity ofdescriptor extraction In the extraction process the order of

1

095

09

Effect of level

085

08

Det

ectio

n ra

te

075

07

065

06

055

0 002 004 006 008 01False positive rate

012 014 016 018 0205

I = 0I = 1I = 2

I = 3I = 4

Figure 2 Performance comparison of different layers of thewavelet transform


wavelet coefficients will be consistent for geometric flowblocks with the same scale and order +erefore the sortingindex can be established in advance according to all possiblesizes such as 4times 4 and 8times 8 +e strip wave blocks geo-metric flow direction which eliminates a considerablenumber of repeated sorting procedures We use a similaroptimization algorithm

Two sort indexes are created

(i) For each possible direction the reordering index ofthe whole two-dimensional wavelet transform co-efficient matrix is established and the two-dimen-sional wavelet transform coefficient matrix isreordered into a one-dimensional vector

(ii) +e second index is set up to rearrange the waveletcoefficients of the one-dimensional vector after theone-dimensional wavelet transform is applied toeach strip wave block

+en the reordered one-dimensional vector is seg-mented (equivalent to the original two-dimensional matrixwhich is divided into blocks) +e Lagrange function valuesin each direction are obtained Finally the direction of theminimum Lagrange function value corresponding to eachvector segment is the best geometric flow direction of thecorresponding block +e strip wave coefficients are ob-tained Using this optimization in the actual experimenteach martial arts imagersquos feature extraction time (the size is192times 64 pixels) is 0138 seconds Compared with the original14 seconds the time consumption is significantly reducedIt is close to the HOG feature extraction time of each sample(012 seconds) +e reduction of time consumption mainlydepends on transforming a one-dimensional wavelettransform into a simple one-dimensional matrix +en thewhole process only needs to implement a one-dimensionalwavelet transform

5 Conclusion

A new method for the feature extraction and detection ofmartial arts is proposed based on the second-generationstrip wave transform To carry out learning information andrecover the three-dimensional posture of martial arts in theimage statistical approaches in band wave transform asimage descriptors are applied Firstly the optimization al-gorithm based on the original second-generation strip waveis used to improve the operation speed +en the relevantoptimal parameters are established through experimentsSome statistical features are selected through the featureselection experiment and feature combination hoof Finallythe maximum value of geometric flow is determined as aneffective global feature representation Different block sizesare used to reduce the dimension of features to furtherreduce the complexity of feature vectors +en the featureextraction method is used to extract the features of thetraining samples +e Gaussian process algorithm is used totrain the predictor +e test image is tested on the databaseusing the obtained predictor model All the results arecompared with feature extraction methods From the resultsit can be found that the maximum geometric flow feature

can effectively represent the posture of martial arts +eimage description ability of simple and basic motion se-quences is better than that of the classical global imagefeatures Different learning methods can obtain bettertracking results and lower tracking errors On the wholefrom the test results of standard deviation we can see thatthe tracking results of the data are relatively stable by usingthe maximum value feature of the strip wave +ey havegood adaptability and robustness in continuous imagetracking with slight fluctuation which is more suitable forthe description of martial arts images

Data Availability

+e data used to support the findings of this study are in-cluded within the article

Conflicts of Interest

All the authors declare no conflicts of interest

Acknowledgments

+is study was supported by Research on Health PromotionMode of Sports and Medical Integration in Urban Com-munities of Anhui Province under the Background ofldquoHealthy Chinardquo (SK 2020A0378)

References

[1] X Lili and Z Yuan ldquoDesign and implementation of remotesensing monitoring system for resources and environmentbased on lidar technologyrdquo J Lasers vol 41 no 8 pp 54ndash582020

[2] L Chang Z Huixin P Qingqing and M Fanyi ldquoDesign ofhigh speed optical fiber video image transmission modulebased on embedded systemrdquo Electron Devices vol 43 no 4pp 882ndash887 2020

[3] B Hu Bin ldquoDesign and application of automatic test systemfor optical transmission equipmentrdquo Decision Exploration(middle) vol 2020 no 8 52 pages 2020

[4] H Hu and W Bo ldquoDesign of 10 GBs SFP+ optical modulebased on CWDMrdquo Communication Technology vol 53 no 8pp 2064ndash2069 2020

[5] T T Shih P H Tseng Y Y Lai and W H Cheng ldquoA 25Gbits transmitter optical sub-assembly package employingcost-effective TO-CAN materials and processesrdquo Journal ofLightwave Technology vol 30 no 6 pp 834ndash840 2011

[6] M Q Tian W Wang J C Song Y Song L Yan and Y XialdquoA dynamic load identification method for rock roadheadersbased on wavelet packet and neural networkrdquo in Proceedingsof the 2015 IEEE 10th Conference on Industrial Electronics andApplications (ICIEA) pp 666ndash670 IEEE Auckland NewZealand June 2015

[7] S Geethalakshmi S Narendran S Ramalingam andN Pappa ldquoOptimization of fed-batch process for recombi-nant protein production in Escherichia coli using geneticalgorithmrdquo in Proceedings of the 2011 International Confer-ence on Process Automation Control and Computing pp 1ndash5IEEE Coimbatore Tamilnadu July 2011

[8] C C Tsai K I Tsai and C T Su ldquoCascaded fuzzy-PIDcontrol using PSO-EP algorithm for air source heat pumpsrdquo


in Proceedings of the 2012 International conference on Fuzzy7eory and Its Applications (iFUZZY2012) pp 163ndash168 IEEETaichung Taiwan November 2012

[9] T M Takala Y Hirao H Morikawa and T Kawai ldquoMartialarts training in virtual reality with full-body tracking andphysically simulated opponentsrdquo in Proceedings of the 2020IEEE Conference on Virtual Reality and 3D User InterfacesAbstracts and Workshops (VRW) p 858 IEEE Atlanta GAUSA March 2020

[10] A K Banerjee and G K Bhattacharyya ldquoBayesian results forthe inverse Gaussian distribution with an applicationrdquoTechnometrics vol 21 no 2 pp 247ndash251 1979

[11] Y Yu B Liu and Z Chen ldquoAnalyzing the performance ofpseudo-random single photon counting ranging lidarrdquo Ap-plied Optics vol 57 no 27 pp 7733ndash7739 2018

[12] N O Sokal and A D Sokal ldquoClass E-A new class of high-efficiency tuned single-ended switching power amplifiersrdquoIEEE Journal of Solid-State Circuits vol 10 no 3 pp 168ndash1761975

[13] B H Tang and Z X Zhou ldquo+e design of communicationnetwork optical fiber cable condition monitoring systembased on distributed optical fiber sensorrdquo in Proceedings of the2018 International Conference on Electronics Technology(ICET) pp 97ndash101 IEEE Chengdu China May 2018

[14] K Diethelm N J Ford and A D Freed ldquoA predictor-cor-rector approach for the numerical solution of fractionaldifferential equationsrdquo Nonlinear Dynamics vol 29 no 1pp 3ndash22 2002

[15] C B Zhao M J Wang and X W Wang ldquoSelf-optimizationcontrol in combustion system using genetic algorithmsrdquo CoalMine Machinery vol 29 no 7 pp 165ndash167 2008

[16] Y Zhang Y Li Y Liu and G Yi ldquoControl of cricket systemusing LQR controller optimized by particle swarm optimi-zationrdquo Journal of Physics Conference Series vol 1670 no 1Article ID 012016 2020

[17] Z Peng ldquoPID control of temperature and humidity in granarybased on improved genetic algorithmrdquo in Proceedings of the2019 IEEE International Conference on Power IntelligentComputing and Systems (ICPICS) pp 428ndash432 IEEE She-nyang China July 2019

[18] E Wu and H Koike ldquoFuturepose-mixed reality martial artstraining using real-time 3d human pose forecasting with a rgbcamerardquo in Proceedings of the 2019 IEEE Winter Conferenceon Applications of Computer Vision (WACV) pp 1384ndash1392IEEE Waikoloa Village HI USA January 2019

[19] L M Zhan B Liu L Fan J Chen and X M Wu ldquoMedicalvisual question answering via conditional reasoningrdquo inProceedings of the 28th ACM International Conference onMultimedia pp 2345ndash2354 New York NY USA October2020

[20] F Xiao B Liu and R Li ldquoPedestrian object detection withfusion of visual attention mechanism and semantic compu-tationrdquo Multimedia Tools and Applications vol 79 no 21pp 14593ndash14607 2020

[21] J Zhang Y Liu H Liu and J Wang ldquoLearning local-globalmultiple correlation filters for robust visual tracking with kfilter redetectionrdquo Sensors vol 21 no 4 Article ID 1129 2021

[22] Y Gu A Chen X Zhang C Fan K Li and J Shen ldquoDeeplearning based cell classification in imaging flow cytometerrdquoASP Transactions on Pattern Recognition and IntelligentSystems vol 1 no 2 pp 18ndash27 2021

[23] J Zhang J Sun J Wang and X G Yue ldquoVisual objecttracking based on residual network and cascaded correlation

filtersrdquo Journal of Ambient Intelligence and HumanizedComputing vol 12 pp 1ndash14 2020

[24] J Zhang W Wang C Lu J Wang and A K SangaiahldquoLightweight deep network for traffic sign classificationrdquoAnnals of Telecommunications vol 75 no 7 pp 369ndash3792020

[25] R Liu X Ning W Cai and G Li ldquoMultiscale dense cross-attention mechanism with covariance pooling for hyper-spectral image scene classificationrdquo Mobile Information Sys-tems vol 2021 Article ID 9962057 2021

[26] Y Ding X Zhao Z Zhang W Cai and N Yang ldquoMultiscalegraph sample and aggregate network with context-awarelearning for hyperspectral image classificationrdquo IEEE Journalof Selected Topics in Applied Earth Observations and RemoteSensing vol 14 pp 4561ndash4572 2021

[27] W Cai and Z Wei ldquoRemote sensing image classificationbased on a cross-attention mechanism and graph convolu-tionrdquo IEEE Geoscience and Remote Sensing Lettersvol 20205 pages In Press Article ID 3026587 2020

[28] L I Huajie and J Wu ldquo+e housing price forecasts of xia menbased on BP neural networkrdquo Journal of Shanxi Radio amp TVUniversity vol 1 pp 102ndash104 2011

[29] Y Li ldquoResearch on house price forecast based on grey systemGM (1 1)rdquo in Proceedings of the 2019 5th InternationalConference on Finance Investment and Law (ICFIL 2019)Colombo Sri Lanka October 2019

[30] Z Wang P Zhang W Sun and D Li ldquoApplication of datadimension reduction method in high-dimensional data basedon single-cell 3D genomic contact datardquo ASP Transactions onComputers vol 1 no 2 pp 1ndash6 2021

[31] W Sun P Zhang Z Wang and D Li ldquoPrediction of car-diovascular diseases based on machine learningrdquo ASPTransactions on Internet of 7ings vol 1 no 1 pp 30ndash352021

[32] Z Huang P Zhang R Liu and D Li ldquoImmature appledetection method based on improved Yolov3rdquo ASP Trans-actions on Internet of 7ings vol 1 no 1 pp 9ndash13 2021

[33] X Ning K Gong W Li L Zhang X Bai and S TianldquoFeature refinement and filter network for person re-identi-ficationrdquo IEEE Transactions on Circuits and Systems for VideoTechnology vol 99 p 1 2020

[34] X Ning D Gou X Dong W Tian L Yu and C WangldquoConditional generative adversarial networks based on theprinciple of homologycontinuity for face agingrdquo Concurrencyand Computation Practice and Experience 2020 In pressArticle ID e5792

[35] X Zhang Y Yang Z Li X Ning Y Qin and W Cai ldquoAnimproved encoder-decoder network based on strip poolmethod applied to segmentation of farmland vacancy fieldrdquoEntropy vol 23 no 4 p 435 2021

[36] M Fan and Y Li ldquo+e application of computer graphicsprocessing in visual communication designrdquo Journal of In-telligent and Fuzzy Systems vol 39 no 8 pp 1ndash9 2020Preprint


the sample category Regression is to make accurate pre-dictions or estimate the labelrsquos continuous real value cor-responding to the data which is the specific real value [3]

For continuous and comprehensive trend estimation inthe real number space the corresponding label in thisproblem is the continuous space such as the posture co-ordinate value of three-dimensional joint points in eachimage of martial arts +e machine learning method ofWushu motion tracking is to estimate the value of the 3Dposture corresponding to the image space +e standardmodels of discriminant tracking methods include para-metric methods and nonparametric methods such as NNKKr local GP shared Golem and skill +e data association isshown in Figure 1 which is the low-dimensional popular orimplicit variable space of data features [4]

However generally speaking many machine learningmethods are derived from some basic ideas+emost typicaland representative is the Gaussian process [5]

2 Machine Learning Methods in Wushu Arts

A Gaussian regression process is a set of random variablesand any numbers of its finite subsets are subject to the jointGaussian distribution+e Gaussian process has been widelystudied in martial arts tracking It poses a recovery in recentyears because of its output probability distribution functioncontinuity and other characteristics However we do notknow the spatial distribution characteristics of a specificsequence of martial arts postures Any traditional datadistribution (martial arts postures are no exception) is in-finitely close to a Gaussian normal distribution under themassive statistical results Learning the Gaussian processmeans learning the superparameters of the method insteadof learning the weights of the primary functions oftencontained in traditional machine learning methods A pa-rameter edge process can eliminate the correspondingweight to reduce the extra parameters In other words thesuperparameters are learned through a maximum likelihoodfunction Here the parameters are not the standard meanand variance matrix of traditional data statistics but themean and variance functions In other words a Gaussianprocess is entirely determined by its mean function andcovariance function As long as the mean function m(x) andcovariance function k(x xprime) are determined the Gaussianprocess is entirely determined In regression estimation akernel function is selected to assume the prior distribution ofthe data A posterior part of the function is obtained bycombining the primary and training data used to estimatethe new data [6ndash8]

+e traditional machine learning methods are studied onthe premise that the number of samples is enough +eperformance of the proposed methods is theoreticallyguaranteed when the total number of samples tends toinfinity when the number of samples is enough However inmost practical cases the number of samples is limited+erefore it is difficult to achieve the desired resultsHowever the Gaussian process finds a method to predictmultiple repetitions through limited data+e predicted dataspace is nearly infinite +erefore GPR can be used to

represent the nonlinear input-output mapping such asmartial arts tracking [9] +e machine learning problem canbe expressed as there is a specific dependence between theknown variable y and the input x In other words there is anunknown joint probability distribution F +e machinelearning problem estimates the maximum posterior prob-ability according to n different samples GPR represents aBayesian function with Gaussian distribution [10] Howevergenerally speaking many machine learning methods arederived from some basic ideas +e most typical and rep-resentative is the Gaussian process

p(f|X) N(0 K) (1)

where f [f1andand fn] is a series of data spaces with functionvalue and f1 f(x1) is the feature vector of the originalimage X [X1 XN]T is a covariance matrix and K isa covariance function Kij k(xi xj) In the actual opera-tion process we have assumed that the data are Gaussiandistribution Hence a radial basis function (RBF) or aGaussian kernel function can be represented by the dataassociation For example consider the functionKR(xi xj) exp(minuscx

xi minus xj

2) + λxδij where cx 0 is the

kernel width parameter λx 0 is the noise difference cx 0is the Kronecker trigonometric function if the value is 1other cases are 0 Since the prior of the joint distribution isGaussian distribution the posterior prediction of the newdata is based on the output value of the observed samples+e mean and variance are as follows

m y(d)

1113872 1113873 Y(d)

Kminus1x K

sx

σ2 y(d)

1113872 1113873 Kx(x x) minus KxX( 1113857

TK

minus1X K

xX

(2)

From the perspective of the graph structure theGaussian process can be recognized as a potential structuralassociation between any observed data pair+e square nodeis the observed vector and the circular node is the unob-served vector Each sample pair obeys Gaussian distributionand the data are also interrelated which affects the esti-mation of other variable functions [11ndash13]

Besides any machine learning process similar to theGaussian process involves the problem of data generaliza-tion or generalization ability +is is how we use the existingdata to fit the data distribution of unknown results as muchas possible and infer the existing observations into a morecomprehensive problem space [14] +e generalizationability of the data is one of the basic requirements to testwhether a machine learning algorithm has real wide avail-ability However it is impossible to know precisely whetherthe test data are consistent with the sample space of thetraining data According to the law of large numbers and thegeneral situation of the whole data most problems can besimplified as a Gaussian distribution or a linear superpo-sition of multiple Gaussian distributions +e posteriordistribution also conforms to the Gaussian theory It can bebelieved that the Gaussian process reflects the complexcorrelation of the sample data to a certain extent [15] +ecore of this correlation is the kernel function and covariancewhich are mainly regulated by the parameters of the kernel












y

x x

x x x xy y y y yz

z z z zx zy












1

095

09

Effect of level

085

08

Det

ectio

n ra

te

075

07

065

06

055


012 014 016 018 0205

I = 0I = 1I = 2

I = 3I = 4








5 Conclusion



Data Availability




Acknowledgments


References



















































y

x x

x x x xy y y y yz

z z z zx zy












1

095

09

Effect of level

085

08

Det

ectio

n ra

te

075

07

065

06

055


012 014 016 018 0205

I = 0I = 1I = 2

I = 3I = 4








5 Conclusion



Data Availability




Acknowledgments


References

















































1

095

09

Effect of level

085

08

Det

ectio

n ra

te

075

07

065

06

055


012 014 016 018 0205

I = 0I = 1I = 2

I = 3I = 4








5 Conclusion



Data Availability




Acknowledgments


References














































5 Conclusion



Data Availability




Acknowledgments


References









































VisualInformationFeaturesandMachineLearningforWushu … · 2021. 8. 5. · Received 3 June 2021;...

Documents

Transcript of VisualInformationFeaturesandMachineLearningforWushu … · 2021. 8. 5. · Received 3 June 2021;...