Ngo Duy Kien _Thesis Version 10_29_Final

download Ngo Duy Kien _Thesis Version 10_29_Final

of 59

Transcript of Ngo Duy Kien _Thesis Version 10_29_Final

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    1/59

    Error! Reference source not found.VIETNAM NATIONAL UNIVERSITY, HA NOI

    UNIVERSITY OF ENGINEERING AND TECHNOLOGY--------

    Ngo Duy Kien

    FALL DETECTION BASED ON

    ACCELEROMETER SENSOR

    Major: Computer Science

    Ha Noi 2012

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    2/59

    VIETNAM NATIONAL UNIVERSITY, HA NOI

    UNIVERSITY OF ENGINEERING AND TECHNOLOGY

    --------

    Ngo Duy Kien

    FALL DETECTION BASED ON

    ACCELEROMETER SENSOR

    Major: Computer Science

    Supervisor:Assoc. Prof. Bui The Duy

    Co-Supervisor:Dr. Vu Thi Hong Nhan

    Ha Noi 2012

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    3/59

    AUTHORSHIP

    I hereby declare that the work contained in this thesis is of my own and has not been

    previously submitted for a degree or diploma at this or any other higher education institution.

    To the best of my knowledge and belief, the thesis contains no materials previously published

    or written by another person except where due reference or acknowledgement is made.

    Signature:

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    4/59

    iv

    SUPERVISORS APPROVAL

    I hereby approve that the thesis in its current form is ready for committee examination as a

    requirement for the Bachelor of Computer Science degree at the University of Engineering

    and Technology.

    Signature:

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    5/59

    v

    ACKNOWLEDGEMENT

    First of all, I wish to express my respect and my deepest thanks to my advisers

    Assoc.Prof. Bui The Duy and Dr. Vu Thi Hong Nhan, University of Engineering and

    Technology, Viet Nam National University, Ha Noi for their enthusiastic guidance, warm

    encouragement and useful research experiences.

    I would like to gratefully thank all the teachers of University of Engineering and

    Technology, VNU for their invaluable knowledge which they gave to me during four

    academic years.

    I would also like to say my special thanks to my friends in K53CA class, University

    of Engineering and Technology, VNU, especially Tran Nguyen Le in the Behavior

    Recognition group for their helpful discussions.

    Last, but not least, my family is really the biggest motivation behind me. My parents

    and my brother always encourage me when I have stress and difficulty. I would like to

    send them my gratefulness and love.

    Ha Noi, May 24th

    , 2012Ngo Duy Kien

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    6/59

    vi

    FALL DETECTION BASED ON ACCELEROMETER SENSOR

    Ngo Duy Kien

    Course: QH-2008-I/CQ, Computer Science

    Abstract:

    The percentage of elderly population is rapidly growing in recent years. Fall-relatedinjuries are a main issue for this population. Falls account for approximately half of all

    admissions- related injuries. Falls cause not only broken bones and other health injuries, but also

    cause psychological trauma which can reduce the independence and confidence in

    communication. Therefore, techniques for detecting and tracking down the movement are

    instrumental in dealing with this problem.

    Elderly people often want to live at home, therefore, new technologies need to be able to

    support automatic fall detection, which guarantee their living independence and security. Thisthesis meets the challenge of different types of motions as part of a system designed to fulfill the

    demand for wearable device to collect data for fall and non-fall analysis. In this thesis, we present

    a machine learning based approach for this problem which takes the data from accelerometers as

    input. First, we have built a database of fall data from real people. We then propose a method to

    extract features from raw accelerometer data, which can be used to differentiate between fall and

    non-fall actions. Finally, a machine learning based classifier is used to detect fall events. Results

    show that falls can be distinguished from non-fall with 94% accuracy, for a total data set of 300

    movements.

    Keyword: fall detection, accelerometer, activity of daily living

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    7/59

    vii

    Table of Contents

    List of Figures ........................................................................................................... ixList of Table ............................................................................................................... xChapter 1Introduction ............................................................................................ 1Chapter 2Related Work .......................................................................................... 3

    2.1. Fall Detection using 3D Head Trajectory Extracted from a Signal CameraVideo Sequence ...................................................................................................... 42.2. Fall Detection using accelerometer sensor ........................................................ 6

    2.2.1 Analytical and Threshold Method ............................................................. 72.2.1.1. Fall detection algorithm using bi-axial gyroscope sensor .................... 72.2.1.2. Fall detection algorithm using tri-axial accelerometer sensor ............ 102.2.1 Learning Methods for fall detection......................................................... 13

    2.3. Summary ....................................................................................................... 14Chapter 3Fall Detection Process ........................................................................... 15

    3.1. Accelerometer Data collection ....................................................................... 163.2. Algorithm for Fall Detection .......................................................................... 18

    3.2.1. Pre-noise processing data ........................................................................ 193.2.2. Action Segmentation ............................................................................... 203.2.3. Feature Extraction ................................................................................... 21

    3.2.3.1. Length of the acceleration vector ...................................................... 213.2.3.2. The Standard Deviation .................................................................... 223.2.3.3. The speed of change in acceleration along the z-axis ........................ 233.2.3.4. The acceleration vector changes (AVC) ............................................ 233.2.3.5. The accelerometer inclination angles ................................................ 24

    3.2.4. Forming feature vector ............................................................................ 243.3. Action learning .............................................................................................. 25

    3.3.1. Support Vector Machine (SVM) .............................................................. 253.3.2. Learning .................................................................................................. 28

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    8/59

    viii

    3.4. Summary ....................................................................................................... 28Chapter 4Experimental Setup and evaluation ..................................................... 30

    4.1. Experiment Setup........................................................................................... 304.1.1. Background of fall and non-fall action .................................................... 30

    4.1.1.1. Study of fall...................................................................................... 314.1.1.2. The simulated-fall study ................................................................... 334.1.1.3. The ADL study (Activity of Daily Living) ........................................ 33

    4.1.2. Data Collection ....................................................................................... 354.2. Performance measures ................................................................................... 374.3. Experimental results of the fall detection system ............................................ 38

    Chapter 5Conclusion............................................................................................. 40Bibliography ............................................................................................................ 42Appendix A .............................................................................................................. 45

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    9/59

    List of Figures

    Figure 2-1: Reading trunk pitches rolocity V v and horizontal velocity Vh. These velocities

    were obtained without markers from the 3D trajectory of a person who brutally sits down andfalls. The different action are: the person(a) stands up,(b)sits down, (c) is seated, (d)stands upagain, (e)remains stand up and (f) falls ................................................................................... 5

    Figure 2-2: Acceleromter Sensor ............................................................................................ 6

    Figure 2-3: Reading trunk pitches and roll gyroscope ............................................................. 8

    Figure 2-4: Overlapping and non-overlapping sampe Fall and ADL data................................ 8

    Figure 2-5: Bi-axial gyroscope fall-detection algorithm flow chart ......................................... 9

    Figure 2-6: Fall detection algorithm operation example for upper and lower fall threshold,using an artificial example signal ......................................................................................... 10

    Figure 3-1: The general model cycle .................................................................................... 16

    Figure 3-2: Fall annotation for a single fall ........................................................................... 17

    Figure 3-3: Fall Detection Flow Chart .................................................................................. 18

    Figure 3-4: Low-pas and high-pass filter algorithm ............................................................. 19

    Figure 3-5: (a) Sliding window technique. (b) Overlap sliding window technique ................ 20

    Figure 3-6: RSS plot of a forward fall .................................................................................. 22

    Figure 3-7: Accelerometer inclination angle ......................................................................... 24

    Figure 3-8: Separating hyper plane ....................................................................................... 26

    Figure 4-1: The stage of fall ................................................................................................. 31

    Figure 4-2: The four phases of a fall event ........................................................................... 32

    Figure 4-3: Two types of non-fall(ADL) .............................................................................. 34

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    10/59

    x

    List of Table

    Table 3.1: Sample data acquired from acceleromter ............................................................. 17 Table 4.1: Events Sequence in the test Scenario ................................................................... 35Table 4.2: The detail of ten young volunteers ....................................................................... 36 Table 4.3: The distribution of the samples ............................................................................ 38Table 4.4: The test result ...................................................................................................... 38

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    11/59

    1

    Chapter 1

    Introduction

    In the modern society, the percentage of elderly people is increasing rapidly.

    The proportion of the population aged more than 65 years old in the developed

    countries is projected to increase from 7.5 % in 2009 to 19.6% in 2030 (United

    Nations 2009).The ratio of elderly people living alone in the house is quite high. Fall-related injury is a main issue of elderly people, both home environment and hospital

    which affects the overall quality of life either at home or hospital. Most of falls occur

    when walking, standing or sitting down, or trying to find something.Peopleusuallythink that falling recognition is the most important in our lives. Fall detection has

    become a major health concern recently. Falls account for approximately half of all

    admissions which are related to injury. Falls cause not only disabling fractures and

    other health injuries but also traumas which can reduce the independence and

    confidence when communicating. Moreover, the elder people always want to live in a

    place which ensures the independence and safety in life. Because of this, new

    technologies to guarantee and assist their lives are absolutely necessary. This is a

    factor to promote the development of technology.

    However, there are many approaches that can be used to construct a fall

    detection systems. The first one is based on image and video system and the second

    one is based on accelerometer sensor. The later is our concern in this study because

    with the rapid advances in wireless network and tiny sensors they can be embedded

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    12/59

    2

    easily in the mobile devices which can be carried anywhere. Moreover, recently,

    smart phone applications are being developed faster and faster. Therefore, with a

    phone equipped sensors, researchers have developed applications for fall detection

    based on accelerometer sensor. These applications are mainly developed based on the

    threshold algorithms, some works involve the use of machine learning techniques for

    the detection of fall and movement classification. This enables the elders status to be

    monitored anytime and anywhere and aided when necessary.

    In this thesis, we present our study on human fall detection. We present a

    machine learning based approach for this problem which takes the data from

    accelerometers as input. First, we have built a database of fall data from real people.

    We then propose a method to extract features from raw accelerometer data, which can

    be used to differentiate between fall and non-fall actions. Finally, a machine learningbased classifier is used to detect fall events.

    The remaining of the thesis is organized as follows: Chapter 2 presents a

    detailed overview of related work in the field of fall detection. Some approaches using

    different types of sensor are discussed and some higher level comparison of the result

    is presented. Chapter 3 describes the processing of the data. This chapter is divided

    into several parts. First, the raw data of the sensors is explained. Then the process of

    filtering and the process of attribute computation. Chapter 4 we give the evaluation of

    the results from different experiments performed in this research. We present about

    building a database of fall data and explain the experimental data.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    13/59

    3

    Chapter 2

    Related Work

    Accelerometer and gyroscopes are the two most popular methods in most of

    the research on fall. Threshold for acceleration, the change of velocity and angles

    were usually applied in some researches. By applying a simple threshold to the

    acceleration and using a tri-axial accelerometer worn on the chest, Lee, Nguyen, Cho

    (2009) detected falls with 98. Bourke and Lyons (2008) usedbiaxial gyroscope worn

    on the chest, they applied threshold to the peaks in the angular velocity, angular

    acceleration and the change of angle.

    In addition to applying threshold for fall detection, machine-learning algorithm

    seems to be a good method instead of threshold algorithm algorithm. One example for

    using machine learning is Shan (2010) and Yuan and Zhang (2006), both studies

    which used tri-axial accelerometer wornon the waist. They used SVM algorithm to

    classify various features from accelerometer sensor data.

    Especially, Particular interest to us is research of Lie (2009). Although their

    result of fall detection is lower than previously research, it may be due to the more

    experimental data. They apply threshold, angles and velocities. Based on this,

    potential fall and the activity will be detected after fall. Their methods sometimes do

    not lie down quickly. And these situations will be addressed by the classify

    accelerometer sensor on this thesis.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    14/59

    4

    2.1. Fall Detection using 3D Head Trajectory Extracted from aSignal Camera Video Sequence

    Falls are one of the greatest dangers for elderly people living alone. They may

    be unconscious after suffering the fall. Computer vision system provides an automatic

    solution to overcome the limitations of researchers. Some research has developed fall

    detection system by using the image sensor. One example is the work of Lee and

    Mihailidis who detect falls by using a camera mounted on the ceiling determine the

    specific silhouette [14]. This is a new method using 3D data of the head trajectory for

    monitoring the movement of the fall.

    The method of this research will be based on three steps:

    Head tracking: because of the fact that the head can be often seen in thepicture and there is a large movement during a fall

    3D tracking: the head is monitored with a particle filter to extract a 3D

    trajectory

    Fall Detecting: a fall is detected by using 3D velocities which are computed

    from the 3D trajectory of the head.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    15/59

    5

    Figure 2-1: Reading trunk pitches rolocity V v and horizontal velocity Vh. These

    velocities were obtained without markers from the 3D trajectory of a person who

    brutally sits down and falls. The different action are: the person(a) stands

    up,(b)sits down, (c) is seated, (d)stands up again, (e)remains stand up and (f) falls

    This method required a set of video in various situations such falls and normal

    situations like sitting down or squatting. Based on the characteristics of the image

    sequences, the result will be showed after analyzing with OpenCV library (Intel Open

    Source Computer Vision Library)

    The result obtained in this research is quite high. However, in some cases it

    depends on the location and specific circumstance. Camera is not always everywhere.

    Moreover, using camera usually makes the user feel unnatural and lose freedom in the

    daily activities. It affects the privacy of individuals. Consequently, using accelerationsensor for fall detection is an effective approach to ensure personal information can be

    kept secure and applied in every situation.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    16/59

    6

    2.2. Fall Detection using accelerometer sensorAccelerometer usually provides the acceleration readings in direction of x, y, z

    axis. Accelerations are these directions are presented by xA , yA , zA respectively. As

    illustrated inFigure 2-2

    Figure 2-2: Acceleromter Sensor

    X-axis has positive direction toward the right side of the device

    Y-axis has positive direction toward the top of the device

    Z-axis has positive direction toward the front of the device

    Noury and Rumeau indicate two approaches to detect fall in their research [4].

    The first place, which is the more common approach, is an analytical method and the

    second place is with machine learning techniques. An example for machine learning

    approach is the research ofMitja and Bostjan [3].This system is a visual basic system.

    By using markers, they tagged different point in body of users and used it as reference

    points. They found that the angle and the reference points between them is very

    reliable data source for extracting features. With other machine-learning algorithm,

    Support Vector machine (SVM) was the best choice for research and followed by

    Random Forest.

    In our research for fall detection is mainly based on the machine learning

    techniques and some use analytical models.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    17/59

    7

    2.2.1 Analytical and Threshold Method

    Analytical method is based on the manual or research experience about data

    sets and set-up our own parameters (thresholds). There is a need for a thorough study

    of falls and ego stage for better understanding and results. What this study focuses onis presented in this section.

    With fall detection - principles and methods [4], many falls end lying on the

    ground, the simplest approach is to detect the lying position, from a horizontal

    inclination sensor. This method is very suitable for monitoring an "isolated a person",

    but less suitable for the detection of falls of an elder person such as the irregular

    sleeping hours. Therefore this method is prone to many "false positives", i.e. detection

    as falls of situations that are not falls.In fall detection system, we can use fall velocity, body orientation, body

    posture, angular acceleration, angular velocity and fall acceleration in the critical

    phases as the determining factors for distinguishing fall and non-falls activities (ADL

    and ADLS).

    The simplest approach to detect a fall is by detecting the lying position (GPS).

    This is based on the fact that most falls, but eventually not all of them, end up falling

    in this lying positing. Both researches of StanKovic, Noury and Fleury indicate thatindirect detection of the lying posture during post-fall can be called [1], [2]. That when

    the foot is no longer in contact with the ground also can be another method to detect

    fall.

    2.2.1.1. Fall detection algorithm using bi-axial gyroscope sensorA threshold-based algorithm to distinguish between Activities of Daily Living

    and falls is described. A threshold-based fall detection algorithm using a bi-axial

    gyroscope sensor is used[13]. With trunk pitch and roll gyroscope are read during

    separate simulatedfall and Activities of Daily Living (ADL) studies, the result will be

    compared through a threshold-based algorithm

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    18/59

    8

    .

    Figure 2-3: Reading trunk pitches and roll gyroscope

    This research focus on studying about the ADL study, Data acquisition set-up,

    sensor location and signal conditioning. In signal conditioning, low-pass filter would

    be used a second-order low-pass Butterworth 2-pass digital filter, with a cut-offfrequency of 100Hz for each pitch and roll angular velocity signal. The resultant

    vector was derived by from taking the root sum of square of roll and pitch angular

    velocities.

    Fall detection will apply a threshold to the Peak value of lowest fall and highest

    ADL from the resultant angular velocity signals. The result of recording data will

    result in one of two scenarios:

    Figure 2-4: Overlapping and non-overlapping sampe Fall and ADL data

    InFigure 2-4(A) the peak values recorded ADL will not overlap with the

    recorded fall peak. Falls from ADL which may be distinguished by a

    single threshold, level of which would be placed at the lowest fall peak

    value.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    19/59

    9

    In Figure 2-4(B) the peak values recorded ADL will overlap with the

    recorded fall values. In this case applying one threshold is not sufficient

    to distinguish falls from ADL, so continuing to investigate additional

    aspects of the signals is required.

    There are three thresholds to determine for a fall which can be distinguished

    from an ADL. If the resultant angular velocity is greater than 3.1 rads/s (Fall

    threshold 1), the resultant angular acceleration is greater than 0.05 rads/s (Fall

    threshold 2) and change the result in changing trunk angle is greater than 0.59 rads

    (Fall threshold 3), a fall will be detected. The results indicate that falls can be

    distinguished from ADL with 100% accuracy for a total data set of 480 movements.

    Figure 2-5: Bi-axial gyroscope fall-detection algorithm flow chart

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    20/59

    10

    2.2.1.2.Fall detection algorithm using tri-axial accelerometer sensorMost previous researchers and others combined the devices total acceleration

    from the X, Y and Z axis [10]. This is called Root Sum of Squares or RSS which is

    presented by the formal:2 2 2

    RSS x y z (2.4)

    This is sometimes referred to as the dynamic total acceleration2

    9.81 / dRSS RSS m s

    Checking the acceleration in early part of the critical phased is the simplest

    algorithm for fall detection. When standing upright, we will have a total acceleration

    (RSS) of 1G. This value will drop to 0G in free fall until finally reaching impact.2

    1 9.81 / g m s

    This method will use a threshold which is set based on training session. There

    are two thresholds to compare with data from ADLs or other non-fall activities. There

    are called Lower Fall Peak (LFP) and Upper Fall Peak (UFP). In training period, LFP

    and UFP will give the Lower Fall Threshold (LFT) and Upper Fall Threshold (UFT),

    respectively.

    Figure 2-6: Fall detection algorithm operation example for upper and lower fall

    threshold, using an artificial example signal

    Normally, in some case, UFT usually have a better result than LFT.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    21/59

    11

    Profiling algorithm in [11] is also to measure the falling edge time and rising

    edge time. With the falling edge time ( FEt ) is the time when the RSS signal last goes

    below the LFT until it reach the UFT. The rising edge time ( REt ) is always a subset

    and smaller than the falling edge time ( FEt ).A specify example, we can imagine thatwith t0 the value of RSS first will go below the Lower Fall Threshold (LFT) and keeps

    under it until0 2 00ms

    t . The time from 0 20 0mst to 0 5 00 mst the RSS value will increase

    until the Upper Fall Threshold (UFT) reached. UFT will reach to peak in 0 5 00mst .

    0 500 0mst t is the galling edge time ( FEt ) and 0 500 0 200ms mst t is the rising edge time. In

    addition we also check for the profile LFT + the falling edge time ( FEt ) and + the rising

    edge time ( REt ).Both LFT and UFT are a prerequisite for using the falling edge time

    and the rising edge time. However, profiling algorithms only use LFT and UFT can be

    further expanded to also use the falling and rising edge time.

    There are some algorithms that is similar with profiling algorithm such as in

    [12].this algorithm use the time window to find the UFP and LFP and take the

    difference between UFP and LFP ( UFP LFPRSS ). If both of them reach to their respective

    threshold and LFP happens before the UFP. Fall will be flagged.

    UFP LFP UFP LFPRSS t t

    with t = time of occurrences (2.5)Improving the result of a fall detection system is extreme important so many

    algorithms check the body posture to improve on the specificity. Most of these

    algorithms assume with many case of posture such as making a threshold similar to

    UFT and LFT when the body is sitting or lying or running. We will compute the

    Lower Sitting threshold and the Upper Sitting threshold. After achieving both LFT and

    UFT 10 or 15 seconds, if both Lower Sitting threshold and Upper Sitting threshold are

    achieved, fall will be flagged.

    Another way for falling detection is measuring the velocity which is computed

    by the formula.

    dRSS dt (2.6)

    The velocity will be added in the profiling algorithm and based on the integral

    from the time of the start of a fall until impact.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    22/59

    12

    In addition, we can also use the fall index such as in 2.7 or make fall thresholds.

    Especially, we can also make use both in three different types of algorithm.

    First algorithm checks posture and impact. With impact, it will be based on Z2,

    UFT, dUFT or UFP LFPRSS .The second algorithm uses LFT + UFT (within frame of 1 second) or threshold

    based on Z2 + monitoring.

    The third algorithm uses LFT + threshold with the velocity and UFT (within

    time frame of 1 second) or threshold based on Z2 + monitoring of posture.

    Postured will be detected 2 seconds after the impact using the vertical axis

    acceleration. The sitting and lying posture are usually lower than 0.4G. The

    result of the first algorithm when using Z2 based threshold + posture or UFT isusually 97%. The fall index will be calculated by the formula:

    2

    1

    , , 19

    (( ) ( ) )i

    i k i k i

    k x y z i

    FI A A (2.7)

    Where x, y, z = acceleration from the X, Y and Z axis

    2 2 2RSS RSS G

    22G

    dZ (2.8)

    The scalar product of two acceleration vectors can also measured for Posture to

    fin the angle between them. It maybe is between the current gravity vector and the

    reference gravity vector or the gravity and the vertical axis (Ay).

    arccos|| || || ||

    referemcegravity currentgravity

    referencgravity x currentgravity (2.9)

    arccos|| || || ||

    d

    y

    A gravityA x gravity

    (2.10)

    Some devices have orientation sensors or more accurate gyroscopes which then

    can be used instead [10]. Both formula 2.5 and 2.11 are used.

    | sin sin cos cos |x z y y z y z

    Orientation A A A (2.11)

    Where x , y , z = the data from the orientation sensor

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    23/59

    13

    X is the front-back axis, y is the horizontal axis (left and right) and z is the

    vertical axis (up and down).

    2.2.1 Learning Methods for fall detectionMachine learning can be broadly classified into two fields, supervised learning

    and unsupervised learning. In front, the machine will attempt to identify the groups of

    similar data from a larger dataset. Besides, it also tries to from clusters of data based

    on some criteria such as cost function. It has not prior knowledge of the data layers. It

    only tries to identify natural clusters or groups of data. With supervised learning, it

    learns from the test containing classified data and predicts data layers invisible. With

    Fall-detection system, it will be more natural when using supervised learning

    techniques.

    Without any analytical algorithm, we can still carry out an intuitive approach

    the development of machine learning based fall detection systems from a training

    period and then classification. However, It may be is necessary to establish criteria for

    classification.

    The most important in supervised learning is the quality of accurate and

    exhaustive of training data. Because of these reasons, in the period of training, it is

    important to be able to simulate how falls as close and how the real falls for othergroups of users who will use the fall detection system.

    With some machine-learning algorithm out there, it may be cause confusing

    about selecting the right one for certain applications. Even thought, some research also

    studies about comparing the result of them to try to choose the best algorithm [18],[19]

    To sum up, a developer can be chosen an application method that is suitable with his

    (her) own discretion. An evident with fall detection system is a sensitive application. It

    is advised to compare and test between a numbers of algorithms to try picking the bestchoice.

    Ralhan in University of Saskachewan compared five different supervised

    learning methods for fall detection system [6]. Naive Bayesian classifier, Radial basis

    Function (RBF), C4.5 and Ripple down Rule Learner, Support Vector machine are

    used. Eight scenarios (four falls, three ADLS and one near fall) are used to getting data

    from test the algorithms against. With Nave Bayes classifier, the highest result is 97%

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    24/59

    14

    and taking the least time for building model and then it was chosen as the best choice

    for research.

    In addition, Fall detection by embedding an accelerometer a method that of

    Ralhan in University of Saskachewan used with machine learning algorithm [5]. Theyused the group integrated a device with tri-axial accelerometer, a MCU device and

    some other peripherals with a mobile phone. One class SVM for preprocessing the

    signals and KFD (Kernel Fisher Discriminant) and K-NN (Nearest neighbor) was used

    with fall detection for precise classification. The result of research is 92.3% for a

    limited number of cases.

    2.3.SummaryStudying and evaluating that system give us the overview about fall detection

    approaches. In this few years, researches in this field are developing more and more.

    There are many approaches for fall detection such as based on image, video etc.

    Besides, the rapid advances in wireless network and tiny sensor promote the

    development of smart phones application. However, these approaches seem to only

    focus on using threshold of the length of acceleration. Therefore, we propose to

    develop a new fall detection system with machine learning based approach for this

    problem.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    25/59

    15

    Chapter 3

    Fall Detection Process

    In the period of studying for fall detection and the installation of experiment,

    we found that the fall detection algorithms based on threshold still have the following

    limitations:

    They have not been built a specific training data on typical actions falls.

    The techniques of experimental process based on the characteristic time

    domain and or frequency domain were not mentioned. They have not

    been proper interest in these methods.

    In addition as mentioned in the previous chapter about the theory of fall

    detection. In the data collection, human undesired movement can cause

    large changes in intensity dramatically at that moment. It affects directly

    on the result obtained when comparing with the threshold.

    To solve the above disadvantages, we propose the method to extract feature

    combined building layer model for fall detection system. In particular, the model will

    be proposed for using extracted techniques based on the information on the time

    domain and the characteristic value in accelerometer sensor. The following thesis will

    present general model for classification

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    26/59

    16

    Figure 3-1: The general model cycle

    The components constitute the layer model using in this thesis. This model isbased on the criteria that the thesis proposed to decrease limitations of the above

    methods. Accelerometers data will be collected directly from the user through specific

    cases which described. Next part is to expand intensity and assign label for datasets. In

    this part, some other information of accelerometer data on the time domain will be

    extracted which are used in the classifying phase.

    3.1.Accelerometer Data collectionIn this section, we introduce first how the data collected. We collected data of

    falls from different directions (x, y, z axes) and different environment (Bedroom,

    kitchen, outdoor garden, living room) with other cases of fall. The data from activities

    of daily living (ADL) such as walking, jogging, sitting or standing are also be

    collected. This below sample data acquired from accelerometer and fall annotation for

    a single fall:

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    27/59

    17

    Table 3.1: Sample data acquired from acceleromter

    x-axis y-axis z-axis time

    0.459687 1.37906 9.80665 57223591098406

    0.459687 1.532289 9.346964 57223619296739

    0.153229 1.225831 9.95987 572236779434050.459687 1.225831 9.040505 57223741578404

    -0.30646 0.459687 10.72602 57223799416739

    0.153229 0.459687 9.500193 57223862593408

    0.153229 0.153229 9.500193 57223921180075

    0.153229 -0.45969 10.5728 57223982330075

    0.153229 -0.30646 9.500193 57224039275074

    -0.15323 -0.76614 9.193734 57224100655074

    -0.1533 -1.53229 9.346964 572241619600720.612916 1.225831 9.959879 57223529871740

    0.459687 1.37906 9.8066 57223591098406

    0.459687 1.532289 9.346964 57223619296739

    0.153229 1.225831 9.959879 57223677943405

    Figure 3-2: Fall annotation for a single fall

    In getting data for fall detection system, the cell phone was put in the shirt

    pocket of participants. In each case of position, every participant falls 7 times in

    different directions and environment. To sum up, we recorded 5 examples of the

    behavior from 15 persons and collecting Activities of daily living for 2 or 3 minutes

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    28/59

    18

    from each person with some cases such as sitting, running, lying or walking. Mitja and

    Igone [17] also gave some multiple activities:

    3 x 15 recordings of falling, consisting of standing/ walking, falling and

    lying. Fall detection is one of the main goals of our project.3 x 10 recordings of lying down, consisting of standing/walking, lying

    down and lying. Lying down is same as falling down, so we wanted to

    verify whether the classification can distinguish between the two.

    3 x 10 recordings of sitting down, consisting of walking, sitting down

    and sitting. Sitting down may also resemble falling and is a common

    action .It is important for the analysis of the users behavior.

    3 x10 recordings of walking. Walking is also popular and we wanted

    clearly examples of it to test the classification. Recognizing word using

    trial cutting and beam search

    3.2.Algorithm for Fall Detection

    Figure 3-3: Fall Detection Flow Chart

    Figure 3-3describes the step of data processing and computation of attributes. It

    is namely pre-noise processing data, attribute computation and feature selection.Especially, Feature selection is an important prior step to any classification problem

    which reduces the dimensionality and thus the amount of data required for training.

    These attributes are later combined to create the final attribute vector which is used in

    the machine learning classification stage .They will be presented as follows:

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    29/59

    19

    3.2.1.Pre-noise processing dataBecause the data is received from sensor (usually is noisy data), so the data

    filter techniques have been applied. There are several techniques that perform signal

    filter. In this thesis, we applied a low-pass and high-pass filter.

    A simple low-pass filter for the time domain is a smoothing function. In other

    work, the signal is smoother and less dependent on short changes. We used low-pass

    filter to reduce the influence of sudden change on the accelerometer data.

    It is also possible to filter a series such that the low-frequency variations are

    reduced and the high-frequency variations unaffected. This type is called a high-pass

    filter. This is particularly important in the acceleration data. This type also allows us to

    remove the gravity component and take into consideration the sudden change in

    acceleration.

    The below algorithm shows the low-pass and high-pass filter that we use in this

    thesis. It uses a low-value filtering factor to generate a value that uses 80% of

    the previously filtered value and 20% of the unfiltered acceleration data. This factor

    was chosen empirically.

    Figure 3-4: Low-pas and high-pass filter algorithm

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    30/59

    20

    3.2.2.Action SegmentationIn this section, we describe the process of action segmentation. Sliding window

    is a common approach to solve the problems of fall recognition. We use an overlap

    sliding window technique for fall detection where different actions are performed

    continuously. The idea is to move the sliding window across the accelerometer in the

    accelerometer sequences and decides what action the actor is performing inside the

    window.

    2 4{ ,..... }

    t w t w t I I I

    With:

    W is the width of sliding window

    t is the index of current acceleration

    Figure 3-5: (a) Sliding window technique. (b) Overlap sliding window technique

    When analyzing series of time, we chose a window size of 10 which is one

    second time interval. We decided for one-second time because some transitional

    activities usually last from one to five seconds.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    31/59

    21

    3.2.3.Feature ExtractionTo detect fall based on accelerometer sensor with machine learning algorithm,

    we will use sliding window to transform stream of acceleration data into instances for

    machine learning. The following attributes were derived from the data within slidingwindow. We will against on some method to calculation:

    Length of the acceleration vector

    The Standard Deviation of the acceleration vector within the window

    The speed of change in acceleration between the maximum and minimum

    along the x, y, z axes.

    The maximum and the minimum acceleration along the x, y, z-axes

    The acceleration vector changes (AVC)

    The accelerometer inclination angles

    3.2.3.1. Length of the acceleration vectorThe first computed attribute is length of the acceleration vector. It is a simple

    but it is very useful attribute. Moreover, it is also used further in the process ofextraction the new attributes. It is not used as a separate attribute in the final vector

    because of the sliding window technique. Its definition is:

    2 2 2RSS x y z

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    32/59

    22

    Figure 3-6: RSS plot of a forward fall

    During static posture this attribute is constant with the value equal to the

    Earchs gravity (RSS = 1g). In dynamic activities the acceleration vector is changing

    the direction and its length.

    3.2.3.2.The Standard DeviationThe Standard Deviation attribute is useful for distinguish the long lasting static

    activities. It can detect when the movement of the sensor its intense. With thesecharacteristics, based on the acceleration value, we compute the Standard Deviation of

    the acceleration vector. The Standard Deviation with in sliding window will be defined

    as follows:

    2

    1

    ( )n

    i

    ii

    a a

    STDN

    (4.1)

    With N is the number of acceleration data within the window, ia is the length

    of the thi acceleration vector and a is the average length of the acceleration vector of

    the person)

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    33/59

    23

    3.2.3.3.The speed of change in acceleration along the z-axisThe speed of change in acceleration along the z-axis was defined as follows:

    max( ) min( )

    tan (max( ) (min( ))

    z z

    zz z

    a a

    spd t a t a (4.2)

    For this attribute, the value of raw materials for the length of the acceleration

    vector is used instead of low-pass value .Max (z

    a ) and min (z

    a ) are the maximum and

    minimum acceleration along the z axis within the window, and t (max ( za )) and t (min

    ( za )) are the time stamp of the data.

    3.2.3.4. The acceleration vector changes (AVC)When the person's body is static, single accelerometer response only to the

    gravity, producing a constant length of acceleration is 1g. In the moving accelerations

    produce a changing in acceleration signal and drastically change the motion. Using

    changes in the acceleration vector, an attribute is calculated to detect the motion

    accelerometer: Acceleration Vector Changes (AVC). The AVC value of this attribute

    increased as the acceleration (walking, going down, stand up, etc ...). This attribute

    take into consideration the data from the current window (ten data samples). It

    combines ten different length of the vector of length acceleration vector and divided

    the sum by the time interval (one second) of data. The AVC is computed as followed:

    1

    1

    0

    | |n

    i i

    i

    n

    length length

    AVCT T

    (4.3)

    0T is the time stamp for the first data sample in the window, and nT is the time

    stamp of the last data sample. The fall can be detected by using this attribute. With this

    attribute, the movement of people can be detected: it distinguishes static from the

    dynamic active. For this attribute, the value of raw materials for the length of the

    acceleration vector is used instead of low-pass value. The reason for this is that we are

    more interested in small changes in the acceleration signal and the smooth of low-pass

    filter.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    34/59

    24

    3.2.3.5. The accelerometer inclination anglesOther important features to be recognized as a static body posture are the

    orientation angles of the accelerometer. The orientation angles are calculated as the

    angles between the actual acceleration and each of the axes (x, y and z axes). (Figure3-7)

    Figure 3-7: Accelerometer inclination angle

    For instance, the angle x between the acceleration vector and the x axis(perpendicular to the ground) is computed as follows:

    2 2 2arccos( )x

    x y z

    ax

    a a a (4.4)

    Where the value xa , ya , za represent the actual acceleration vector. In this

    attributes, lower-pass filter should be used. Because of the changes of the angle is less

    and less variation. Without the low-pass filter, the angles were sensitive angles to each

    small change of the accelerometer. These angles improve the classification of activities

    that have different accelerometer angle inclinations.

    3.2.4.Forming feature vectorIn our work, feature vector is formed for each sliding window (sub-regions) in the fall

    data sequence and thus, each sequence is represented by a set of feature vector. The process in

    which feature are formed is illustrated in the following:

    From each input sample, we use low-pass and high-pass filter as described in

    section 3.2.1

    Each feature is then divided into 10 non-overlapping sub-regions.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    35/59

    25

    The value of attributes in section 3.2.3 is calculated for each sub-region.

    The feature vector is created by concatenating all attribute computed from sub-

    regions ( each sliding window)

    3.3.Action learningIn this thesis, we make use of the famous Support Vector Machine (SVM) for action

    learning and recognition.

    3.3.1.Support Vector Machine (SVM)Support Vector Machine is one of the most famous machine learning techniques

    which has been used widely in not only Computer Vision but also in Artificial Intelligent and

    other fields. Especially, with classification, SVM is the tool for finding good hyperplanes for

    separating different classes of instance.

    Considering binary classification problem with L training point { ix , iy } where:

    1, 1( ),....,( , ) {1, 1}m mx y x y X

    and test set1,.....,

    mx x X

    With a given training set1 2{( , ) | 1... }, ( , ,... ) , { 1, 1}

    n

    i i i i i in iD x y i m x x x x R y which is

    separable, we need to find a hyper plane which separates two object classes +1, -1 and

    predicts bets with data not in the training set. Figureillustrates a hyper plane separating two

    types of data objects.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    36/59

    26

    Figure 3-8: Separating hyper plane

    Optimal hyper plane has the equation:

    0)(1

    m

    i

    ii bxwbxwxf

    Satisfying constraints:

    mibxwyi ,...,1,1)(

    Then, an objectx is classified into +1 class iff(x) 0 and into1 class otherwise.

    The distance from margin to optimal hyper plane is w1

    . The problem here is to find w

    so that ||w|| has a minimum value satisfying constraints ,1)( bxwyi i= 1 m

    with the hope that the larger margin, the better classifier.

    This problem can be solved by solving its dual problem. Lagrangian for the

    primal problem is:

    m

    i

    iii xwywwbwL

    1

    ]1)([

    2

    1),,( (3.3.1)

    where i 0 is Lagrange multiplier. We can transform the primal into a dual

    by simply setting to zero the derivatives of the Lagrangian with respect to the primal

    variables, and substituting the relations so obtained back into the Lagrangian, hence

    removing the dependence on the primal variables:

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    37/59

    27

    m

    i

    ii

    m

    i

    iii

    yb

    bwL

    xyww

    bwL

    1

    1

    ),,(

    ),,(

    Then, we have:

    01

    1

    i

    m

    i

    i

    m

    i

    iii

    y

    xyw

    Replace them into (3.3.1), we have:

    m

    i

    iii xwywwbwL1

    ]1)([2

    1),,(

    m

    ji

    jijiji

    m

    i

    i

    m

    i

    i

    m

    ji

    jijiji

    m

    ji

    jijiji

    xxyy

    xxyyxxyy

    1,1

    11,1,

    2

    1

    2

    1

    The formula above is the dual representation of primal optimization problem. In

    optimization theory, solving primal problem is equivalent to solving its dual problem.

    Assuming * is the solution of the following dual optimization problem:

    Maximizei ji

    jijijiiD xxyyL,2

    1

    where i 0, i = 1m and 01

    i

    m

    i

    iy

    Then vector of weights ism

    i

    iii xyw1

    ** and

    Ii

    n

    j

    jijji xxyy

    I

    b1||

    1

    whereIis set ofi so that i > 0. The classification function is sgn(f(x)) where:

    *)(*

    bxxyxfIi

    iii

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    38/59

    28

    3.3.2. LearningIn this work, we also apply grouping technique for the learning phase:

    Assume training example iX has iN sub-regions which are represented by iN different feature vectors. Each of these feature vectors will be labeled with the corresponding

    class of the sample. Therefore, with n samples of two action class (fall and non-fall), a

    training set D is described as:

    { , , }iNn

    ij ij ij i

    i j

    D x y y y

    With:

    ijx is theth

    j feature vector of the thi example

    ijy is the label for

    ijx

    iy is the action class of the thi example ( {0,1}iy )

    3.4.SummaryIn this section, we presented an overview from data collection to feature extraction.

    Pre-processing data is one of the most important parts in the process of feature extraction.

    Especially, we also presented overlapping window technique which is usually used in

    classifying activity sequences or image series, etc. Based on the overlapping window

    techniques, the typical attribute for accelerometer vector are computed and created feature

    vector for classification.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    39/59

    29

    .

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    40/59

    30

    Chapter 4

    Experimental Setup and evaluation

    This chapter presents the experiments we carry out to build a database of fall

    data from real people and analyze the performance of the fall detection method. In this

    section, we first describe how we set up our experiments, explain the choice of

    parameters for fall detection and finally show the result of our method in comparison

    with other related methods.

    4.1. Experiment SetupThe following briefly presents the basic concept of fall in our daily activities and

    describe the process of data collection which is used in our experiments.

    4.1.1.Background of fall and non-fall actionAs mentioned in Chapter 2, we have presented some fall detection system that

    uses various approaches in the study. Their disadvantages and advantages were

    discussed. There is need for a precision of fall detection system that wishes train and

    test. Because of this, studying for falls is an important issue with the thesis.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    41/59

    31

    4.1.1.1.Study of fallDefinition a fall

    Many definitions about a fall are given. This is because fall for one specific

    group of users such as young people, elder people or different other groups. This thesis

    focuses on distinguishing fall and non-fall activities (it may be called Activity of Daily

    Livings or ADLs). Now as mentioned, types of fall of each group of people always are

    different (elderly people or childetc).

    Figure 4-1: The stage of fall

    In our thesis, we will distinguish between two kinds of falls. The first place is

    falls which are caused by external factor (such as behavioral or environmental). The

    second place is falls that caused by internal factors (such as biology).

    Lying down/ Standing/ Walking

    Fall Forward/Backward/Left/Right

    Near Fall

    Cause of falling

    The normal changes of aging, like poor vision and poor hearing. They can make

    you more likely to fall.Diseases and physical conditions can affect your strength and

    balance.

    There are some risk factors for cause of falling. In this thesis, we divide these

    factors into intrinsic (biology) and extrinsic (behavioral and environmental). We use

    both of factors to detect falling.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    42/59

    32

    Firstly, biological factors may be are:

    1. Age2. Medical condition such as Parkinsons disease, patients.3. Muscle Weakness4. Visual impairment.5. Foot problem.Secondly, Behavioral and Environment factors are:

    1. Sedentary2. Medication intake.3. Alcohol misuse.

    Beside two types of groups for falling, we also research some trends of falls.

    They are as follows:

    1. Time2.

    Climate/weather

    3. Location4. Race5. Depression

    Phases of falling

    There are four phases of falling: the pre-fall, the critical phases, the post-fall phase

    and the recovery phase

    Figure 4-2: The four phases of a fall event

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    43/59

    33

    The pre-fall phase corresponding movement of daily life, sometimes

    with sudden movements directed towards the ground such as sitting

    down or bending over. These activities should not create an alarm with a

    fall detection system.

    The critical phase, corresponding to the fall, is highly short. This phase

    can be detected by the motion of the body toward the ground or by the

    shock of impact with the floor

    The post-fall phase is usually characterized by a real person on the

    ground immediately after the fall. It can be detected by a lying position

    or by an absence of significative motion

    A recovery phase may eventually occur if the person can stand alone or

    with the help of another person

    4.1.1.2.The simulated-fall studyThe simulated fall study related to healthy young subjects and carrying out

    under the supervision. Tri-axial accelerometers signals recorded from the trunk and

    thigh in a fall event simulation. Each subject performed eight different kinds of fall

    and each type was repeated three times.

    The fall types used in the testing process for the current study were selected tosimulated common types of fall in elderly people. The most common causes of falls

    are the trips, slips, and loss of balance. ONeill et al.[20]indicated that 60% of falls in

    older people were falling in the forward direction. Laterally directed fall also were a

    major pose of threat. Laterally directed fall causes the bad impact such as the potential

    to fracture every time it happens. Therefore, falls from all directions should be

    considered when validating fall detection system. We also should attempt to make

    mimicking the realistic falls. Thus, simulation of performance were backward falls,

    lateral falls right, lateral fall left and forward falls.

    4.1.1.3.The ADL study (Activity of Daily Living)The second study involved older people performing ADL, in their own homes,

    while equipped with the same sensor configuration. Ten subjects of the elderly

    community, four females and six males, were monitored. They ranged in age from 60

    to 80 years old. Each ADL was performed three times by each older person. The ADL

    were a task that could produce impacts or sudden changes in movement and result of a

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    44/59

    34

    mistake caused by a threshold-based fall detection algorithm. This below picture

    shows activities of daily living:

    Figure 4-3: Two types of non-fall(ADL)

    The second study involved every activity of living from children to elderly

    people. Each ADL was performed three times by each person. To evaluate fall

    detection algorithm, Bourke, OBrien and Lyons showed some types of ADL in their

    research [7]:

    1 Sitting down and standing up from an arm-chair2 Sitting down and standing up from a kitchen-chair.3 Sitting down and standing up from a toilet seat.4 Sitting down and standing up from a low stool5 Getting in and out of a car seat6 Sitting down on and standing up from a bed7 Lying down and standing up from a bed8 Walking 5m9 Cycling

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    45/59

    35

    4.1.2.Data CollectionWith intent and purposes of the system that we want to make. We define a fall

    to be a sudden change of body position coming to rest on the ground, it does not

    include intentional change, and then is inactivity. This means that the most serious

    falls where user loses a balance after hitting the ground. They also lose the ability for

    getting help.

    To have database to train and test the character recognizer, we designed a

    number of this scenario was designed to investigate the events which can be difficult

    to recognize such as fall and non-fall. It includes 9 different events which are showed

    in Table 4.1. They are recorded in single recordings that it interspersed with shortperiods of walk, each record lasted from 3 to 5 minutes. An example for specific

    image of the event can be view inhttp://dis.ijs.si/confidence/iaai.html

    Table 4.1: Events Sequence in the test Scenario

    No Description Fall Ending position

    1 Sitting down normally on the chair No

    2 Tripping, landing normally on the

    bed

    Yes

    3 Lying down normally on the bed No

    4 Falling slowly( trying to hold onto

    furniture),landing flat on the ground

    Yes

    5 Sitting down quickly on the chair No

    6 Falling when trying to stand up,

    landing sitting of the ground

    Yes

    7 Lying down quickly on the bed No

    8 Falling slowly when trying to stand

    up(trying to hold onto furniture),

    landing sitting of the ground

    Yes

    9 Searching for something on the

    ground on all fours any lying

    No

    http://dis.ijs.si/confidence/iaai.htmlhttp://dis.ijs.si/confidence/iaai.htmlhttp://dis.ijs.si/confidence/iaai.htmlhttp://dis.ijs.si/confidence/iaai.html
  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    46/59

    36

    Some types of falls were selected from a list of typical falls. As shown in the

    choosing on related work. Accelerometer sensor can accurately detect typical falls.

    Because of this, we gave a fall (such as event 2) to prove that the system can recognize

    .In addition, we also choose three atypical falls (event 4, 6 and 8) to test the use of

    information in each circumstance. A specific example that a person is not expected to

    lie or sit on the ground (in contrast with the bed or the chair). They are atypical in

    speed (event 4 and 8) and starting/ending posture (event 6 and 8). With each subject,

    we would repeat 2 or 3 times for each. Therefore, there were about 300 sequences of

    data (120 fall data and 180 ADL data)

    The experiments with groups of real persons are conducted. Both activities of daily

    living and falls are tested and training. However, the fall situation is dangerous to

    human body, especially to the elderly people, so we cannot test falls with elderlypeople, the elderly volunteers only attended with non-fall (ADL). We received data

    from 2 groups of subject:

    Ten young subjects (6 male and 4 female, age 256 years, body mass

    5311.5 kg and height 16610.5 cm)

    Table 4.2: The detail of ten young volunteers

    Volunteers Age(years)

    Weight(kg)

    Height(cm)

    Subject 1 22 53 169

    Subject 2 19 54 168

    Subject 3 33 58 164

    Subject 4 25 56 165

    Subject 5 28 63 174

    Subject 6 21 55 170

    Subject 7 25 59 166

    Subject 8 23 60 171

    Subject 9 22 50 167

    Subject 10 22 58 165

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    47/59

    37

    Ten elderly subjects and children (6 subjects for age 646 years, body mass

    525kg, and height 1645cm,4 subjects for age 104 years old).

    4.2.Performance measuresFall detection is either positive if the detector properly recognizes a fall event or

    negative if it does not. The quality of fall detection cannot be evaluated from a single

    test; instead, it is necessary carry out from a series of test. It will include four possible.

    True positive (TP): fall occurs, the algorithm detects the fall

    True Negative (TN): a normal movement (non-fall) is performed. Thealgorithm does not detect a fall.

    Fall Positive (TP): The algorithm detect a fall, but fall does not occur

    Fall Negative (FN): a normal movement (non-fall) is performed. The

    algorithm detects a fall.

    Based on four situations, we proposed three criteria to evaluate the fall

    detection system:

    Sensitivity is the capability to detect a fall. It can be expressed:

    ensTP+FN

    TPS itivity

    Specificity is the capability to detect an ADL (Activity of Daily Living). It

    can be expressed:

    TNSpecificityTN FP

    F measure is harmonic mean of the specificity and the precision

    2sensitivity specificity

    F measureSensitivity specificity

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    48/59

    38

    4.3. Experimental results of the fall detection systemWe tried various machine learning algorithms to train classifiers for classifying

    the fall and non-fall. In our experiment, 20 volunteers were selected to attend in the

    experiments. The detail of them was mentioned in the previous section. The

    experiment was performed with the scenario in Table 4.1.From 300 samples received

    from the testing scenario, we conducted experiment with four categories. With these

    samples, we randomly selected 3/4 of samples for training and other model for testing.

    It was classified with libSVM function in the software of toolkit WEKA. As shown in

    Table 4.3 is the detail of four categories.

    Table 4.3: The distribution of the samples

    Category total samples samples for

    training

    sample for

    test

    1 90 66 24

    2 70 52 18

    3 60 45 15

    4 80 60 20

    The test results are shown in Table 4.4, and the mean ratio of accuracy mr is 92.2

    percent, where:

    1

    1

    Nm i

    i

    r r

    N is the number of categories, ir is ratio of accuracy ofthi category.

    Table 4.4: The test result

    Category 1 2 3 4

    Sensitivity (%) 91.3 94 91.3 92.6

    Specificity (%) 91.4 93.9 90.3 92.6

    F-measure (%) 91.4 94.0 90.8 92,6

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    49/59

    39

    When using an accelerometer for fall detection, the machine learning based

    methods somewhat outperformed threshold algorithm. Considering the simplicity of

    these methods, threshold may be simpler than machine learning methods, but the

    related work shows that in some case, machine learningbased methods gave the better

    result because this may seem surprising, but the related work shows that threshold-

    based methods were apparently able to learn the pattern of acceleration during the

    falling.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    50/59

    40

    Chapter 5

    Conclusion

    Because our goal is to detect fall, most of the non-fall (ADL) we have obtained

    samples are more specific than daily activities (such as cooking, sleeping). In other

    words, the probability of fall-actions is higher than in everyday of life. Compare with

    other methods in [21]. The result of our study is more specific. The correct ratio is

    94%.

    This thesis studies to build a standard database and typical attributes for fall

    events based on accelerometer sensor. We studied the fundamental knowledge about

    fall detection such as data feature extraction, machine learning models, support vector

    machine and how to combine them.

    This thesis has achieved the following results:

    Study to build a model of fall detection based on accelerometer sensor

    Studying and learning the typical case of falls, the data processing techniques

    and the characteristic of the accelerometer sensor are used.

    Build a special dataset for testing falls.

    About theory and methods, in terms of the thesis, we explore several detection

    techniques identify different falls including pre-processing techniques

    accelerometer sensor, feature extraction from raw accelerometer data and

    classification techniques using in signal classification problems.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    51/59

    41

    In particular, we proposed a model to improve the accuracy of fall detection

    based on accelerometer sensor to ensure objectivity, the proposed model was

    compared with the evaluation results of other fall detection algorithms in

    previous studies.

    In the future, we will continue to improve the recognition quality with a large

    enough data for a fully evaluated general methods and experimental models.

    Especially, we also try to fix the standard signal from many different users more fully.

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    52/59

    42

    Bibliography

    [1] Li,Q.,Stankovic,J.,Hanson,M.,Barth,A.,Lach,J.,&Zhou,G.2009.Accurate,fast fallDetection using gyroscopes and accelerometer-derived posture information. In

    Wearable and Implantable Body Sensor Networks, 2009. BSN2009. Sixth

    International Workshop on, 138143

    [2] Noury,N.,Fleury,A.,Rumeau,P.,Bourke,A.,Laighin,G.,Rialle,V.,&Lundy,J.2007.Fall detection-principles and methods. In Engineering in Medicine And

    BiologySociety,2007. EMBS2007.29th Annual International Conference of the

    IEEE,16631666.

    [3] Mitja, L. &Bostjan, K. Fall detection and recognition with machine learning.Technical report, Joef Stefan Institute, Department of Intelligent Systems, 2009.

    http://dis.ijs.si/mitjal/documents/Fall_detection_and_activity_recognition_with_m

    achine_learning-Informatica-09.pdf.

    [4] Noury, N., Fleury, A., Rumeau, P., Bourke, A., Laighin, G., Rialle, V., & Lundy, J.2007. Fall detection - principles and methods. In Engineering in Medicine and

    Biology Society, 2007. EMBS 2007. 29th Annual International Conference of the

    IEEE, 16631666.

    [5] Zhang, T., Wang, J., Liu, P., &Hou, J. 2006. Fall detection by embedding anaccelerometer in cell phone and using kfd algorithm. In International Journal of

    Computer Science and Network Security, Vol. 6 no. 10 pp 227284.

    [6] Ralhan, A. S. A study on machine learning algorithms for fall detection andmovement classification.Masters thesis, University of Saskatchewan,

    http://library2.usask.ca/ theses/available/etd-12222009-144628/, 2009.

    [7] A.K.Bourke,J.V.OBrien,G.M.Lyons Evalutaion of a threshold-based tri-axialaccelerometer fall detection algorithm

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    53/59

    43

    http://www.eee.nuigalway.ie/documents/go_gait_posture_2007.pdf

    [8] Ubisense. http://www.ubisense.net, 2008-09-15[9] Maybeck, P. S. Stochastic models, estimation, and control. Mathematics in Science

    and Engineering 141, 1979.

    [10] Dai,J.,Bai,X.,Yang,Z.,Shen,Z.,&Xuan,D.292010.Perfall: A pervasive fall Detection system using mobile phones. In Pervasive

    Computing and Communications Workshops (PERCOM Workshops), 2010 8th

    IEEE International Conference on.

    [11] Bourke,A.,vandeVen,P.,Gamble,M.,OConnor,R.,Murphy,K.,Bogan,E.,McQuade,E., Finucane, P., Laighin, G.,&Nelson,J.2010.Evaluationofwaist-mountedtri-axial

    Accelerometer based fall-detection algorithms during scripted and continuous

    unscripted activities. Journal of Biomechanics, 43(15), 30513057

    [12] Jantaraprim,P.,Phukpattaranont,P.,Limsakul,C.,&Wongkittisuksa,B.may2010.Improving the accuracy of a fall detection algorithm using free fall

    characteristics. In Electri-Cal Engineering/Electronics Computer

    Telecommunications and Information Technology (ECTI-CON), , 501504.[13] A threshold-based fall-detection algorithm using a bi-axial gyroscope sensor

    http://www.eee.nuigalway.ie/documents/go_med_eng_phys_2008_a.pdf

    [14] Fall detection Using 3D Head Trajectory Extracted From a Signal Camera VideoSequence

    http://www.computer-vision.org/4security/pdf/montreal-fall_detection.pdf

    [15] Boyle, J. &Karunanithi,M.2008. Simulated fall detection via accelerometers. InEngineering in Medicine and Biology Society, 2008.EMBS 2008.30

    th

    AnnualInternationalConference of the IEEE, 12741277.

    [16] Dai,J. ,Bai,X., Yang,Z., Shen,Z., &Xuan, D.292010.Perfalld: Apervasive falldetection system using mobile phones. In Pervasive Computing and

    Communications Workshops (PERCOM Workshops), 2010 8th IEEE International

    Conference on.

    [17] Mitja LUTREK1, Matja GAMS1, Igone VLEZ2, IST-Africa 2009 ConferenceProceedings Paul Cunningham and Miriam Cunningham (Eds) IIMC International

    http://www.eee.nuigalway.ie/documents/go_med_eng_phys_2008_a.pdfhttp://www.eee.nuigalway.ie/documents/go_med_eng_phys_2008_a.pdfhttp://www.computer-vision.org/4security/pdf/montreal-fall_detection.pdfhttp://www.computer-vision.org/4security/pdf/montreal-fall_detection.pdfhttp://www.computer-vision.org/4security/pdf/montreal-fall_detection.pdfhttp://www.eee.nuigalway.ie/documents/go_med_eng_phys_2008_a.pdf
  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    54/59

    44

    Information Management Corporation, 2009. Posture and Movement Monitoring

    for Ambient Assisted Living

    [18] Caruana,R.&Niculescu-mizil,A.2006.An empirical comparison of supervisedlearning algorithms. In InProc.23rdIntl.Conf.Machine

    learning(ICML06),161168.

    [19] Caruana,R.,Karampatziakis,N.,&Yessenalina,A.2008.Anempirical evaluationof supervised learning in high dimensions. In International Conference on

    Machine Learning (ICML) 96103.

    [20] ONeill TW, Varlow J, Silman AJ, Reeve J, Reid DM,Todd C,etal.Age and sex in fluences on fall characteristics. Ann Rheum Dis1994;53(11):7735

    [21] Zhang, T., et al., Fall Detection by Wearable Sensor and One-Class SVMAlgorithm. Lecture Notes in Control and Information Science, 2006. 345: p.

    858-863

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    55/59

    45

    Appendix A

    ADL (Non-Fall)

    1. From standing position, sit down (normal speed) on chair , remain 5 seconds andthen stand up

    2. Kneel down and pick up an item

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    56/59

    46

    3. Standing position, bend-down and pick up an item on the floor. Rise-up

    4. Standing position, quickly fall down facing the floor, push up 3-5 times then rise up

    5. Lie down and remain 5 seconds. Rotate 180 degree, remain 5 second , then rise up

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    57/59

    47

    6. Standing position, back to original position

    7. Standing position and jump

    8. Standing position and run from A to B.

    9. Lie down on the table

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    58/59

    48

    10.Climb up on the table, remain 5 second and then jump down

  • 7/31/2019 Ngo Duy Kien _Thesis Version 10_29_Final

    59/59