the application of fuzzy granular computing for the
TRANSCRIPT
University of Texas at El PasoDigitalCommons@UTEP
Open Access Theses & Dissertations
2012-01-01
The Application Of Fuzzy Granular ComputingFor The Analysis Of Human Dynamic Behavior In3D SpaceMurad Mohammad AlaqtashUniversity of Texas at El Paso, [email protected]
Follow this and additional works at: https://digitalcommons.utep.edu/open_etdPart of the Artificial Intelligence and Robotics Commons, Biomedical Commons, and the
Computer Engineering Commons
This is brought to you for free and open access by DigitalCommons@UTEP. It has been accepted for inclusion in Open Access Theses & Dissertationsby an authorized administrator of DigitalCommons@UTEP. For more information, please contact [email protected].
Recommended CitationAlaqtash, Murad Mohammad, "The Application Of Fuzzy Granular Computing For The Analysis Of Human Dynamic Behavior In 3DSpace" (2012). Open Access Theses & Dissertations. 2025.https://digitalcommons.utep.edu/open_etd/2025
THE APPLICATION OF FUZZY GRANULAR COMPUTING FOR THE
ANALYSIS OF HUMAN DYNAMIC BEHAVIOR IN 3D SPACE
MURAD MOHAMMAD SULEIMAN ALAQTASH
Department of Electrical and Computer Engineering
APPROVED:
Thompson Sarkodie-Gyan, Ph.D., FInstMC, Chair
Amr Abdelgawad, M.D.
Noe Vargas Hernandez, Ph.D.
Richard Brower, M.D.
Scott Starks, Ph.D.
Vladik Kreinovich, Ph.D.
Benjamin C. Flores, Ph.D.
Acting Dean of the Graduate School
Copyright ©
by
Murad Mohammad Suleiman Alaqtash
2012
Dedication
This work is dedicated to:
My wife, Muzaina, who supported me each step of the way
My parents, Haijar and Mohammad, who prayed for me every day
THE APPLICATION OF FUZZY GRANULAR COMPUTING FOR THE
ANALYSIS OF HUMAN DYNAMIC BEHAVIOR IN 3D SPACE
by
MURAD MOHAMMAD SULEIMAN ALAQTASH, M.Sc.
DISSERTATION
Presented to the Faculty of the Graduate School of
The University of Texas at El Paso
in Partial Fulfillment
of the Requirements
for the Degree of
DOCTOR OF PHILOSOFY
Department of Electrical and Computer Engineering
THE UNIVERSITY OF TEXAS AT EL PASO
May 2012
v
Acknowledgements
I would like to acknowledge my mentor, my committee members, my colleagues, my friends and
my family for the help and the support to accomplish this work.
I would like to extend my sincere thanks to my professor and my mentor Dr. Thompson
Sarkodie-Gyan without whose effort my accomplishment wouldn’t have been a reality, special thanks
for his help, support, guidance, and kindness.
I would like to thank each member of my committee. I would like to thank Dr. Vladik
Kreinovich, for his distinguished knowledge and expertise in the area of fuzzy logic. I would also like to
thank Dr. Scott Starks, for supporting this research with the sensors. I would like to thank Dr. Noe
Vargas Hernandez, for participating and guiding of this work. I would further like to thank Dr. Richard
Brower, for providing the impaired subjects and the great financial support of my travel to WCNR2012
conference which will be held in Melbourne, Australia. I would like to thank Dr. Amr Abdelgawad, for
serving this research with the impaired subjects and encouraging feedback. I would like to thank them
for their membership of my dissertation committee and their valuable times.
I would like to express my special thanks to Dr. Huiying Yu, who supervised each step of this
work, as a student and as a professor, to assure the success of the experiments.
I would like to thanks my wonderful colleagues, Luis Sagarnaga, Ryan Price, Julio Torres,
Chethan Keladi, Melaku Bogale, Pablo Rangel, Jorge Garza Ulloa, Mario Renteria, Oscar Espino
(graduated in Dec. 2011), for the massive help during system development and data collection. Their
unlimited effort enriches the research with so much data. . I would further like to thank each member of
LIMA lab for their support and sharing their ideas and food as one family in our second home LIMA.
I would further like to thank Stern Foundation for the financial support of this research over the
past years.
I would never forget to thank my family for unfailing and lasting support.
vi
Abstract
Human dynamic behavior in space is very complex in that it involves many physical, perceptual
and motor aspects. It is tied together at a sensory level by linkages between vestibular, visual and
somatosensory information that develop through experience of inertial and gravitational reaction forces.
Coordinated movement emerges from the interplay among descending output from the central nervous
system, sensory input from the body and environment, muscle dynamics, and the emergent dynamics of
the whole neuromusculoskeletal system.
There have been many attempts to directly capture the activities of the neuronal system in human
locomotion without the ability to clarify how the nervous system adaptively functions as a dynamic
system and how it effectively coordinates adaptive interactions with the musculoskeletal system during
locomotion. Other studies have tried to artificially emulate locomotion by using mathematical models
and robots based on control theory, without any success in the understanding of biological locomotor
mechanisms, as the control laws are artificially constructed solely based on an engineering perspective
independent of actual biological mechanisms.
This research applies fuzzy granular computing and the fusion of multiple wearable sensor data
to analyze human dynamic behavior in 3D space. The outcome of this system may yield cues for the
estimation of external stimuli and/or compensatory responses depending upon which factors are
controlled.
This novel system has been tested and evaluated using two groups of subjects, able-bodied and
neurological patients. The results show a high ability of the system to facilitate an objective and
quantitative assessment of functional gait impairment.
This study provides a potential base for a flexible, efficient and cost-effective system for clinical
gait assessment, a system to assist both doctors and clinicians in the diagnosis of pathological gait
impairments, prescribe treatment, and assess the improvements in response to therapeutic intervention.
vii
Table of Contents
Acknowledgements ................................................................................................................................ v
Abstract ................................................................................................................................................. vi
Table of Contents ................................................................................................................................ vii
List of Tables ........................................................................................................................................ ix
List of Figures ....................................................................................................................................... xi
Chapter 1: Introduction .......................................................................................................................... 1
1.1 Background and Significance .............................................................................................. 1
1.2 Motivation ............................................................................................................................ 3
1.3 Goals and Specific Aims ..................................................................................................... 3
1.4 Scope .................................................................................................................................... 4
1.5 Hypothesis ........................................................................................................................... 4
Chapter 2: State of the Art in Human Motion Analysis ........................................................................ 5
2.1 Summary .............................................................................................................................. 5
2.2 State of the art technologies ................................................................................................. 5
2.3 State of the art techniques .................................................................................................. 11
Chapter 3: Experimental Design and Methodology ............................................................................ 18
3.1 Summary ............................................................................................................................ 18
3.2 Participants ........................................................................................................................ 18
3.3 Experimental design .......................................................................................................... 19
3.4 Data acquisition ................................................................................................................. 20
3.5 Data processing .................................................................................................................. 23
3.6 Intelligent systems and soft computing techniques ........................................................... 26
3.7 Software development ....................................................................................................... 34
Chapter 4: Application of Wearable Sensors for Human Gait Analysis using Fuzzy
Computational Algorithm ........................................................................................................... 35
4.1 Summary ............................................................................................................................ 35
4.2 Methods and materials ....................................................................................................... 36
4.3 Experimental results and discussion .................................................................................. 39
4.4 Conclusion ......................................................................................................................... 45
viii
Chapter 5: Automatic Classification of Pathological Gait Patterns using Ground Reaction Forces
and Machine Learning Algorithms ............................................................................................. 46
5.1 Summary ............................................................................................................................ 46
5.2 Methods and materials ....................................................................................................... 47
5.3 Experimental results and discussion .................................................................................. 51
5.4 Conclusion ......................................................................................................................... 55
Chapter 6: Cost-Effective Device for Measurement of Joint Angles (LIMA Smart Goniometer) ..... 57
6.1 Summary ............................................................................................................................ 57
6.2 Introduction ........................................................................................................................ 57
6.2 Methods and materials ....................................................................................................... 59
6.3 Experimental results .......................................................................................................... 62
6.4 Conclusion ......................................................................................................................... 64
Chapter 7: The Applications of Fuzzy Granular Computing for Assessment of Functional
Impairments in Human Locomotion ........................................................................................... 65
7.1 Summary ............................................................................................................................ 65
7.2 Introduction ........................................................................................................................ 65
7.3 Methods and materials ....................................................................................................... 66
7.4 Experimental results and discussion .................................................................................. 69
7.5 Conclusion ......................................................................................................................... 86
Chapter 8: Research Outcomes, Recommendations for Future Works, and List of Publications ....... 87
8.1 Research outcomes ............................................................................................................ 87
8.2 Recommendations for future works ................................................................................... 88
8.3 List of publications ............................................................................................................ 88
References ............................................................................................................................................ 90
Vita .................................................................................................................................................. 100
ix
List of Tables
Table 4.1: Anthropometric data for 14 male subjects. ............................................................................... 36
Table 4.2: Mean(std) of the rule-based matrices Pref(x,y) and Ptest(x,y) for healthy and MS subjects
respectively and the grade of similarity between them. ............................................................................ 42
Table 4.3: Mean(std) of the rule-based matrices Qref(z,y) and Qtest(z,y) for healthy and MS subjects
respectively and the grade of similarity between them. ............................................................................ 43
Table 5.1: Summary of classification accuracy using the different approaches and classifiers ................ 51
Table 5.2: The classification performance measures using M-shaped, FX and FY features .................... 53
Table 5.3: The classification performance measures using the optimal feature set .................................. 55
Table 6.1: mean±std of correlation coefficient (CORRCOEF) and root mean square error (RMSE)
for knee angles in sagittal and transverse planes. ...................................................................................... 63
Table 7.1: Anthropometric data and speed for twenty able-bodied subjects (12 males and 8 females),
and a group of three females with different gait disorders. ....................................................................... 67
Table 7.2: The degree of similarity DS(G,H) between CP patient and the able-bodied reference for
(A) GRFs, (B) EMG, and (C) ankle, knee, and hip joint angles in 3D. DS is calculated using
different granulation window sizes (wsize) 5, 10, 14, and 20. The mean and standard deviation of DS
are reported. p is the number of fuzzy granules representing the original data series. .............................. 75
Table 7.3: The degree of similarity DS(G,H) between MS patient and the able-bodied reference for
(A) GRFs, (B) EMG, and (C) ankle, knee, and hip joint angles in 3D. DS is calculated using
different granulation window sizes (wsize) 5, 10, 14, and 20. The mean and standard deviation of DS
are reported. p is the number of fuzzy granules representing the original data series. .............................. 80
Table 7.4: The degree of similarity DS(G,H) between CD patient and the able-bodied reference for
(A) GRFs, (B) EMG, and (C) ankle, knee, and hip joint angles in 3D. DS is calculated using
x
granulation window size (wsize) 10. DS are reported for two trials, without insoles and with 1 cm
insole on right shoe. ................................................................................................................................... 85
xi
List of Figures
Figure 2.1: Vicon® Motion System, record human motion by tracking reflective markers using high-
speed cameras. ............................................................................................................................................. 6
Figure 2.2: MEMS inertial sensors, accelerometer and gyroscope, shown respectively. ........................... 7
Figure 2.3: 9 DoF IMU units, MTw and shimmer, shown respectively. ..................................................... 8
Figure 2.4: The rotary encoders, magnetic and optical, respectively. ......................................................... 9
Figure 2.5: The mechanical and electrical goniometer for joint angles measurement, respectively. .......... 9
Figure 2.6: Rigid-body joint angle measurement using accelerometers and gyroscopes .......................... 10
Figure 2.7: Surface EMG sensors to record the muscles activity. ............................................................. 11
Figure 2.8: Functional gait phases and events within a full gait cycle. Initial contact (IC), loading
response (LR), midstance (MSt), terminal stance (TSt), preswing (PSw), initial swing (ISw),
midswing (MSw) and terminal swing (TSw). The time is expressed in percentage of a gait cycle.......... 12
Figure 2.9: IDEFA module for monitoring physical activity using inertial sensors. ................................ 14
Figure 3.1: The experimental design and methodology of the proposed system ...................................... 20
Figure 3.2: The dual-belt instrumented treadmill. ..................................................................................... 21
Figure 3.3: The Myomonitor wireless EMG systems. ............................................................................... 21
Figure 3.4: The Trigno wireless EMG systems. ........................................................................................ 22
Figure 3.5: The placement of the surface EMG sensors. ........................................................................... 22
Figure 3.6: EMG data processing. (A) Row EMG data, (B) band pass filtered EMG Data, (C) Full-
wave rectified EMG data, and (D) Linear envelope of EMG data, X-axis is the time in seconds and
Y-axis is amplitude in volts. ...................................................................................................................... 25
Figure 3.7: A granule in the universe of discourse (U) ............................................................................. 27
xii
Figure 3.8: The design of fuzzy information granule (fuzzy set) A of type triangular. a and b are the
left and right bounds of the support of A, and m is the mode of A. ........................................................... 28
Figure 3.9: A multi-layer feedforward neural network. ............................................................................ 33
Figure 3.10: a LabVIEW application for acquisition, processing, and visualization of multiple
sensors data. ............................................................................................................................................... 34
Figure 4.1: The placement of the sensors. ................................................................................................. 37
Figure 4.2: 3-D GRFs for healthy and MS subjects within full gait cycle. ............................................... 40
Figure 4.3: 3-D accelerations of foot, shank, thigh, and hip for healthy and MS subjects within full
gait cycle. ................................................................................................................................................... 41
Figure 5.1: GRFs parameters. .................................................................................................................... 49
Figure 5.2: Classification accuracy for all classes using NNC. ................................................................. 53
Figure 5.3: Features selection order and the corresponding classification accuracy. ................................ 54
Figure 6.1: LIMA smart goniometer design and methodology ................................................................. 59
Figure 6.2: The circuit design of a single axis goniometer. ....................................................................... 60
Figure 6.3: A servo motor, used for system calibration. ........................................................................... 61
Figure 6.4: A printed circuit board (PCB) consists of tri-axis goniometers. ............................................. 62
Figure 6.5: Knee joint angles in sagittal plane (KneeY) and transverse plane (KneeZ) measured by
LIMA smart goniometer and compared to a reference optical system (VICON). .................................... 63
Figure 7.1: (A) The original data points of GRF-Z and (B) The fuzzy rule-base of GRF-Z within the
seven gait phases. ....................................................................................................................................... 66
Figure 7.2: two data series f1 and f2 with the same mean. ......................................................................... 66
Figure 7.3: The experimental setup and data acquisition using LIMA smart goniometer, wireless
EMG, and instrumented treadmill. ............................................................................................................ 68
xiii
Figure 7.4: (A) GRFs, (B) electrical muscles activities EMG, and (C) joint angles of the reference
represented by the mean±std of 20 able-bodied subjects. ......................................................................... 70
Figure 7.5: Ankle angles in 3D, two groups of reference, A and B, are shown for the coronal and the
transverse planes. ....................................................................................................................................... 71
Figure 7.6: The series of fuzzy granules are constructed from vertical GRF Fz, tibialis
anterior muscle signal TA, and ankle sagittal angle AnkleY, using window size (wsize) of 5, 10, 14,
and 20, respectively. .................................................................................................................................. 72
Figure 7.7: A comparison between CP patient and the able-bodied reference (normal) for (A) GRFs,
(B) EMG, and (C) ankle, knee, and hip joint angles in 3D, within the full gait cycle. ............................. 74
Figure 7.8: A comparison between MS patient and the able-bodied reference (normal) for (A) GRFs,
(B) EMG, and (C) ankle, knee, and hip joint angles in 3D, within the full gait cycle. ............................. 79
Figure 7.9: A comparison between CD patient and the able-bodied reference (normal) for (A) GRFs,
(B) EMG, and (C) ankle, knee, and hip joint angles in 3D, within the full gait cycle. ............................. 84
1
Chapter 1: Introduction
1.1 Background and Significance
According to a study from World Health Organization (WHO) (2006), neurological
disorders are an important cause of mortality and constitute 12% of total deaths globally, they
constitute 16.8% of the total deaths in lower middle income countries compared with 13.2% of
the total deaths in high income countries. Moreover, neurological disorders contribute to 92
million DALYs, an estimate of Disability-Adjusted Life Years, in 2005 projected to increase to
103 million in 2030. Thus, neurological disorders constitute 6.3% of the global burden of
disease, compared with other diseases such as HIV/AIDS and malignant neoplasm each
constitute slightly over 5% of total burden.
The effect of neurological disorders can be mitigated by applying an efficient
rehabilitation. WHO defines rehabilitation as an active process by which those affected by injury
or disease achieve a full recovery or, if a full recovery is not possible, realize their optimal
physical, mental and social potential and are integrated into their most appropriate environment
(WHO, 2001). The rehabilitation strategy determined by a process called Rehab-CYCLE (Stucki
et al., 2002). It involves four steps: assessment, assignment, intervention and evaluation. An
effective neurorehabilitation is based on the involvement of expert and multidisciplinary
assessment, realistic and goal-oriented programmes, and evaluation of the impact on the patient’s
rehabilitation achievements (WHO, 2006).
Recently, gait analysis has become a crucial assessment tool in clinical rehabilitation
process. It provides new insights to help understand various human movement patterns
corresponding to different gait pathology. Gait analysis is used to provide quantitative
information about the mobility state of disorder in order to help in the prescription of treatment
2
and the assessment of its outcome (Simon, 2004), and to allow the selection from amongst
treatment options including the possibility of not intervening (Baker, 2006). The reasons for
performing clinical gait analysis are: diagnosis between disease entities; assessment of the
severity, extent or nature of a disease or injury; monitoring of the progress in the presence or
absence of intervention; and prediction of the outcome of intervention or the absence of
intervention (Brand, 1987; Brand et al., 1981; and Baker, 2006).
Gait analysis involves the measurements of dynamic gait kinematics, kinetics, and
muscles activity, electromyography (EMG). Brand (1987) proposed a number of criteria for
useful biomechanical measurements in clinical gait analysis: reproducible; stable (independent
of mood, motivation and pain); accurate; appropriately validated; capable of distinguishing
between normal and abnormal, also between the characteristics of one disease entity and another;
must not alter the function it is measuring; reported in form analogous to accepted clinical
concepts; cost-effective; and not-observable by the skilled clinician
According to Simon (2004), current limitations of gait analysis are variability and lack of
reproducibility of gait measurements due to technical factors or subjective clinical interpretation,
in addition to, the time and expense of gait studies where the most common gait study protocol
uses the standard instruments of optical motion system in a standard manner. These limitations
can be corrected according to the same study by the following:
Increasing subject testing efficiency. An alternative method for makers based optical
motion system can reduce the time and the cost of experimental setup and
preparations.
Computer assisted gait data analysis and report generation. Development a tool for
objective assessment and rapid production of a clinical report. Moreover, eliminating
3
the subjectivity in current gait analysis reports and enables to establish standards for
interpretation of gait analysis data.
Analytical techniques for gait assessment. New nontraditional techniques such as
fuzzy logic and artificial intelligence are to be applied for representation and analysis
of gait data.
1.2 Motivation
Neurological disorders affect the human locomotion system. The effect of neurological
disorders can be decrease significantly by an efficient rehabilitation. Recently, gait analysis has
become a crucial assessment tool in clinical rehabilitation process. Gait analysis is a prominent
method for: diagnosis between disease entities, assessment of the severity or nature of a disease
or injury, monitoring of the progress in the presence or absence of intervention, and prediction of
the outcome of treatments. Gait analysis involves the measurements of dynamic gait kinematics,
kinetics, and EMG. Therefore, there is a need for an automated gait assessment tool to
manipulate the massive kinematic, kinetic, and EMG gait data in 3D, and hence, providing an
objective and quantitative assessment of gait disorders. This algorithmic tool may be helpful to
doctors and clinicians in the diagnosis of pathological gait impairments, prescribe treatment, and
assess the improvements in response to therapeutic intervention.
1.3 Goals and Specific Aims
The overall goal of this research is to investigate the application of fuzzy granular
computing and wearable inertial sensors for the analysis of human dynamic behavior in 3D
space. The specific aims of this research are:
4
Develop a flexible, cost-effective and reliable system for the measurement of joint angles
in 3D.
Analyze the gait data using a computational intelligence system based on fuzzy granular
computing and multi-sensor data fusion algorithms.
Facilitate an objective and quantitative assessment of functional gait impairment.
Explore the potential for eventual application in clinical human gait analysis.
1.4 Scope
The scope of this study includes the development of wearable inertial sensory system and
a computational intelligence system based on fuzzy granular computing and multi-sensor data
fusion.
1.5 Hypothesis
We hypothesize that the application of fuzzy granular computing to the fusion of multi-
sensor data facilitates the analysis of the dynamic behavior of human in 3D space.
5
Chapter 2: State of the Art in Human Motion Analysis
2.1 Summary
Recently, different technologies have been used to acquire gait data consist kinematic,
kinetic, and muscles activities. The predominant method for human motion analysis is the optical
motion system. The optical motion systems record human 3D motion using high-speed cameras
and reflective markers attached to human’s body, although it requires expensive equipment, large
space, and extrapolation of the obscuration due to limb movements. An alternative method is
wearable inertial sensors. Gait kinematics data in 3D can be conveniently collected using tri-
axial inertial sensors. The inertial sensors are small, low cost, and provide a direct measurement
of 3D gait kinematics data. Moreover, wearable sensors tolerate measurements outside the
laboratory. Electromyography (EMG) is a method for measurement the dynamic neuromuscular
activity.
Current studies have used different techniques to develop an automated tool for human
motion analysis. These techniques have been implemented for three major interests: detection of
gait phases, monitoring of daily physical activities, and representation and classification of gait
data.
2.2 State of the art technologies
2.2.1 Optical motion system
The predominant method for human motion analysis is the optical motion system
(Karaulovaa et al., 2002; Gavrila and Davis, 1996; Holt et al., 2000). The optical motion systems
record human 3D motion using high-speed cameras and reflective markers attached to human’s
6
body. The displacements of body segments and joints angle in 3D are measured by tracking the
positions of markers for each frame in the recorded video, whereas 4 to 8 cameras are required
for 3D motion. Moreover, many commercial optical motion analysis systems such as Vicon
(Vicon Motion System Inc., USA, www.vicon.com ) shown in Figure 2.1, and SIMI (SIMI
Reality Motion Systems GmbH, Germany, www.simi.com ) use inverse dynamics to estimate the
forces and torques (moments of forces) in body joints. Inverse dynamics is a method for
calculating the joint forces and torques based on the motion of body segments (kinematics),
body’s properties such as mass and external forces such as ground reaction forces (GRFs).
Figure 2.1: Vicon® Motion System, record human motion by tracking reflective markers using
high-speed cameras.
However, there are number of limitations to this technique: the optical motion analysis
system requires expensive equipment and a relatively large amount of space; it requires data
processing times that preclude clinical application in real time; and limb movements cause
obscuration of the reflective markers, requiring reliance on extrapolation by auto-tracking
paradigms. Furthermore, to determine the accelerations required for the forces, differential
7
calculations of the position and angle trajectories of the body segments are involved. This
differentiation introduces unwanted distortions and noise.
2.2.2 Wearable inertial sensors
An alternative method for measuring gait kinematics data is using body-worn inertial
sensors. The data in 3D can be conveniently collected using a tri-axial inertial sensors system
consists of accelerometers to measure body segments accelerations, gyroscopes to measure
angular velocity/rotational speed. Micro-electromechanical (MEMS) inertial sensors and electro-
goniometers are shown in Figure 2.2.
Figure 2.2: MEMS inertial sensors, accelerometer and gyroscope, shown respectively.
The MEMS inertial sensors are small, lightweight, portable, and low cost. Moreover,
wearable sensors are useful because they tolerate measurements in free-space and it is not limited
to the laboratory and could be performed in simplified settings or while performing practical
tasks (Kavanagh and Menz, 2008). The application of inertial sensors in human motion analysis
has a significant advantage by providing direct measurement of 3D gait kinematics data.
Recently, researchers have developed wearable sensory systems for human motion
analysis based on miniature wearable inertial sensors (Bonato, 2005; Giansanti et al., 2005;
Jasiewicz et al., 2006; Luinge and Veltink, 2005; Mayagoitaia et al., 2002; Pappas et al., 2001;
Sarkodie-Gyan et al., 2011; Selles et al., 2005; Takeda et al., 2009; Tao et al., 2009). Data fusing
of different wearable sensors have been implemented for quantitative human motion analysis
(Tao et al., 2009, Sarkodie-Gyan et al., 2011).
8
The orientation of body segments can be calculated in 3D using sensor fusing algorithms
and inertial measurement unit (IMU) consists of tri-axial accelerometer, gyroscope, and/or
magnetometer. Different Kalman filters have been developed for estimating the orientation of
body segments by fusing inertial sensors output (Luinge et al., 2004; Rehbinder et al., 2004; Zhu
et al., 2004; Luinge et al., 2005). Moreover, many commercial products have been developed by
industry for the same purpose using wireless IMU and sensor fusion algorithms performed by
special software. Figure 2.3 depicts wireless motion tracker (MTw) (xsens, Netherlands) and
Shimmer (shimmer research, Dublin, Ireland). MTw is a 9 degree of freedom (DoF) IMU
consists of tri-axial accelerometer, gyroscope, and magnetometer. MTw calculates the 3D
orientation, acceleration, angular velocity, and earth-magnetic field using. In a similar manner,
Shimmer 9DoF module provides the static and dynamic orientation.
Figure 2.3: 9 DoF IMU units, MTw and shimmer, shown respectively.
A major drawback of calculating the orientation of body segments using sensor fusing
algorithms is the inaccuracy in the estimated orientation resulted from the offset errors in the
angles calculated by angular velocity integration due to gyroscope drift, and the errors in the
inclination estimate for segments with large accelerations, like shank, as the accelerometers have
a slow response and sensitive to linear accelerations (Luinge et al., 2004; Rehbinder et al., 2004).
In addition to the magnetometer output can be affected by an existing magnetic field.
9
2.2.3 Goniometers
The most common sensors for measurement of joint angles are rotary encoders include
magnetic and optical rotary encoders (Kikuchi et al., 1997; Kojima et al., 2004). The rotary
encoders, presented in Figure 2.4, include two parts, one at each joint segment, and record the
rotation angle between them based on multiple magnet in case of magnetic encoder, LED light
source emitted through a disk and detected by photo detector in case of optical encoders. A
major drawback of rotary encoders is that the installation is difficult or impossible for such joints
as human body joints. In addition, optical encoders are expensive.
Figure 2.4: The rotary encoders, magnetic and optical, respectively.
Figure 2.5 depicts the conventional mechanical and electrical goniometers used for
human joint angle measurement. This type of goniometers is inflexible and unreliable as it
requires an installation at the joint center and tight coupling between the two terminals which
may alter the joint motion. It also suffers from a drift.
Figure 2.5: The mechanical and electrical goniometer for joint angles measurement, respectively.
10
An alternative to the conventional goniometers is using MEMS accelerometers and
gyroscopes. The inertial sensors exhibit the advantages of low-cost, flexibility and reliability as
the installation does not require tight coupling between around the joint. Although, the accuracy
of inertial sensors is usually less than the optical encoders for the reasons stated in the previous
section.
Four different methods for body joint angle measurement using accelerometers and
gyroscopes were presented an analyzed in a survey (Cheng et al., 2010). The four methods,
common mode rejection (CMR), CMR with gyro-integration (CMRGI), and CMR with gyro-
differentiation (CMRGD), and distributed CMR (DCMR), were reported in ten studies
(Ghassemi et al., 2008; Ghassemi et al., 2008; Dejnabadi et al. 2006; Dejnabadi et al., 2005;
Williamson et al., 2001; Ohtaki et al., 2001; Mayagoitia et al., 2002; Kurata et al., 1998;
Willemsen et al, 1991; Dong et al., 2007). All the described methods rely on the property of
rigid-body kinematics for joint angle measurement. As a result, this would lead to inaccuracy of
measurement in case of non-rigid human body because of the dynamic movement and the
installed positions of sensors (Cheng et al., 2010).
Figure 2.6: Rigid-body joint angle measurement using accelerometers and gyroscopes
11
2.2.4 Surface electromyography (EMG)
Electromyography (EMG) is a method for measurement the dynamic neuromuscular
activity. Surface EMG sensors are illustrated in Figure 2.7. EMG data provides important
information regarding muscle recruitment patterns and neuromuscular control of walking.
Consistent phase-specific patterns of muscle activation occur during gait cycling at preferred
walking speeds in normal adults (Arsenault et al., 2001; Winter et al., 1991). The EMG behavior
associated with the dynamics of gait can be used for diagnostic assessment and treatment
decisions (Granata et al., 2004). EMG has been used in biomedical applications for identifying
neuromuscular disease, assessing lower back pain, biofeedback of prosthetic devices, and in
sports applications.
Figure 2.7: Surface EMG sensors to record the muscles activity.
2.3 State of the art techniques
Recent research on developing wearable inertial sensors for human motion analysis
intends to develop an automated motion assessment tool to manipulate the massive kinetic and
kinematic gait data in a real-time and provide an objective and quantitative assessment. Current
studies have developed different techniques for quantitative human gait analysis. These
techniques have been implemented for three major interests: detection of gait phases (Jasiewicz
et al., 2006; Pappas et al.,2001; Selles et al., 2005; Tao et al., 2009), monitoring of daily
physical activities (Lau et al., 2008; Parkka et al., 2006; Salarian et al., 2007), and representation
12
and classification of gait data (Ismail and Asfour, 1999; Jonesa et al. 2006; Kohle and Merkl,
2000; Mezghani et al., 2008; Takahashi et al., 2004).
2.3.1 Detection of gait phases
It has been shown in the literature that a full gait cycle has seven phases/ eight events
(Perry, 1992; Ayyappa, 1997; and Pasparkis et al., 2009). Figure 2.8 illustrates the functional
gait phases/ events within a full gait cycle.
Figure 2.8: Functional gait phases and events within a full gait cycle. Initial contact (IC), loading
response (LR), midstance (MSt), terminal stance (TSt), preswing (PSw), initial swing (ISw),
midswing (MSw) and terminal swing (TSw). The time is expressed in percentage of a gait cycle.
A gait cycle/stride is defined as the time interval between the heel strike of one foot and
the next heel contact of the same foot. Each gait cycle is divided into consecutive phases/periods
measured in percent of the gait cycle. It is divided into two main phases: stance phase starts from
initial contact/ heel strike of one foot and ends at the same foot toe-off, and swing phase starts
from the end of stance phase and ends at the initial contact again.
IC LR MSt TSt PSw ISw MSw TSw
0~10% 10%~30% 30%~50% 50%~60% 60%~70% 70%~85% 85%~100%
Stance Phase Swing Phase
13
The stance phase lasts for 60% of the gait cycle, approximately, and is subdivided into
four phases as follows (Perry, 1992): loading response (0~10%) starts from the heel strike/ initial
contact of one foot( reference foot) and ends at the other foot toe-off; mid stance (10~30%) starts
from the end of loading response and ends when the body weight is aligned over the forefoot;
terminal stance (30~50%) starts from the end of the mid stance and ends at the opposite foot
initial contact; pre swing (50~60%) starts from the end of terminal stance and ends at the
reference foot toe-off.
The swing phase lasts for the remaining 40%, approximately, and is subdivided into three
phases as follows: initial swing (60~70%) starts from the toe-off even and ends when the
maximum knee flexion occurs; mid swing (70~85%) starts from the end of the initial swing and
ends when the tibia is in vertical position; terminal swing (85~100%) starts from the end of the
mid swing and ends at the reference foot heel strike.
The use of inertial sensors, accelerometers and gyroscopes, to detect gait phases has been
investigated in (Tao et al., 2009; Jasiewicz et al., 2006; Pappas et al., 2001; Selles et al., 2005).
Tao et al. (2009) reported that all eight gait events can be detected for healthy subjects using
wearable inertial sensors system. In the developed system, three single-axis gyroscopes and one
two-axis accelerometer were attached to the foot, shank, and thigh segments. A calibration
process was implemented by fusing data of gyroscopes and accelerometers. Jasiewicz et al.
(2006) proves that the toe-off and heel-strike events can be correctly indentified using foot linear
accelerations or foot angular velocity in both normal and spina-cord injured subjects. Similarly,
Selles et al. (2005) detected the same gait events using shank accelerations in both healthy and
transtibial amputees subjects. Pappas et al. (2001) designed a gait phase detection system to
detect stance, heel-off, swing, and heel-strike gait events using a gyroscope to measure the
14
angular velocity of the foot and three force sensitive resistors in both healthy and impaired gait
subjects.
2.3.2 Monitoring of daily physical activities
Monitoring of physical activities has been investigated for healthy subjects by Lau et al.
(2008); a small sensor unit compromising an accelerometer and a gyroscope was developed to
detect motions of segments, the captured kinematic data were used to classify five walking
conditions: stair ascent, stair descent, level ground, upslope and downslope. Classification of
walking, running, and cycling conditions has been carried out by Parkka et al. (2006). The
physical activities for Parkinson’s disease patients were monitored by Salarian et al. (2007),
where periods of sitting and standing were detected using three inertial sensors units: one sensor
unit was attached to the trunk and included a bi-axial accelerometer and a gyroscope, the two
other sensors were attached to the shanks and each consisted of a gyroscope. Additionally, a
commercial product, Intelligent Device for Energy Expenditure and Activity (IDEFA®)
(Minisun, California, USA) (www.minisun.com ), Figure 2.9, has been developed to monitor
physical activity using a portable data logger and five inertial sensors.
Figure 2.9: IDEFA module for monitoring physical activity using inertial sensors.
15
2.3.3 Representation and classification of gait data
Several studies have investigated different techniques for the representation and
classification of data for the analysis of gait. Such techniques include: statistical analysis,
mathematically transforms, and machine learning classifiers, e.g. support vector machines
(SVM), nearest neighbor classifier (NNC), artificial neural networks (ANN), and fuzzy logic.
Statistical methods are the most widely applied and understood for gait analysis. The
most common adopted methods to represent kinetic data (dynamic data) in gait are the statistical
methods (Lai et al., 2009; Begg et al., 2005a; Takahashi et al., 2004; Gok et al., 2002; and
Lafuente et al., 1997). Statistical technique was performed in which the representative features
were selected based on the peak levels of the GRFs and their corresponding time of occurrence.
This approach has the benefit of selecting the most significant data points and hence reducing the
massive data extracted from the gait kinetics (dynamics) in 3D. Drawbacks of this approach are
that it is sensitive to noise and it assumes that all subjects’ data within the same class exhibit
similar patterns and similar features. To cope with this issue, other studies used the
mathematically transforms such as Fast Fourier Transform (FFT) and Discrete Wavelet
Transform (DWT) (Mezghani et al., 2008; Kohle and Merkl, 2000; and Ismail and Asfour,
1999). In these methods, the transforms compute approximation coefficients that are used as
feature vectors to represent gait data. An advantage of this technique is that a massive data
reduction is achieved without the need to a selection process of specific data points as performed
by statistical methods. A limitation of the mathematical transforms approach based on its
frequent application to univariate signal and little guidelines on selection of wavelet basis for gait
data (Chau, 2001b).
Recently, machine learning techniques have been used to represent and analyze gait data.
16
This approach demonstrates the ability to capture patterns and to model the complex non-linear
relationships in gait data. Several machine learning classifiers were investigated for the
automatic recognition of gait patterns including: SVM (Lai et al., 2009; Begg et al. 2005a; and
Begg et al. 2005b); NNC (Mezghani et al.,2008 and Wang et al., 2003); ANN (Kohle and Merkl,
2000; Hanson et al., 2009; Parkka et al., 2006; Wu et al., 2001); Fuzzy logic (Salarian et al.,
2007; O’Malley et al., 1997; Yu et al. 2010). ANN was found to be the most prevalent non-
traditional methodology for gait analysis recently (Simon, 2004 and Chau, 2001b). However, the
performance of ANN depends on internal learning parameters, weights and biases, which are
difficult to interpret. Fuzzy classifier has been used by Salarian et al. (2007) to provide an
objective quantification of physical activities in both normal and Parkinson’s disease patients.
O’Malley et al. (1997) adopted fuzzy clustering to identify five clusters for 88 children with
cerebral palsy. Yu. et al. (2010) developed a fuzzy rule-based reasoning algorithm for
recognizing the activation patterns of the lower extremity muscles during normal walking within
the seven gait phases. Relational matrices were developed to represent a rule-base then a
comparison between the reference rule-base and the actual measurements of a subject was
implemented through fuzzy inferential reasoning system. The output of the system expressed
through similarity measure provides a quantitative assessment of the muscle activities of the
subject during gait cycle.
The applications of machine learning techniques for gait data have been evaluated in
many reviews (Simon, 2004; Chau, 2001a; Chau, 2001b; Dobson et al., 2007; and Preece et al.,
2009). Preece et al. (2009) investigated the different classification algorithms which have been
used to classify normal activities to identify falls from body-worn sensor data. Although few
investigated comparisons between different techniques suggested that ANN may give the highest
17
classification accuracy, the differences are often small. SVM have shown promise in small pilot
studies but have yet to be tested in large scale studies. Furthermore, recent work in combining
different classifiers has been examined in the review. It has been found that the performance was
improved by combining the output of different classifiers. Dobson et al. (2007) presented a
systematic review examining 18 studies for gait classification in children with cerebral palsy
(CP). Half of the studies used qualitative pattern recognition method, whereas the remaining
studies used quantitative method based on cluster analysis techniques. The study found that most
classifications were constructed using only sagittal plan gait data, only one study used gait data
in all three plans, which may limit classifications validity and restrict the applications.
Although, many techniques have been investigated for gait classification, the fact that
different gait pathologies have different gait patterns implies that more investigation and work
are necessary in this field.
18
Chapter 3: Experimental Design and Methodology
3.1 Summary
The proposed system includes the acquisition, processing, and analysis of gait data in 3D
space. The data acquisition system consists of an instrumented treadmill for GRFs, surface EMG
system for muscles activity, and wearable inertial sensors for gait kinematics. The analysis of
data was performed using intelligent systems involves the methods of feature extraction, multi-
sensor data fusion, knowledge base system, and fuzzy-based inference system. In addition,
machine learning algorithms such as artificial neural networks (NN) and nearest neighbor
classifier (NNC) have been investigated for the classification of gait impairments. Two groups of
subjects, able-bodied and impaired, were recruited to create the knowledge base. A comparison
between the reference rule-based data, able-bodied, and an input test data, impaired, was
evaluated using fuzzy granular computing and fuzzy algorithm. A similarity measure was used
to provide a quantitative assessment of mobility state of the impaired subjects.
3.2 Participants
This study was approved by the Institutional Review Board (IRB) of the University of
Texas at El Paso (UTEP). All subjects are asked to sign an informed consent form prior to
participation. Able-bodied adult subjects of age 18 or above were recruited from campus of
University of Texas at El Paso (UTEP) and community in El Paso. Patients with neurological
disorders were recruited from the local hospitals and rehabilitation clinics (Paul L. Foster School
of Medicine at Texas Tech University Health Sciences Center and MENTIS
NeuroRehabilitation).
19
The anthropometric data were recorded for each subject. The anthropometric
measurements including; gender, age, height, weight, and Body Mass Index (BMI), are required
for the purpose of group classification of subjects. Furthermore, the length, breadth, and
circumference of both lower and upper extremity segments are required for determination the
mass of body segments involved in the calculation of forces (Winter, 2009).
3.3 Experimental design
Figure 3.1 illustrates the experimental design and methodology of the proposed system.
The development of the system includes the acquisition, processing, and analysis of gait data.
The gait kinematics, muscles activity, and GRFs are acquired, processed, and analyzed using an
intelligent system. The proposed intelligent system involves the methods of feature extraction,
multi-sensor fusion, knowledge base system, and a fuzzy-based inference system. In addition,
machine learning algorithms such as artificial neural networks (NN) and nearest neighbor
classifier (NNC) have been investigated for the classification of gait impairments.
The proposed system provides an automated gait assessment tool that may be useful to
the clinicians in the identification of pathological gait impairments, in conjunction with a
quantitative assessment of mobility state of the impaired subject.
20
Figure 3.1: The experimental design and methodology of the proposed system
3.4 Data acquisition
The data acquisition system consists of three main hardware devices: instrumented
treadmill for GRFs, surface EMG system for muscles activity, and wearable inertial sensors for
gait kinematics.
3.4.1 Instrumented treadmill
Each subject was asked to walk on instrumented treadmill (Bertec Corporation, USA) at a
self-selected natural speed continuously for three minutes. The self-selected natural speed may
be a subjective speed selection (comfortable speed) by the subjects. The instrumented treadmill
is a dual-belt type with two independent force plates mounted beneath the belts as shown in
Figure 3.2. Each force plate consists of strain gauge load transducers that precisely measure
GRFs in three directions (FX: mediolateral, FY: anterioposterior, and FZ: vertical).
Gait Kinematics
( Inertial Sensors )
Muscles Activity
( Surface EMG )
GRFs
( Instrumented
Treadmill )
Data
Processing
Intelligent System
Features
Extraction
Knowledge Base
Fuzzy-based
Inference
System
Classification Quantitative
Assessment
Data Acquisition
Multi-Sensors
Data Fusion
21
Figure 3.2: The dual-belt instrumented treadmill.
3.4.2 Surface electromyography (EMG)
Figure 3.3 illustrates the experimental design for EMG data acquisition and the
placement of surface EMG electrodes. Tow wireless surface EMG systems were used to measure
the dynamic muscles activities for both side of the lower extremity, Myomonitor® and Trigno®
wireless EMG systems (Delsys Inc., Boston, MA, USA) (www.delsys.com ). Myomonitor
system, shown in Figure 3.3, has 16 channels acquire the EMG data at sampling rate of 1 kHz.
Trigno system, shown in Figure 3.4, consists of 16 wireless sensors, each sensor has an EMG
sensor for EMG acquisition at 2 kHz sampling rate and tri-axial accelerometer.
Figure 3.3: The Myomonitor wireless EMG systems.
22
Figure 3.4: The Trigno wireless EMG systems.
One or two muscles of each muscle group are usually selected to represent the whole
group during walking. The following eight muscles for each side were selected as illustrated in
Figure 3.5: soleus (Sol), tibialis anterior (TA), gastrocnemius lateralis (LG), vastus lateralis
(VL), rectus femoris (RF), biceps femoris (BF), gluteus medius (Gmed), and erector spinae (ES).
Figure 3.5: The placement of the surface EMG sensors.
23
3.4.3 Wearable inertial sensors
The developed wearable inertial sensors system consists of miniature analog inertial
sensors, accelerometers and gyroscopes, as well as analog to digital (A/D) converter. The
accelerometer ADXL335 (Analog Devices Inc.) is a small, low power, and tri-axis accelerometer
measures the acceleration with a full-scale range of ±3g. It can measure the static acceleration of
gravity and the dynamic acceleration resulting from motion. Two MEMS gyroscopes were used
to measure the angular rate in 3D. LPR530AL (STMicroelectronics Inc.) is a dual-axis analog
gyroscope capable of measuring angular rate along pitch and roll axes with a range of ±300°/s, as
well as, LY530ALH (STMicroelectronics Inc.) is a single-axis analog gyroscope capable of
measuring angular rate along yaw axis with a range of ±300°/s. A 12-bit A/D converter
USB6008 (National Instruments Inc.) was used to acquire the analog sensors output and converts
it in digital format.
3.5 Data processing
The data processing and filtering methods were performed based on the type of acquired
data. Three types of raw data were recorded for each subject; GRFs, EMG, and kinematics data.
3.5.1 Ground Reaction Forces (GRFs)
The GRFs data have been acquired at frequency of 100Hz and filtered using 20Hz low
pass filter. The GRFs amplitudes are normalized based on body mass (Lai et al., 2009; Lafuente
et al., 1997; and Begg et al., 2005b). The gait cycle/stride time and phases are determined based
on the vertical component of GRFs FZ (Pasparkis et al., 2009). Time-normalization of strides is
accomplished by re-sampling and expressing each stride in percentage rather than time (Winter,
2009). In addition, the GRFs amplitudes were normalized based on the subject’s body weight.
24
3.5.2 Surface EMG
Figure 3.6 illustrates the EMG data processing steps.
Filtering:
As most of the EMG data are concentrated in the band between 20Hz and 200Hz (Winter,
2009), a band pass filter (20Hz-200Hz) has been applied to the raw EMG data to reduce the
noise effects such as motion artifact noise.
Full-wave rectification:
The filtered data were then full-wave rectified to generate the absolute value of the EMG.
Linear envelope:
Linear envelope is a common way to manipulate EMG signal (De-Luca, 2006). It is also
known as moving average. The linear envelope is produced by applying a second order
Butterworth low pass filter with a cutoff frequency of 7Hz to the full-wave rectified EMG signal.
Normalization:
EMG Amplitude normalization is performed because of the variability between
individuals such as the inherent physiological variability and the variability associated with
electrode placement (Allison et al., 1993). The normalization is carried out for each subject
based on the peak or the mean of the EMG signal (Ricamato and Hidler, 2005). Time-
normalization is also performed for the strides time similar to the GRFs.
25
Figure 3.6: EMG data processing. (A) Row EMG data, (B) band pass filtered EMG Data, (C)
Full-wave rectified EMG data, and (D) Linear envelope of EMG data, X-axis is the time in
seconds and Y-axis is amplitude in volts.
3.5.3 Kinematics
The output of inertial sensors has been acquired at 100Hz sampling rate. A low pass filter
with cutoff frequency of 6 Hz is typically applied to lower the noise and improve the resolution
of the accelerometers (Winter, 2009). Time-normalization is applied for the strides time as well.
26
3.6 Intelligent systems and soft computing techniques
Soft computing methodology is essential to the design and implementation of intelligent
systems (Zadeh, 1998). Fuzzy logic is a principal component of soft computing. It exploits the
tolerance for imprecision and uncertainty to achieve tractability, robustness, and low solution
cost (Zadeh, 1994).
Fuzzy logic has been investigated and applied in the proposed intelligent system for the
purpose of gait assessment and classification. Fuzzy logic exhibits the benefits of: 1) reducing
the massive data extracted from gait kinetics and kinematics in 3D, 2) easy to interpret and
understand, 3) offering insight into non-linear relationships among gait variables, 4) providing
quantitative comparison, 5) less complexity and fast processing time, and therefore, offering a
possibility for real time applications.
According to Zadeh (1998), fuzzy logic has two core concepts: First, Fuzziness/
Fuzzification. Fuzziness is a condition which relates to sets (classes) whose boundaries are
unshapely defined, while fuzzification refers to replacing a crisp set with a set whose boundaries
are fuzzy. The second concept is Granularity/ Granulation. Granularity relates to clumpiness of
structure while granulation refers to partitioning an object into a collection of granules. A
combination of fuzzification and granulation is referred by fuzzy granulation (Zadeh, 1997).
3.6.1 Fuzzy granular computing
Zadeh (1997) defined granulation as “Granulation of an object A results in a collection of
granules of A, with a granule being a clump of objects (or points) which are drawn together by
indistinguishability, similarity, proximity or functionality” as illustrated in Figure 3.7. Typically,
a granule is a fuzzy set.
27
Figure 3.7: A granule in the universe of discourse (U)
Granular computing (GrC) covers theories, methodologies, techniques, and tools that make
use of granules, i.e., groups, classes, or clusters of a universe, in the process of problem solving
(Yao, 2000). In Granular computing (GrC), the granules may be crisp or fuzzy (Zadeh, 1997 and
1998). Crisp granules (c-granular) play important roles in a wide variety of methods, approaches
and techniques such as interval analysis, rough set theory, decision tree, cluster analysis,
Dempster-Shafter theory, etc.
Based on Zadeh’s (1997), a granule is defined and constructed by a generalized
constraint:
G= {X| X isr R}
where, X is a variable takes values in a universe of discourse U, R is the constraining relation, isr
is a variable copula and r is a discrete variable whose value defines the way in which R
constrains X. type of constraints are equality, possibilistic, veristic, probabilistic, and fuzzy
constraints. The type of granules is determined by type of constraint.
Fuzzy information granules
The formation of fuzzy information granules is based on the following principle; given a
collection of numeric data confined to the granulation window W, construct an information
granule (fuzzy set) A based on a certain family of fuzzy set (membership function is triangular,
Universe
Granule
28
parabolic, etc.) such that it is experimentally highly legitimate and retains high specificity
(Bargiela et al., 2003 and Pedrycz et al., 2002).
To construct such fuzzy information granule (fuzzy set) A, two conflicting requirements
should be taken into account:
1) Experimental evidence requirement: the fuzzy information granule A should embrace
enough experimental data, to be experimentally valid.
2) Specificity requirement: the support (size) of A should be as small as possible, to be
specific enough.
For a fuzzy information granule A shown in Figure 3.8, the above requirements can be
satisfied by maximizing the sum of membership grades,
Maximize ∑
and minimizing the support of A,
Minimize supp(A)= b - a
Figure 3.8: The design of fuzzy information granule (fuzzy set) A of type triangular. a and b are
the left and right bounds of the support of A, and m is the mode of A.
An optimization method combining the two requirements based on (Pedrycz et al., 2002
and Yu. et al., 2009) suggests to maximize the performance index Q defined as,
∑
A
a bm x
29
The Optimization of two parameters a and b attained by two performance indices Q(a)
and Q(b) for both left and right sections of A, respectively, as follows:
∑
(1)
∑
(2)
where m is the mode of fuzzy set A that can be determined by,
Minimize ∑ | | (3)
k1 and k2 are the numbers of data points located at the left section [a , m] and the right
section [m , b] of A, respectively.
Yu and Pedrycz (2009) proved that for a triangular membership function the maximum of
Q(a) is attained at ∑ . In a similar manner, b can be calculated
for the right section.
Although the granulation process is affected by the size and the type of granule
(membership function), experimental evidence indicates that different membership functions
have no significant impact on the performance of granulation (Yu and Pedrycz, 2009).
The granulation window size can be selected based on a tradeoff between specificity and
generality. Larger granulation window leads to more general representation of the data, yet, it
reduces the level of detailed description. Although the performance index Q exhibits a
monotonic increase with respect to the increase of granulation window for some data series, the
behavior of Q is based on the nature of data series, generally (Yu and Pedrycz, 2009). Therefore,
an experimental adjustment of granulation window is suggested.
30
Fuzzy granulation similarity
Assessment of functional impairments in human locomotion using fuzzy granulation
approach is presented and justified in our previous paper (Alaqtash et al., 2012). The proposed
method is described briefly in the following steps:
Step 1: Rescaling:
Because of the variability between individuals, e.g. walking speeds variability,
physiological variability, and the variability associated with sensors placement. Gait data may
differ by speed and by intensity. To reduce the effect of this difference, two normalization
methods, described in the previous section, are used. First, re-scaling stride time so that it is
measured in terms of gait cycle percentage reduces the effect of different motion speed. Second,
re-scaling the signal amplitude such that
where, is the scaled signal, is the minimum value of the original signal , while, is
determined based on the nature of data, e.g. the body weight for GRFs and the signal mean for
EMG.
Step 2: Fuzzy granules:
After re-scaling, fuzzy granule series are built as follows:
Given two data series, and .
Segment the data series X into segments, each segment has a number of data points specified
by a window size wsize and expressed as of k segments.
Construct a fuzzy granule g for each segment W and determine the three parameters (ai, mi,
bi) for each using Equations 1, 2, and 3.
31
Build a fuzzy granule series of p granules and each is expressed by
represent three parameters (ai, mi, bi) of the fuzzy granule .
Similarly, build a fuzzy granule series of p granules for the data series .
Step 3: Degree of Similarity:
The degree of similarity (DS) between two fuzzy granule series, G and H, is defined as
follows (Yu et al., 2005):
∑ ∑
∑ ∑
(4)
where, gh = min {g , h}, g h = max {g , h}.
The fuzzy granulation similarity measure offers a comparison between two series of data
that may represent a reference (able-bodied) data and a test (patient) data.
3.6.2 Fuzzy Relational matrix (Rule-base)
In our previous papers (Alaqtash et al., 2011; Sarkodie-Gyan et al., 2001; Yu et al.,
2010), we proposed a fuzzy-based semi-heuristic method for gait assessment using fuzzy
relational matrices, rule-base, and fuzzy similarity algorithm.
Fuzzy Sets and relational matrix (Rule-base)
In Fuzzy set theory, an element (member) x of a set X is represented by a “membership
function” µX(x). The degree of membership shows the grade of belonging of the member to its
set and has a real value in the interval [0, 1].
Fuzzy relations represent and quantify the associations among the elements of Fuzzy sets.
Let x an element of a set X expressed as where },...,,{ 21 nxxxX and y an element of a
32
set Y expressed as where },...,,{ 21 myyyY . The fuzzy relation R may be obtained using
the membership function described as µR(x,y). Fuzzy relational matrix of size n×m can be
developed as follows:
[
]
The described relational matrix depicts a rule-based system representing the strength of
association or interaction amongst the elements of Fuzzy sets.
Fuzzy relational matrix, rule-base, can be used to represent the strength of association or
interaction amongst the elements of gait functions, and it depicts a rule-base that can be used to
provide a model of feature matrix.
Fuzzy rule-base similarity
The fuzzy rule-base similarity algorithm offers a comparison between the reference rule-
base and a test rule-base as depicted by the equation.
),(),,(max
),(),,(min*
yxyx
yxyx
testref
testref
testref
testref
testref
where * represents the Fuzzy cross-correlation operator, (min) represents the Fuzzy logic
intersection, (max) represents the Fuzzy logic union, and the μref * μtest is the grade of
similarity between the test rule-base and the reference rule-base, where the grade of similarity
ranges between 0 and 1.
33
3.6.3 Neural Networks (NN)
NN is among the most effective learning methods for learning to interpret complex real-
world sensor data (Mitchell, 1997). A multi-layer feedforward neural network shown in Figure
3.9 is one of the most common ANNs (Haykin, 1999). A multi-layer feedforward neural network
has been a standard method for various applications including gait analysis (Chau, 2001b). It
consists of interconnected set of simple nodes, neurons, distributed in a hidden layer or layers.
The inputs are mapped to the nodes through an input layer and the outputs are controlled by
transfer function within each node and by adjusting the weights of links between nodes. The
learning process in neural networks consists of adjusting the weights through training in which
the actual output approximately matches the desired output. Although NNs can have multiple
hidden layers, research suggests that there should be one hidden layer with a number of nodes
less than input nodes (Simon, 2004).
Figure 3.9: A multi-layer feedforward neural network.
.
.
.
.
.
.
.
.
.
Hidden Layer
Output Layer
Input Layer
34
3.7 Software development
The software has been developed for real-time data acquisition and processing, as well as
analysis of gait parameters based on an intelligent system. The development of the system has
been achieved through NI LabVIEW, MATLAB®.
NI LabVIEW development system is a graphical programming environment used to
develop measurement and control systems. It offers integration with hardware data acquisition
devices and built-in libraries for advanced data analysis, processing and visualization. The
LabVIEW platform is scalable across multiple targets and OSs. Figure 3.10 shows a LabVIEW
application for acquisition, processing, and visualization of multiple sensors data.
MATLAB® is a high-level technical computing language and interactive environment
for algorithm development, data analysis, 2D and 3D data visualization, and numeric
computation. MATLAB can be used for data processing and analysis and integrated with other
programming languages and applications.
Figure 3.10: a LabVIEW application for acquisition, processing, and visualization of multiple
sensors data.
35
Chapter 4: Application of Wearable Sensors for Human Gait Analysis using
Fuzzy Computational Algorithm
4.1 Summary
This work has been published in (Alaqtash et al., 2011). A wearable inertial sensor
system has been developed and tested for the acquisition of gait features. The sensors were
placed on anatomical segments of the lower limb: foot, shank, thigh, and hip, and the motion
data were then captured in conjunction with 3D ground reaction forces (GRFs). The method of
relational matrix was applied to develop a rule-based system, an intelligent fuzzy computational
algorithm. The rule-based system provides a feature matrix model representing the strength of
association or interaction amongst the elements of the gait functions (limb-segments
accelerations and GRFs) throughout the gait cycle. A comparison between the reference rule-
based data and an input test data was evaluated using a fuzzy similarity algorithm. This system
was tested and evaluated using two subject groups: 10 healthy subjects were recruited to
establish the reference fuzzy rule-base, and 4 relapsing remitting multiple sclerosis subjects were
used as an input test data; and the grade of similarity between them was evaluated. This
similarity provides a quantitative assessment of mobility state of the impaired subject.
This chapter is organized as follows: Section 4.2 describes the methods and materials
including subjects, data acquisition and processing, fuzzy rule-based relational matrix, and fuzzy
similarity algorithm. The experimental results and discussion are presented in Section 4.3.
Section 4.4 illustrates the conclusion.
36
4.2 Methods and materials
4.2.1 Subjects
14 male subjects were recruited for the purpose of the study: 10 healthy adults (age
26.2±5.2 years, weight 71.5±12.6 kg, height 172.5±9.5 cm, and BMI 23.9±2.6 kg/m2), and 4
relapsing remitting multiple sclerosis (RRMS) patients (age 43.5±14.5 years, weight 103.7±27.6
kg, height 178.9±13.3 cm, and BMI 32.6±8.6 kg/m2). Table 4.1 illustrates the anthropometric
data for all subjects.
Table 4.1: Anthropometric data for 14 male subjects.
Subject
Age
(years)
Weight
(kg)
Height
(cm)
BMI
(kg/m2)
Healthy
subjects
1 29 88 192 23.9
2 31 88.3 174 29.2
3 38 70.8 171 24.2
4 25 56.6 157.5 22.8
5 21 61.2 163 23.0
6 22 67 176 21.6
7 23 71.6 171 24.5
8 24 66.6 170 23.0
9 26 56.6 169 19.8
10 23 88.1 181 26.9
Mean± std 26.2±5.2 71.5±12.6 172.5±9.5 23.9±2.6
MS
subjects
1 63 77.7 162 29.6
2 37 130.7 181 39.9
3 45 124.3 178 39.2
4 29 82.3 194.5 21.8
Mean± std 43.5±14.5 103.7±27.6 178.9±13.3 32.6±8.6
MS: Multiple Sclerosis
BMI: Body Mass Index = Weight (kg) / (Height (m) × Height (m) )
37
4.2.2 Data Acquisition and Processing
A wearable inertial sensor system was developed to acquire the kinematic data
(accelerations) of body segments. The system consists of analog MEMS accelerometers
ADXL330 (Analog Devices Inc) and 12-bits A/D (National Instruments Inc). The ADXL330 is a
low power and tri-axis accelerometer. An array of eight accelerometers was used for the lower
extremity and placed on both right and left sides at the foot, shank, thigh and hip. Figure 4.1
depicts the placement of the sensors. The motion data (accelerations) of body segment in three
directions were then captured in conjunction with 3D GRFs measured using an instrumented
treadmill (Bertec Corporation, USA). The output of the sensors was acquired at a frequency of
100 Hz, a 6 Hz second-order Butterworth low pass filter was applied to lower the noise and
improve the resolution of the accelerometers. In a similar way to GRF data processing, the
acceleration data were averaged and normalized across all strides. The filtered data were
averaged and normalized across all strides.
Figure 4.1: The placement of the sensors.
38
4.2.3 Fuzzy rule-based relational matrix
A fuzzy rule-based system was developed provides representing the strength of
association or interaction amongst the elements of the gait functions (limb-segments
accelerations and GRFs) throughout the gait cycle.
Three fuzzy sets were developed to represent 3D GRFs, gait phases, and accelerations.
}{, ZYX FFFXXx
,Yy Y={phase1 phase2 phase3 phase4 phase5 phase6 phase7};
,Zz Z={FootX FootY FootZ ShankX ShankY ShankZ ThighX ThighY ThighZ HipX HipY HipZ}
The fuzzy relational matrix may be obtained using the membership function described as
n
i
p in
yxyxP1
)(1
),(),(
where ),( yxp depicts the membership function, n is the number of data samples in a fuzzy set,
and )(i is the normalized data of that fuzzy set from 0 to 1. The membership function ),( yxp
represents the mean of the data samples of gait function within each gait phase. Two fuzzy
relational matrices were developed. The first relational matrix was defined as P(x,y) – the
relation between the GRF data and gait phases, the second relational matrix was defined as
Q(z,y) – the relationship between accelerations and gait phases.
P(x,y) = [
]
Q(z,y) = [
]
39
4.2.4 Fuzzy similarity algorithm
A comparison between the reference rule-based data and an input test data was evaluated
using a fuzzy similarity algorithm.
),(),,(max
),(),,(min
*
yxyx
yxyx
testref
testref
testref
testref
testref
This similarity provides an evaluation methodology for determining the behavior of each
individual force or acceleration in a corresponding gait phase.
4.3 Experimental results and discussion
The proposed methodology was tested and evaluated using two subject groups: 10
healthy subjects were recruited to establish the reference fuzzy rule-base, and 4 relapsing
remitting multiple sclerosis subjects were used as an input test data.
Figure 4.2 depicts the 3-D GRFs for both healthy and MS subjects within the full gait
cycle, whilst Figure 4.3 illustrates the 3-D accelerations of the lower extremity segments for both
healthy and MS subjects within the full gait cycle.
40
Figure 4.2: 3-D GRFs for healthy and MS subjects within full gait cycle.
41
Figure 4.3: 3-D accelerations of foot, shank, thigh, and hip for healthy and MS subjects within
full gait cycle.
Utilizing the fuzzy relational matrices described above, a rule-base, relational matrix, was
established to represent the feature matrix of acquired data within gait cycle for both healthy and
MS subjects. The reference rule-based matrix Pref(x,y) was established to represent the GRFs of
42
healthy subjects, whilst the test matrix Ptest(x,y) was established to represent the GRFs of MS
subjects. The fuzzy similarity between Pref(x,y) and Ptest(x,y) was determined. An example of the
relational matrices, Pref(x,y) and Ptest(x,y) for individual subject, and the similarity matrix
between them is expressed as:
Pref(x,y) * Ptest(x,y) =
[
] [
]
= [
]
Table 4.2 exhibits the mean and standard deviation (std) of the rule-based matrices,
Pref(x,y) and Ptest(x,y) for all subjects, and the grade of similarity between them within the gait
cycle. Similarly, the mean and std of the rule-based matrices representing the segmental
acceleration, Qref(z,y) and Qtest(z,y), and the grade of similarity between them are shown in Table
4.3.
Table 4.2: Mean(std) of the rule-based matrices Pref(x,y) and Ptest(x,y) for healthy and MS
subjects respectively and the grade of similarity between them.
Phase 1 Phase 2 Phase 3 Phase 4 Phase 5 Phase 6 Phase 7
FX
Pref 0.626(0.089) 0.140(0.054) 0.125(0.054) 0.292(0.083) 0.652(0.100) 0.660(0.105) 0.660(0.105)
Ptest 0.748(0.048) 0.164(0.044) 0.092(0.079) 0.221(0.122) 0.602(0.265) 0.820(0.151) 0.925(0.077)
Sim. 0.836 0.850 0.738 0.759 0.923 0.805 0.713
FY
Pref 0.282(0.049) 0.168(0.044) 0.545(0.054) 0.921(0.032) 0.528(0.073) 0.474(0.039) 0.474(0.040)
Ptest 0.446(0.133) 0.192(0.159) 0.555(0.168) 0.826(0.151) 0.742(0.169) 0.557(0.132) 0.495(0.051)
Sim. 0.634 0.874 0.982 0.897 0.711 0.849 0.956
FZ
Pref 0.421(0.049) 0.869(0.045) 0.887(0.031) 0.722(0.073) 0.066(0.043) <.001(0.001) <.001(0.001)
Ptest 0.278(0.080) 0.805(0.089) 0.962(0.017) 0.803(0.044) 0.369(0.196) 0.101(0.141) 0.002(0.004)
Sim. 0.659 0.926 0.922 0.899 0.179 0.008 0.289
43
Table 4.3: Mean(std) of the rule-based matrices Qref(z,y) and Qtest(z,y) for healthy and MS
subjects respectively and the grade of similarity between them.
Phase 1 Phase 2 Phase 3 Phase 4 Phase 5 Phase 6 Phase 7
FootX
Pref 0.580(0.117) 0.588(0.142) 0.577(0.157) 0.646(0.153) 0.630(0.196) 0.455(0.207) 0.683(0.060)
Ptest 0.616(0.108) 0.571(0.064) 0.555(0.055) 0.640(0.178) 0.651(0.207) 0.474(0.176) 0.528(0.222)
Sim. 0.942 0.971 0.960 0.992 0.967 0.961 0.774
FootY
Pref 0.478(0.080) 0.761(0.040) 0.768(0.039) 0.762(0.044) 0.741(0.066) 0.910(0.028) 0.420(0.077)
Ptest 0.596(0.052) 0.746(0.032) 0.741(0.033) 0.732(0.040) 0.704(0.069) 0.753(0.157) 0.522(0.132)
Sim. 0.802 0.980 0.965 0.960 0.950 0.828 0.804
FootZ
Pref 0.326(0.071) 0.387(0.133) 0.371(0.132) 0.526(0.100) 0.848(0.075) 0.395(0.207) 0.604(0.077)
Ptest 0.475(0.147) 0.463(0.195) 0.464(0.169) 0.549(0.189) 0.640(0.248) 0.399(0.227) 0.576(0.148)
Sim. 0.687 0.835 0.800 0.958 0.755 0.989 0.953
ShankX
Pref 0.555(0.147) 0.314(0.084) 0.431(0.068) 0.459(0.144) 0.323(0.200) 0.502(0.191) 0.453(0.126)
Ptest 0.533(0.079) 0.285(0.168) 0.224(0.209) 0.328(0.276) 0.368(0.135) 0.344(0.197) 0.467(0.232)
Sim. 0.960 0.908 0.520 0.715 0.879 0.685 0.969
ShankY
Pref 0.577(0.137) 0.418(0.064) 0.616(0.072) 0.727(0.114) 0.588(0.160) 0.278(0.079) 0.227(0.076)
Ptest 0.628(0.064) 0.575(0.223) 0.602(0.245) 0.592(0.228) 0.508(0.275) 0.298(0.342) 0.194(0.100)
Sim. 0.919 0.727 0.978 0.815 0.863 0.935 0.853
ShankZ
Pref 0.669(0.069) 0.531(0.031) 0.554(0.055) 0.751(0.109) 0.731(0.099) 0.129(0.080) 0.586(0.135)
Ptest 0.739(0.103) 0.615(0.082) 0.619(0.081) 0.719(0.122) 0.721(0.166) 0.362(0.276) 0.493(0.138)
Sim. 0.905 0.862 0.896 0.958 0.986 0.356 0.842
ThighX
Pref 0.590(0.206) 0.355(0.199) 0.340(0.180) 0.302(0.194) 0.581(0.241) 0.757(0.173) 0.704(0.130)
Ptest 0.399(0.125) 0.340(0.033) 0.163(0.060) 0.139(0.072) 0.341(0.266) 0.543(0.241) 0.742(0.146)
Sim. 0.675 0.958 0.481 0.461 0.588 0.718 0.949
ThighY
Pref 0.712(0.105) 0.239(0.096) 0.310(0.069) 0.421(0.115) 0.357(0.106) 0.272(0.067) 0.352(0.103)
Ptest 0.785(0.062) 0.260(0.071) 0.188(0.028) 0.192(0.155) 0.182(0.164) 0.238(0.068) 0.385(0.131)
Sim. 0.907 0.920 0.605 0.456 0.508 0.873 0.915
ThighZ
Pref 0.585(0.119) 0.465(0.096) 0.383(0.042) 0.495(0.143) 0.661(0.099) 0.578(0.135) 0.132(0.059)
Ptest 0.462(0.042) 0.338(0.095) 0.272(0.105) 0.285(0.152) 0.401(0.259) 0.467(0.235) 0.283(0.231)
Sim. 0.791 0.726 0.710 0.576 0.606 0.809 0.467
HipX
Pref 0.698(0.109) 0.299(0.116) 0.181(0.131) 0.375(0.165) 0.573(0.218) 0.634(0.125) 0.668(0.131)
Ptest 0.648(0.204) 0.450(0.229) 0.256(0.332) 0.392(0.305) 0.579(0.181) 0.702(0.083) 0.678(0.198)
Sim. 0.928 0.666 0.707 0.957 0.990 0.903 0.985
HipY
Pref 0.743(0.114) 0.448(0.101) 0.582(0.082) 0.581(0.095) 0.235(0.141) 0.640(0.150) 0.766(0.090)
Ptest 0.703(0.056) 0.236(0.078) 0.322(0.186) 0.316(0.206) 0.178(0.158) 0.348(0.309) 0.663(0.198)
Sim. 0.946 0.526 0.553 0.545 0.760 0.544 0.865
HipZ
Pref 0.691(0.099) 0.338(0.119) 0.310(0.087) 0.505(0.146) 0.466(0.109) 0.321(0.111) 0.206(0.076)
Ptest 0.672(0.188) 0.331(0.155) 0.305(0.183) 0.411(0.272) 0.249(0.191) 0.411(0.130) 0.354(0.305)
Sim. 0.972 0.980 0.982 0.813 0.534 0.781 0.583
44
Based on the recorded results, significant outcomes can be observed. The vertical
component of GRF, FZ, shown in Figure 4.2 depicts that the MS subjects have a significant
longer stance phase. Moreover, the contact force was relatively flat with a single peak force as
shown in the same force curve, whereas the same curve of healthy subjects exhibits the M shape
with two peaks to reflect the weight transfer from the heel to the mid-foot and to the ball of the
foot for push-off. These differences were represented by the low grades of similarity for FZ
within the swing phases (Similarity = 0.179 for initial swing (phase5), 0.008 for midswing
(phase6), and 0.289 for terminal swing (phase7)), and the moderate similarities for both FY and
FZ within the loading response phase (phase1) (Similarity= 0.634 and 0.659, respectively) as
shown in Table 4.2.
In MS subjects, the long stance phase, and therefore lack of sufficient swing phase, is
caused by the abnormal muscle activities of shank (tibialis anterior), thigh (quadriceps), and hip
(gluteus medius). This abnormality also leads to insufficient accelerations (forces) of the thigh
and the hip as shown in Figure 4.3. The acceleration curves of the thigh and the hip, HipY,
demonstrate low value of accelerations for MS compared to healthy subjects throughout most of
the gait phases with some variations. This can be observed in Table 4.3 by the low grades of
similarity for the thigh within some gait phases (the lowest grades of similarity are 0.461 for
ThighX and 0.456 for ThighY within the preswing phase (phase4), and 0.467 for ThighZ within
the terminal swing phase (phase7)), and the moderate similarity for the anterioposterior
acceleration of Hip, HipY ,(the average similarity for 4 gait phases is 0.542).
The behavior of the foot acceleration depicts that the MS subjects exhibit sharp increased
acceleration at the initial contact and lack of foot clearance at the initial swing phase as shown in
Figure 4.3.
45
In view of these results, a quantitative assessment of the neurological state of the subject
can be evaluated. This algorithm may serve as an assessment tool for clinician and doctors to
gain substantial insight into the neurological state of gait and evaluate the outcomes of surgery
and/or therapy.
4.4 Conclusion
In this chapter, gait data including limb-segment accelerations and GRFs in 3D were
collected simultaneously during the experimentation. The collected data served as input into a
rule-based relational matrix built upon the principle of fuzzy computational algorithm. Fuzzy
similarity algorithm was used to perform the comparison between the reference rule-based data
and an input test data. The proposed methodology was tested and evaluated using two groups of
subjects; healthy subjects were recruited to establish the reference rule-base, and MS patients to
establish the input test rule-base, and the grade of similarity between them was evaluated (Tables
4.2 and 4.3). Significant differences were observed between the two groups with respect to limb-
segments accelerations and GRFs data throughout the gait phases.
This study demonstrates an efficient method of gait characterization utilizing wearable
sensors through building and comparing the reference rule-base of healthy subjects with an input
rule-base of impaired subjects using a fuzzy similarity algorithm. This similarity can provide a
quantitative assessment of the neurological state of the subject which may be useful for clinician
and doctors to identify pathological gait impairments, prescribe treatment, and assess the
improvements in response to therapeutic intervention.
46
Chapter 5: Automatic Classification of Pathological Gait Patterns using
Ground Reaction Forces and Machine Learning Algorithms
5.1 Summary
An automated gait classification method is developed in this study (Alaqtash et.al.,
2011b), which can be applied to analysis and to classify pathological gait patterns using 3D
ground reaction force (GRFs) data. The study involved the discrimination of gait patterns of
healthy, cerebral palsy (CP) and multiple sclerosis subjects. The acquired 3D GRFs data were
categorized into three groups. Two different algorithms were used to extract the gait features; the
GRFs parameters and the discrete wavelet transform (DWT), respectively. Nearest neighbor
classifier (NNC) and artificial neural networks (ANN) were also investigated for the
classification of gait features in this study. Furthermore, different feature sets were formed using
a combination of the 3D GRFs components (mediolateral, anterioposterior, and vertical) and
their various impacts on the acquired results were evaluated. The best leave-one-out (LOO)
classification accuracy 85% was achieved. The results showed some improvement through the
application of a features selection algorithm based on M-shaped value of vertical force and the
statistical test ANOVA of mediolateral and anterioposterior forces. The optimal feature set of six
features enhanced the accuracy to 95%. This work can provide an automated gait classification
tool that may be useful to the clinician in the diagnosis and identification of pathological gait
impairments.
Although, many techniques have been investigated for gait classification, the fact that
different gait pathologies have different gait patterns imposes that more investigation and work
are necessary in this field. The aim of this study is to develop an automated gait classification
47
tool, which can be used in gait analysis application to classify pathological gait patterns, using
3D GRFs data and by utilizing different machine learning algorithms. Particularly, the proposed
method was tested and evaluated using three groups of subjects: healthy, cerebral palsy (CP), and
multiple sclerosis (MS). CP and MS are neurological diseases causing gait impairments.
5.2 Methods and materials
5.2.1 Participants
Twenty subjects were recruited for the purpose of the study: Twelve healthy adults
(mean± std: age 27.1±5.9 years, weight 71.4±11.5 kg, height 171.6±8.3 cm, and body mass index
(BMI) 24.2±2.8 kg/m2), four spastic diplegic cerebral palsy patients (age 29.5±17.5 years, weight
67.8±17.9 kg, height 162.3±8.7 cm, and BMI 25.4±4.2 kg/m2), and four relapsing remitting
multiple sclerosis patients (age 50.3±11.5 years, weight 99.7±32.5 kg, height 167.8±14.5 cm,
and BMI 34.5±5.8 kg/m2).
Although the database is rather modest which may limit the generalization of extracted
features and then affect the classifications accuracy, this work represents a preliminary study for
investigation different machine learning algorithms that may be used for classification of
pathological gait impairments.
5.2.2 Data acquisition
An instrumented treadmill was used in this study (Bertec Corporation, USA) to measure
the GRFs in three directions (FX: mediolateral, FY: anterioposterior, and FZ: vertical) while
subject walking at a self-selected natural speed for three minutes continuously. The GRFs data
were acquired at frequency of 100 Hz and filtered using 20 Hz second order Butterworth low
pass filter. The GRFs amplitudes were normalized based on body mass. The stride time was
48
specified based on FZ and time-normalization was accomplished by re-sampling and expressing
each stride in percentage rather than time.
5.2.3 Feature extraction
Two different feature extraction approaches have been evaluated in this study.
Approach I - GRFs Parameters
A common approach to represent gait data is based on the amplitudes and temporal
parameters of data (Lafuente et al., 1997). The maximum and minimum GRF amplitudes and
their time of occurrence, in percentage, were computed for the three GRFs components. A total
of 19 features were extracted to represent each subject as shown in Figure 5.1. Within the three
minutes walking, the average value of each feature was calculated across all strides. Moreover,
left and right values were transformed into a common mode value represent the average value for
both sides, this done to avoid feature dependences on the laterality of the disease (Lafuente et al.,
1997).
Approach II - Discrete Wavelet Transform (DWT)
Wavelets are basis functions obtained from a prototype called mother function and it can
be used to approximate data. The DWT compute the approximation coefficients using digital
filtering to represent the feature vector for GRFs data. A mother wavelet function Symlet3 can be
used to calculate the feature vector for GRFs data (Mezghani et al., 2008). In this work, a mother
wavelet Symlet3 and a feature vector of length 16 coefficients were determined experimentally.
49
Figure 5.1: GRFs parameters.
5.2.4 Classification
Nearest Neighbor classifier (NNC)
The principle of nearest neighbor or k-nearest-neighbor algorithm is that the properties of
any particular input point x are likely to be similar to those of points in the neighborhood of x
(Russell et al., 2003). The neighborhood is defined to include k points, and a distance metric, e.g.
Euclidean distance, is used to identify the nearest neighbors of a query point. Nearest neighbor
algorithm is very simple to implement and often performs quit well (Russell et al., 2003). The
typical value of k varies from 1 to a small percentage of the training data and is selected using
trial and error (Preece et al., 2009).
TZ3TZ2TZ1
FY
(N/K
g)
FX
(N/K
g)
0 10 20 30 40 50 60 700
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Medio
late
ral fo
rce (
Fx)
0 10 20 30 40 50 60 700
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Ante
rioposte
rior
forc
e (
Fy)
Ver
tica
l fo
rce
(Fz)
FZ
(N/K
g)
Stride time (%)
Stride time (%)
Stride time (%)
FZ1 FZ2 FZ3
TSTEP
FY1
TY1
FY2
TY2
TY3
FY3
TX1
FX1
FX2
TX2
FX3
TX3
50
Artificial Neural Network (ANN)
ANN is an efficient learning method to interpret complex real-world sensor data
(Mitchell, 1997). A description of ANN is in Section 3.6.3. In this study, a multi-layer feed-
forward neural network of one hidden layer is used.
Classification evaluation and performance measures
Cross validation is a common method to evaluate the accuracy of classifiers (Duda et al.,
2001). The classifier is trained with training dataset which includes most of the subjects and then
tested with testing dataset which has few subjects. In Leave-One-Out (LOO) cross validation,
one subject is used for testing and the rest are used for training. The classification result is then
computed and repeated until all subjects have participated in the testing dataset. The overall
classification result is then computed as the average of all testing subjects.
A typical measure of classification algorithms performance is a confusion matrix
(Chawla et al., 2002). An example of a confusion matrix with two classes, positive and negative,
has four elements: TP, true positive, is the number of correctly classified positive instances; FN,
false negative, is the number of positive instances incorrectly classified as negative; TN, true
negative, is the number of correctly classified negative instances; and FP, false positive, is the
number of negative instances incorrectly classified as positive. The classification accuracy is
defined as the ratio of correctly classified instances to all instances.
Accuracy = (TP+TN) / (TP+FN+TN+FP)
Precision and recall are another typical classification performance measures defined as
(Chawla et al., 2002):
Precision = TP / (TP+FP)
Recall = TP / (TP+FN)
51
Precision is the ratio of correctly classified positive instances to both correctly and
incorrectly classified instances as positive. Recall is the ratio of correctly classified positive
instances to positive instances.
5.3 Experimental results and discussion
5.3.1 Feature sets
In order to evaluate the effect of each GRFs component on the classification accuracy,
different feature sets were formed using a combination of the three GRFs components. Seven
feature sets (FZ; FY; FX; FXFY; FYFZ; FXFZ; FXFYFZ) were examined.
A summary of LOO classification accuracy using different approaches and classifiers is
presented in Table 5.1. The highest achieved classification accuracy 85% corresponded to:
FXFYFZ using NNC and approach I (GRFs parameters); the same set of features using ANN and
approach II (DWT); FXFY using ANN and approach I. The results on a single force component
indicate that FX is the most discriminatory and FZ is the worst.
In general, the feature extraction approach I, GRFs parameters, depicts better accuracy
than DWT associated with NNC, whereas the result varies in the case of ANN. Both NNC and
ANN have the same range of classification accuracy among all feature sets.
Table 5.1: Summary of classification accuracy using the different approaches and classifiers
Classifier Features FZ FY FX FX,FY FY,FZ FX,FZ FX,FY,FZ
NNC
Approach I a 65% 70% 80% 75% 75% 75% 85%
Approach II b 65% 65% 70% 60% 65% 75% 60%
ANN
Approach I 70% 70% 75% 85% 75% 70% 80%
Approach II 60% 75% 80% 80% 70% 70% 85%
a GRFs parameters
b DWT
52
5.3.2 Features selection
In view of previous results, further investigation has been made to enhance the results by
applying effective features selection techniques. In order to select the most significant features,
GRFs parameters, it is important to investigate the contribution of each feature set to
classification for each class of gait patterns: healthy, CP, and MS. Figure 5.2 shows the LOO
classification accuracy for all classes using NNC.
Alaqtash et al. (2011) recorded that FZ of MS patients shows a significant longer stance
phase. Moreover, the contact force was relatively flat with a single peak force as shown in the
same force curve, whereas the same curve of healthy subjects exhibits the M shape with two
peaks to reflect the weight transfer from the heel to the mid-foot and to the ball of the foot for
push-off.
Although FZ provides full discrimination accuracy between healthy and pathological gait
pattern, FZ leads to lower classification accuracy for both pathological gaits CP and MS as
shown in Figure 5.2. Therefore, another representation of FZ has been investigated based on the
M shape (Takahashi et al., 2004). FZ force was represented as either M-shaped or non M-shaped.
M-shaped was defined as both ratios FZ2 / FZ1 and FZ2 / FZ3 are less than 0.9; others were defined
as non M-shaped. The number 0.9 was determined experimentally.
53
Figure 5.2: Classification accuracy for all classes using NNC.
As a result, using an input feature set consists of the complete FX and FY features sets and
one feature represents FZ (1: M-shaped, 0: non M-shaped), the classification outcomes are
recorded in Table 5.2. The result shows 100% for precision and recall measures of healthy
subjects. This result demonstrated that full discrimination between healthy and pathology gait
can be achieved using one feature, M-shaped, represents FZ in addition to the complete FX and FY
features sets. Moreover, the confusion matrix depicts that all misclassified instances are located
within the pathological gait classes. Therefore, further enhancement can be made if a certain
features selection technique applied for FX and FY sets.
Table 5.2: The classification performance measures using M-shaped, FX and FY features
A. Confusion matrix
CP MS Healthy
CP 2 2 0
MS 1 3 0
Healthy 0 0 12
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
FZ FY FX FX,FY FY,FZ FX,FZ FX,FY,FZ
25%
75%
0%
50% 50%
25%
50%
0%
50%
100%
50% 50% 50%
75%
100%
75%
100%
92% 92%
100% 100%
CP
MS
Healthy
Features Set
Cla
ssif
ica
tio
n A
ccu
ra
cy
54
B. Precision and recall
Precision Recall
CP 75% 50 %
MS 60% 75%
Healthy 1 1
The statistical significant difference test, analysis of variance (ANOVA), was employed in
this study to evaluate the effect of each feature on the difference between CP and MS classes.
The p-value depicts the significant difference between CP and MS classes corresponding to a
certain feature. The p-value was calculated for each feature in FX and FY sets and therefore
features were sorted based on p-value. Figure 5.3 demonstrates the features selection order and
the corresponding classification accuracy.
Figure 5.3: Features selection order and the corresponding classification accuracy.
It can be observed that the best LOO classification accuracy 95% was achieved using the
optimal feature set of six features; one feature from FZ : M-shaped; three from FX : FX1, TX3, and
FX3; two from FY: TY3 and TY2. The classification performance measures: confusion matrix,
precision, and recall, using the optimal feature set are shown in Table 5.3. The results clearly
55
indicate that the classification performance significantly increased using the optimal feature for
the three classes. The overall performance measures are 96% and 95% for precision and recall,
respectively.
Table 5.3: The classification performance measures using the optimal feature set
A. Confusion matrix
CP MS Healthy
CP 3 1 0
MS 0 4 0
Healthy 0 0 12
B. Precision and recall
Precision Recall
CP 100% 75%
MS 80% 100%
Healthy 100% 100%
Weighted Average a
96% 95%
a Average of 12 healthy, 4 CP, and 4 MS.
5.4 Conclusion
In this study, an automated gait classification tool was developed. A classification of
healthy, CP, and MS gait patterns, was performed using the three components of GRFs data (FX,
FY, and FZ) and by utilizing different machine learning algorithms. The proposed method was
tested and evaluated using three groups of subjects: healthy, CP, and MS. Two feature extraction
algorithms (GRFs parameters and DWT) and two classifiers (NNC and ANN) have been
56
investigated and the result was evaluated. Further enhancement on classification accuracy has
been made applying a features selection algorithm based on M-shaped value of FZ and the
statistical test ANOVA of FX and FY. The best LOO classification accuracy 85% was achieved
without features selection. An improvement of the result was achieved using features selection.
The accomplished classification accuracy was 95% corresponding to the optimal set of six
features.
This study has demonstrated some significant points: 1) a comparison between two
different feature extraction approaches depicts that GRFs parameters approach associated with
NNC has higher classification accuracy than DWT; 2) both NNC and ANN have the same range
of classification accuracy among all feature sets; 3) full discrimination between healthy and
pathology gait can be achieved using one feature, M-shaped, represents FZ in addition to the
complete FX and FY features sets; and 4) a features selection method can be used to enhance the
classification accuracy and the statistical test ANOVA is a possible technique for that.
This work can provide an automated gait classification tool that may be useful to the
clinician in the diagnosis and identification of pathological gait impairments.
57
Chapter 6: Cost-Effective Device for Measurement of Joint Angles (LIMA
Smart Goniometer)
6.1 Summary
In this chapter, a cost-effective sensor called LIMA smart goniometer has been developed
for the measurement of joint angles. LIMA smart goniometer use inertial gyroscopes and a
hardware circuit of operational amplifiers. The joint angles are derived by performing hardware
integration of the angular rates obtained from two gyroscopes attached to joint segments.
Normal walking trials of five able-bodied subjects were conducted and knee joint angle
was acquired using our LIMA smart goniometer and a reference optical system (VICON)
simultaneously. A comparison has made between the two systems. The result shows high
correlation coefficients (0.9834±0.0054 and 0.8651±0.0428) and low root mean square errors
(4.0762±0.7351 and 3.1968±1.1366). The proposed system provides a cost-effective and reliable
device for the measurement of joint angles in 3D.
6.2 Introduction
The most common sensors for measurement of joint angles are rotary encoders include
magnetic and optical rotary encoders. The rotary encoders include two parts, one at each joint
segment, and record the rotation angle between them based on multiple magnet in case of
magnetic encoder, LED light source emitted through a disk and detected by photo detector in
case of optical encoders. A major drawback of rotary encoders is that the installation is difficult
or impossible for such joints as human body joints. In addition, the optical encoders are
expensive. Conventional mechanical and electrical goniometers have been used for human joint
angle measurement. This type of goniometers is inflexible and unreliable as it requires an
58
installation at the joint center and tight coupling between the two terminals which may alter the
joint motion. It also suffers from a drift.
An alternative to the conventional goniometers is using MEMS accelerometers and
gyroscopes. The inertial sensors exhibit the advantages of low-cost, flexibility and reliability as
the installation does not require tight coupling around the joint. Four different methods for body
joint angle measurement using accelerometers and gyroscopes were presented an analyzed in a
survey (Cheng et al., 2010). The four methods, common mode rejection (CMR), CMR with
gyro-integration (CMRGI), and CMR with gyro-differentiation (CMRGD), and distributed CMR
(DCMR), were reported in ten studies. All the described methods rely on the property of rigid-
body kinematics for joint angle measurement. As a result, this would lead to inaccuracy of
measurement in case of non-rigid human body because of the dynamic movement and the
installed positions of sensors (Cheng et al., 2010).
Many recent studies have calculated the joint angles by estimation the orientation of body
segments using sensor fusing algorithms and inertial measurement unit (IMU) consists of tri-
axial accelerometer, gyroscope, and/or magnetometer. Different Kalman filters have been
developed for the same purpose ((Luinge et al., 2004; Rehbinder et al., 2004; Zhu et al. 2004;
Luinge et al. 2005).). A major drawback of calculation the orientation of body segments using
sensor fusing algorithms is the inaccuracy in the estimated orientation resulted from the offset
errors in the angles calculated by angular velocity integration due to gyroscope drift, and the
errors in the inclination estimate for segments with large accelerations, like shank, as the
accelerometers have a slow response and sensitive to linear accelerations (Luinge et al., 2004;
Rehbinder et al., 2004). In addition to the magnetometer output can be affected by an existing
magnetic field.
59
In general, previous work used sensor fusing algorithms and numerical integration
performed by software to evaluate the angles. However, it has been shown that joint angles
derived by software algorithms suffer from distortion caused by offsets and drifts due to the
nature of accelerometers and gyroscopes. In this work, joint angles were derived by performing
hardware integration of gyroscopes outputs using operational amplifiers.
6.2 Methods and materials
A cost-effective sensor called LIMA smart goniometer has been developed for the
measurement of joint angles. The design and methodology of LIMA smart goniometer are
illustrated in Figure 6.1.
Figure 6.1: LIMA smart goniometer design and methodology
The main function of the LIMA smart goniometer is based on the principle that joint
angles can be derived by integration the difference of angular velocities obtained from two
gyroscopes that are attached to joint segments as described in the equation below.
∫ ( )
where, is the joint angle, and are the angular velocities obtained from the two
gyroscopes.
A hardware integration of gyroscopes outputs is performed using operational amplifiers
(Op-Amps). The developed system consists of three main hardware devices; tri-axis analog
Tri-Axis
Gyro1
Tri-Axis
Gyro2
Differential
Amplifier
Integrator
Amplifier
Active
Filter3D Joint Angle
60
gyroscopes, common Op-Amp UA741, as well as, a 12-bit A/D converter, USB6008 (National
Instruments Inc.), to acquire the analog output and converts it in digital format. Since that an
analog tri-axis gyroscope within a single chip does not available in the market, two MEMS
analog gyroscopes were used to make a tri-axis analog gyroscope module. LPR530AL
(STMicroelectronics Inc.) is a dual-axis analog gyroscope capable of measuring angular rate
along pitch and roll axes with a range of ±300°/s, while, LY530ALH (STMicroelectronics Inc.)
is a single-axis analog gyroscope capable of measuring angular rate along yaw axis with a range
of ±300°/s.
The circuit design of a single axis goniometer is shown in Figure 6.2.
Figure 6.2: The circuit design of a single axis goniometer.
Angles are derived from single axis goniometer two gyroscopes by three phases:
61
Differentiation: subtraction the two angular velocities using an instrumentation
Op-Amp. Instrumentation Op-Amp is high gain differential amplifier which has
high input impedance.
Integration: Op-Amp Integrator is used to perform the mathematical operation of
Integration. The frequency of the integrator is 1.03 Hz.
Filtering: active high pass filter with a very low cutoff frequency 0.033 Hz is used
to remove the slight drift on the output caused by the inclination of body segments
(Tong et al., 1999).
A common Op-Amp, UA741, is used in the design and powered by +12 and -12 volts.
Both integrator and filter frequencies were selected experimentally for normal walking. The
calibration of the system has been performed using a servo motor, Figure 6.3. A controlled
rotation was used to calculate the scale factor. For different application, the frequency can be
adjusted easily by changing either the value of capacitor or resistor.
Figure 6.3: A servo motor, used for system calibration.
The circuit design has been implemented and replicated on a printed circuit board (PCB)
so that multiple tri-axis goniometers can be fitted on one board as shown in Figure 6.4.
62
Figure 6.4: A printed circuit board (PCB) consists of tri-axis goniometers.
6.3 Experimental results
To validate the proposed system, normal walking trials of five able-bodied subjects were
conducted and knee joint angle was acquired using our LIMA smart goniometer and a reference
optical system (VICON) for comparison simultaneously. Two planes were used to validate our
system, sagittal for flexion, extension angle, and transverse for interior/exterior angle. The output
signals for both systems were filtered using a low pass filter 3rd
order Butterworth at 6 Hz cutoff
frequency. Figure 6.5 shows the knee joint angles in sagittal plane (KneeY) and transverse plane
(KneeZ) measured by LIMA smart goniometer and compared to the same angles measured by a
reference optical system (VICON).
63
Figure 6.5: Knee joint angles in sagittal plane (KneeY) and transverse plane (KneeZ) measured by
LIMA smart goniometer and compared to a reference optical system (VICON).
The correlation coefficient (CORRCOEF) is used to measure the relationship between the
signals, while the root mean square error (RMSE) is used to measure the average difference
between the signals. The results of the comparison between LIMA smart goniometer and the
reference system represented by the mean and standard deviation (mean±std) of both
CORRCOEF and RMSE in Table 6.1.
Table 6.1: mean±std of correlation coefficient (CORRCOEF) and root mean square error
(RMSE) for knee angles in sagittal and transverse planes.
Knee Angle CORRCOEF RMSE
Sagittal 0.9834±0.0054 4.0762±0.7351
Transverse 0.8651±0.0428 3.1968±1.1366
0 50 100 150 200 250-10
0
10
20
30
40
50
60
70
Time (s)
Kn
ee
Y A
ng
le (
de
g)
0 50 100 150 200 250-20
-15
-10
-5
0
5
Time (s)
Kn
ee
Z A
ng
le (
de
g)
Reference (VICON)
Goniometer
64
The results show an excellent CORRCOEF 0.9834 and very low RMSE 4.0762 for the
sagittal angle, a high correlation 0.8651 and very low RMSE 3.1968 for transverse angles. The
difference on the correlation between the two angles is due to the fact that the joint angles during
walking depicts a large range of degrees and smooth shape in the sagittal plane, whereas, a low
range of degrees and wrinkled shape which decrease the correlation factor as shown in Figure
6.5, despite that, a very low difference still exists represented by RMSE.
6.4 Conclusion
A cost-effective and reliable device called LIMA smart goniometer has been developed
for the measurement of joint angles in 3D. LIMA smart goniometer used inertial gyroscopes and
a hardware circuit of operational amplifiers. The main function of the LIMA smart goniometer is
based on the principle that joint angles can be derived by integration the difference of angular
velocities obtained from two gyroscopes that are attached to joint segments.
The performance of LIMA smart goniometer has been validated by conduction
experiments of normal walking trials and recording knee joint angles using our LIMA smart
goniometer and a reference optical system (VICON) for comparison. The experimental results
show a high correlation and low differences between the two systems. LIMA smart goniometer
provides a cost-effective and reliable device for the measurement of joint angles in 3D.
65
Chapter 7: The Applications of Fuzzy Granular Computing for Assessment of
Functional Impairments in Human Locomotion
7.1 Summary
In this chapter, fuzzy granular computing algorithm has been used for the analysis of
human dynamic behavior in 3D space for the assessment of functional impairments. GRFS,
electrical muscles activity (EMG), and joint angles were collected from a group of twenty able-
bodied subjects as reference. The analysis of gait data using fuzzy granular computing were
applied for three case studies of three patients with different gait disorders. An evaluation of the
dynamic behavior and a quantitative assessment has been presented. The results show a high
efficiency of the proposed algorithm.
7.2 Introduction
A fuzzy similarity algorithm based on rule-base, relational matrices, has been used for
gait assessment in (Alaqtash et al., 2011; Sarkodie-Gyan et al., 2001; Yu et al., 2010). Although
it offers a comparison between the reference rule-base and a test rule-base within the seven gait
phases, it exhibits certain shortcomings; the fuzzy rule-base has a low specificity (the level of
detailed description is low) as illustrated in Figure 7.1. Most importantly, the fuzzy sets were
constructed based on a single value, the mean, whereas the mean does not reflect the variance in
data points, as a result inaccurate similarity is inferred. This can be clearly observed as shown in
Figure 7.2, two data series with the same mean value results in similarity equals to 1.
A fuzzy granulation algorithm is proposed to cope with these shortcomings by adjustable
granule window size, as well as, a degree of similarity measure based on three parameters
66
(a,m,b) for each granule. A theoretical justification for the efficiency of the proposed method has
been provided in (Alaqtash et al., 2012).
(A) (B)
Figure 7.1: (A) The original data points of GRF-Z and (B) The fuzzy rule-base of GRF-Z within
the seven gait phases.
Figure 7.2: two data series f1 and f2 with the same mean.
7.3 Methods and materials
7.3.1 Participants:
Twenty able-bodied adults were recruited for this study, 12 males and 8 females. Three
case studies have been presented for 3 females with different gait disorders, hemiplegic cerebral
palsy (CP), relapsing remitting multiple sclerosis (MS), and congenital dislocation of left hip
(CD). Table 7.1 illustrates the anthropometric data and the speed for all subjects.
0 10 20 30 40 50 60 70 80 90 1000
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1FP-Z
Gait Cycle [%]
Am
plit
ude [
%]
0 10 20 30 40 50 60 70 80 90 1000
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1FuzzyRB for FP-Z
Gait Cycle [%]
Am
plit
ude [
%]
x
f1
f2
67
Table 7.1: Anthropometric data and speed for twenty able-bodied subjects (12 males and 8
females), and a group of three females with different gait disorders.
Subject
Age
(years)
Weight
(kg)
Height
(cm)
BMI
(kg/m2)
Speed
(m/s)
Able-
bodied
subjects
MH01 32 90 192 24.41 1.2
MH02 38 88 174 29.07 1.0
MH03 23 60.5 167 21.69 0.99
MH04 24 65.1 178 20.55 1.0
MH05 31 84.3 171 28.83 1.0
MH06 26 76.3 175 24.91 1.2
MH07 26 77.9 168 27.60 1.0
MH08 19 81.4 185 23.78 1.15
MH09 22 70 163 26.35 1.05
MH10 24 83 178 26.20 1.0
MH11 25 81.1 173 27.10 1.1
MH12 20 61.7 167 22.12 1.0
FH01 25 53.1 155 22.10 1.0
FH02 20 64.5 170 22.32 1.0
FH03 27 63.3 168 22.43 1.0
FH04 20 40.8 162 15.55 0.81
FH05 25 54.8 155 22.81 0.85
FH06 23 55.9 159 22.11 1.0
FH07 20 57.8 155 24.06 0.95
FH08 40 62 165 22.77 1.2
Mean±std 25±4.49 67±12.95 167.55±8.45 23.72±3.26 1.03±0.10
Subjects
with gait
disorders
CP01 17 46.2 163 17.39 0.70
MS01 49 53 152 22.94 0.81
CD01 22 58.5 163 22.02 0.90
BMI: Body Mass Index = Weight (kg) / (Height (m) × Height (m))
MH denotes able-bodied males, FH for able-bodied females, CP01, MS01, and
CD01 refer to the subjects with gait disorders.
68
7.3.2 Data Acquisition
Joint angles, muscles activity (EMG), and GRFs were acquired simultaneously in this
experiment. A LIMA smart goniometer system has been developed to measure the joint angles
for the lower extremity of human body; ankle, knee and hip joints for both right and left sides.
Trigno® wireless EMG systems has been used to measure eight muscles signals for each; soleus
(Sol), tibialis anterior (TA), gastrocnemius lateralis (LG), vastus lateralis (VL), rectus femoris
(RF), biceps femoris (BF), gluteus medius (Gmed), and erector spinae (ES). GRFs in 3D has
been acquired using instrumented treadmill (Bertec Corporation, USA). Figure 7.3 depicts the
experimental setup and data acquisition.
Figure 7.3: The experimental setup and data acquisition using LIMA smart goniometer, wireless
EMG, and instrumented treadmill.
LIMA Smart
Goniometer V1.0
Developed by
LIMA 2012
Foot Gyro
Shank Gyro
Thigh Gyro
Hip Gyro
Instrumented
Treadmill
(GRFs)
EMG
Sensors
Joint Angles
Sensors
(Gyros)
69
Three minutes trials were conducted while subjects walk on instrumented treadmill at
their natural speed. Data were captured at 100 Hz sampling rate for joint angles and GRFs,
while, EMG acquired at 2 kHz. The data were processed as described in Section 3.5.
7.3.3 Methodology
The proposed fuzzy granulation algorithm and the degree of similarity measure were
described in details in Section 3.6.1. A series of fuzzy granules { } is
constructed to represent the reference (able-bodied) data. While is
constructed to represent patient’s data. Each fuzzy granule (fuzzy set) and are expressed as
and to represent the three parameters (ai, mi, bi) of a triangular fuzzy
set (membership function). The degree of similarity between the fuzzy granules has
been calculated for different granulation window sizes of 5, 10, 14, and 20.
7.4 Experimental results and discussion
7.4.1 Reference system
The average of 20 able-bodied subjects represents a reference for comparison. The
mean±std of GRFs, electrical muscles activities EMG, and Joint angles are shown in Figure 7.4.
(A)
0 20 40 60 80 100-4
-2
0
2
4
Gait Cycle (%)
FX (
N/k
g)
0 20 40 60 80 100-4
-2
0
2
Gait Cycle (%)
FY
(N
/kg
)
0 20 40 60 80 1000
5
10
15
Gait Cycle (%)
FZ
(N
/kg
)
mean
mean+std
mean-std
70
(B)
(C)
Figure 7.4: (A) GRFs, (B) electrical muscles activities EMG, and (C) joint angles of the
reference represented by the mean±std of 20 able-bodied subjects.
0 20 40 60 80 1000
1
2
3
Gait Cycle (%)
So
l
0 20 40 60 80 1000
2
4
Gait Cycle (%)
TA
0 20 40 60 80 1000
2
4
Gait Cycle (%)
GL
0 20 40 60 80 1000
2
4
6
Gait Cycle (%)
VL
0 20 40 60 80 1000
1
2
3
Gait Cycle (%)
RF
0 20 40 60 80 1000
2
4
6
Gait Cycle (%)
BF
0 20 40 60 80 1000
1
2
3
Gait Cycle (%)
GM
ed
0 20 40 60 80 1000
1
2
3
Gait Cycle (%)
ES
0 50 100-20
0
20
40
An
kle
X (
de
g)
Gait Cycle (%)
0 50 100-20
-10
0
10
20
An
kle
Y (
de
g)
Gait Cycle (%)
0 50 100-20
-10
0
10
An
kle
Z (
de
g)
Gait Cycle (%)
0 50 100-10
0
10
20
30
Kn
ee
X (
de
g)
Gait Cycle (%)
0 50 100-50
0
50
100
Kn
ee
Y (
de
g)
Gait Cycle (%)
0 50 100-20
-10
0
10
Kn
ee
Z (
de
g)
Gait Cycle (%)
0 50 100-15
-10
-5
0
5
Hip
X (
de
g)
Gait Cycle (%)
0 50 100-20
0
20
40
Hip
Y (
de
g)
Gait Cycle (%)
0 50 100-15
-10
-5
0
5
Hip
Z (
de
g)
Gait Cycle (%)
71
For the ankle, two groups of reference, denoted by A and B, are resulted for the coronal
and the transverse planes as shown in Figure 7.5.
Figure 7.5: Ankle angles in 3D, two groups of reference, A and B, are shown for the coronal and
the transverse planes.
Figure 7.6 shows the series of fuzzy granules are constructed from vertical
GRF Fz, tibialis anterior muscle signal TA, and ankle sagittal angle AnkleY, using window size
(wsize) of 5, 10, 14, and 20, respectively. As the size of granulation window increased fuzzy
granules embrace more data points, whereas, the specificity is decreased.
0 20 40 60 80 100-50
0
50
An
kle
X A
(de
g)
Gait Cycle (%)
0 20 40 60 80 100-20
0
20
An
kle
X B
(de
g)
Gait Cycle (%)
0 20 40 60 80 100-20
0
20
An
kle
Z A
(de
g)
Gait Cycle (%)
0 20 40 60 80 100-10
0
10
An
kle
Z B
(de
g)
Gait Cycle (%)
0 20 40 60 80 100-20
0
20
An
kle
Y (
de
g)
Gait Cycle (%)
72
Figure 7.6: The series of fuzzy granules are constructed from vertical GRF Fz,
tibialis anterior muscle signal TA, and ankle sagittal angle AnkleY, using window size (wsize) of
5, 10, 14, and 20, respectively.
0 50 100-2
0
2
4
6
8
10
12
Gait Cycle (%)
FZ
0 50 1000
0.5
1
1.5
2
2.5
3
3.5Fuzzy Granules g
i(a
i,m
i,b
i), wsize=5, p=20
Gait Cycle (%)
TA
0 50 100-5
0
5
10
15
20
25
An
kle
Y
Gait Cycle (%)
ai
mi
bi
0 50 100-2
0
2
4
6
8
10
12
Gait Cycle (%)
FZ
0 50 1000
0.5
1
1.5
2
2.5
3
3.5Fuzzy Granules g
i(a
i,m
i,b
i), wsize=10, p=10
Gait Cycle (%)
TA
0 50 100-5
0
5
10
15
20
25
An
kle
Y
Gait Cycle (%)
ai
mi
bi
0 50 100-2
0
2
4
6
8
10
12
Gait Cycle (%)
FZ
0 50 100-0.5
0
0.5
1
1.5
2
2.5
3
3.5Fuzzy Granules g
i(a
i,m
i,b
i),wsize=14, p=7
Gait Cycle (%)
TA
0 50 100-5
0
5
10
15
20
25
An
kle
Y
Gait Cycle (%)
0 50 1000
2
4
6
8
10
12
Gait Cycle (%)
FZ
0 50 1000
0.5
1
1.5
2
2.5
3
3.5
4Fuzzy Granules g
i(a
i,m
i,b
i), wsize=20, p=5
Gait Cycle (%)
TA
0 50 100-5
0
5
10
15
20
25
An
kle
Y
Gait Cycle (%)
ai
mi
bi
73
7.4.1 Gait Analysis - Case studies
Three case studies for three female subjects with gait disorders are presented; hemiplegic
cerebral palsy (CP), relapsing remitting multiple sclerosis (MS), and congenital dislocation of
left hip (CD). Both left and right sides of human body were compared to the normal reference of
twenty able-bodied subjects within the full gait cycle. GRFs, EMG, and 3D joint angles of ankle,
knee, and hip were analyzed using fuzzy granular computing and the degrees of similarity to the
normal reference are presented.
Case A: hemiplegic cerebral palsy (CP)
A 17-year old female patient suffers from hemiplegic cerebral palsy and had treatment in
the form of foot surgery.
Results:
(A)
0 20 40 60 80 100-5
0
5
Gait Cycle (%)
FX (
N/k
g)
0 20 40 60 80 100-2
-1
0
1
Gait Cycle (%)
FY
(N
/kg
)
0 20 40 60 80 1000
5
10
15
Gait Cycle (%)
FZ
(N
/kg
)
Normal
Right
Left
74
(B)
(C)
Figure 7.7: A comparison between CP patient and the able-bodied reference (normal) for (A)
GRFs, (B) EMG, and (C) ankle, knee, and hip joint angles in 3D, within the full gait cycle.
0 20 40 60 80 1000
2
4
Gait Cycle (%)
So
l
0 20 40 60 80 1000
2
4
Gait Cycle (%)
TA
0 20 40 60 80 1000
2
4
Gait Cycle (%)
GL
0 20 40 60 80 1000
5
Gait Cycle (%)
VL
0 20 40 60 80 1000
2
4
Gait Cycle (%)
RF
0 20 40 60 80 1000
5
Gait Cycle (%)
BF
0 20 40 60 80 1000
2
4
Gait Cycle (%)
GM
ed
0 20 40 60 80 1000
2
4
Gait Cycle (%)E
S
Normal
Right
Left
0 20 40 60 80 100-5
0
5
10
15
An
kle
X (
de
g)
Gait Cycle (%)
0 20 40 60 80 100-20
-10
0
10
20
An
kle
Y (
de
g)
Gait Cycle (%)
0 20 40 60 80 100-5
0
5
10
An
kle
Z (
de
g)
Gait Cycle (%)
0 20 40 60 80 100-5
0
5
10
15
20
Kn
ee
X (
de
g)
Gait Cycle (%)
0 20 40 60 80 100-20
0
20
40
60
80
Kn
ee
Y (
de
g)
Gait Cycle (%)
0 20 40 60 80 100-10
-5
0
5
10
Kn
ee
Z (
de
g)
Gait Cycle (%)
0 20 40 60 80 100-15
-10
-5
0
5
Hip
X (
de
g)
Gait Cycle (%)
0 20 40 60 80 100-30
-20
-10
0
10
20
Hip
Y (
de
g)
Gait Cycle (%)
0 20 40 60 80 100-10
-5
0
5
Hip
Z (
de
g)
Gait Cycle (%)
Normal
Right
Left
75
Table 7.2: The degree of similarity DS(G,H) between CP patient and the able-bodied reference
for (A) GRFs, (B) EMG, and (C) ankle, knee, and hip joint angles in 3D. DS is
calculated using different granulation window sizes (wsize) 5, 10, 14, and 20. The
mean and standard deviation of DS are reported. p is the number of fuzzy granules
representing the original data series.
(A)
wsize p FX FY FZ
Left Right
Left Right
Left Right
5 20
0.890 0.872
0.871 0.846
0.898 0.903
10 10
0.885 0.874
0.878 0.850
0.893 0.904
15 6
0.874 0.863
0.864 0.856
0.905 0.921
20 5
0.884 0.868
0.870 0.848
0.883 0.915
mean
0.883 0.869
0.871 0.850
0.895 0.911
std 0.007 0.005 0.006 0.005 0.009 0.009
(B)
wsize p Sol TA GL VL
Left Right
Left Right
Left Right
Left Right
5 20
0.764 0.721
0.608
0.529 0.655
0.714 0.755
10 10
0.743 0.709
0.615
0.547 0.682
0.729 0.743
15 6
0.754 0.711
0.600
0.529 0.654
0.688 0.757
20 5
0.738 0.745
0.603
0.508 0.655
0.673 0.747
mean
0.750 0.721
0.607
0.528 0.662
0.701 0.750
std 0.012 0.016 0.007 0.016 0.014 0.025 0.006
wsize p RF BF Gmed ES
Left Right
Left Right
Left Right
Left Right
5 20
0.593 0.769
0.514 0.657
0.710 0.560
0.583
10 10
0.609 0.771
0.503 0.663
0.729 0.552
0.588
15 6
0.582 0.766
0.454 0.616
0.723 0.546
0.542
20 5
0.562 0.757
0.470 0.629
0.723 0.586
0.533
mean
0.587 0.766
0.485 0.641
0.721 0.561
0.562
std 0.020 0.006 0.028 0.022 0.008 0.018 0.028
76
(C)
wsize p AnkleX AnkleY AnkleZ
Left Right
Left Right
Left Right
5 20
0.700 0.596
0.806 0.815
0.690 0.641
10 10
0.708 0.603
0.811 0.822
0.675 0.641
15 6
0.695 0.607
0.805 0.834
0.663 0.642
20 5
0.680 0.589
0.811 0.850
0.679 0.696
mean
0.696 0.599
0.808 0.830
0.677 0.655
std 0.012 0.008 0.003 0.015 0.011 0.027
wsize p KneeX KneeY KneeZ
Left Right
Left Right
Left Right
5 20
0.602 0.699
0.738 0.789
0.654 0.632
10 10
0.602 0.696
0.737 0.790
0.667 0.628
15 6
0.598 0.703
0.740 0.795
0.680 0.654
20 5
0.603 0.704
0.721 0.787
0.667 0.651
mean
0.601 0.700
0.734 0.790
0.667 0.641
std 0.002 0.004 0.009 0.004 0.011 0.013
wsize p HipX HipY HipZ
Left Right
Left Right
Left Right
5 20
0.839 0.692
0.872 0.719
0.741 0.617
10 10
0.836 0.687
0.872 0.719
0.748 0.610
15 6
0.834 0.687
0.868 0.715
0.744 0.612
20 5
0.846 0.692
0.872 0.715
0.715 0.628
mean
0.839 0.690
0.871 0.717
0.737 0.617
std 0.005 0.003 0.002 0.002 0.015 0.008
Discussion:
As shown in Table 7.2, the degree of similarity (DS) does not change significantly with
the change of granulation window size (wsize). Very small variation of DS among different
wsize values exists as represented by standard deviation. Therefore, as suggested in Chapter 3,
the granulation window size can be adjusted experimentally.
Although patient’s GRFs exhibit similar patterns to the reference of able-bodied group as
shown in Figure 7.7 (A) and represented by the high degree of similarity in Table 7.2 A,
significant outcomes can be observed. GRFs indicate slightly longer stance phase and a relatively
77
flat contact force for the vertical component FZ, whereas the reference of able-bodied group
exhibits a clear M shape with two peaks reflect the weight transfer from heel to mid-foot. In
addition, lower force in coronal plane FX is noticed for both right and left legs, this difference
may results from the slower speed for the patient. The average degree of similarity of 0.883 and
0.869 for left and right FX reflect this difference.
For the electrical muscles activity EMG represented by Figure 7.7 (B) and Table 7.2 (B),
the most obvious discrepancy is in the activity for most of the muscles during initial contact
phase; decreased of TA and ES, increased of Sol, GL, VL, RF, and Gmed, this may be a result of
toe walking. The other discrepancy is the increased of TA and GL during the swing phase, and
peak activity for BF at pre-swing phase. These differences are represented by low degree of
similarity presented in Table 7.2 (B). (e.g. 0.485 for left BF, 0.528 for GL, 0.562 for left ES, and
0.607 for TA)
In general, joint angles show a pattern similar to the reference. However, there are
substantial differences linked to the differences in muscles activity. From the sagittal plan data in
Figure 7.7 (C), it is evident that the patient’s left ankle has higher plantarflexion degree than
able-bodied one, this difference is a result of the changes in the muscles associated with muscles
associate with ankle as Sol and TA. Both knee and hip have lower minimum extension degree in
the sagittal plane. The hip joint angles demonstrate a significant difference in 3D, while right
side exhibits lower value than left. A clear demonstration of the changes in joint angles is
presented by the DS figures in Table 7.2 (C). e.g. HipY has DS of 0.717 for right and 0.871 for
left, HipZ has DS of 0.617 for right and 0.737 for left.
78
Case 2: relapsing remitting multiple sclerosis (MS)
Results:
(A)
(B)
0 20 40 60 80 100-5
0
5
Gait Cycle (%)
FX (
N/k
g)
0 20 40 60 80 100-2
-1
0
1
FY
(N
/kg
)
Gait Cycle (%)
0 20 40 60 80 1000
5
10
15
Gait Cycle (%)
FZ
(N
/kg
) Normal
Right
Left
0 20 40 60 80 1000
2
4
Gait Cycle (%)
So
l
0 20 40 60 80 1000
2
4
Gait Cycle (%)
TA
0 20 40 60 80 1000
2
4
Gait Cycle (%)
GL
0 20 40 60 80 1000
2
4
Gait Cycle (%)
VL
0 20 40 60 80 1000
2
4
Gait Cycle (%)
RF
0 20 40 60 80 1000
2
4
Gait Cycle (%)
BF
0 20 40 60 80 1000
2
4
Gait Cycle (%)
GM
ed
0 20 40 60 80 1000
2
4
Gait Cycle (%)
ES
Normal
Right
Left
79
(C)
Figure 7.8: A comparison between MS patient and the able-bodied reference (normal) for (A)
GRFs, (B) EMG, and (C) ankle, knee, and hip joint angles in 3D, within the full gait cycle.
0 50 100-20
0
20
40
An
kle
X (
de
g)
Gait Cycle (%)
0 50 100-20
-10
0
10
20
An
kle
Y (
de
g)
Gait Cycle (%)
0 50 100-20
-10
0
10
An
kle
Z (
de
g)
Gait Cycle (%)
0 50 100-10
0
10
20
Kn
ee
X (
de
g)
Gait Cycle (%)
0 50 100-50
0
50
100
Kn
ee
Y (
de
g)
Gait Cycle (%)
0 50 100-20
-10
0
10
20
Kn
ee
Z (
de
g)
Gait Cycle (%)
0 50 100-10
-5
0
5
Hip
X (
de
g)
Gait Cycle (%)
0 50 100-20
0
20
40
Hip
Y (
de
g)
Gait Cycle (%)
0 50 100-20
-10
0
10
20
Hip
Z (
de
g)
Gait Cycle (%)
Normal
Right
Left
80
Table 7.3: The degree of similarity DS(G,H) between MS patient and the able-bodied reference
for (A) GRFs, (B) EMG, and (C) ankle, knee, and hip joint angles in 3D. DS is
calculated using different granulation window sizes (wsize) 5, 10, 14, and 20. The
mean and standard deviation of DS are reported. p is the number of fuzzy granules
representing the original data series.
(A)
wsize p FX FY FZ
Left Right
Left Right
Left Right
5 20
0.914 0.889
0.867 0.850
0.949 0.923
10 10
0.914 0.891
0.861 0.852
0.953 0.919
15 6
0.908 0.883
0.861 0.844
0.948 0.935
20 5
0.909 0.888
0.856 0.847
0.945 0.921
mean
0.911 0.888
0.861 0.848
0.949 0.925
std 0.003 0.004 0.004 0.003 0.003 0.007
(B)
wsize p Sol TA GL VL
Left Right
Left Right
Left Right
Left Right
5 20
0.628 0.826
0.726 0.827
0.699 0.765
0.564 0.769
10 10
0.642 0.823
0.737 0.814
0.698 0.774
0.559 0.776
15 6
0.618 0.824
0.732 0.822
0.684 0.776
0.550 0.755
20 5
0.658 0.807
0.732 0.826
0.711 0.798
0.604 0.756
mean
0.637 0.820
0.732 0.822
0.698 0.778
0.569 0.764
std 0.017 0.009 0.005 0.006 0.011 0.014 0.024 0.010
wsize p RF BF Gmed ES
Left Right
Left Right
Left Right
Left Right
5 20
0.627 0.644
0.631 0.587
0.655 0.607
0.614 0.702
10 10
0.642 0.641
0.627 0.590
0.649 0.601
0.629 0.712
15 6
0.618 0.672
0.586 0.543
0.702 0.606
0.639 0.685
20 5
0.642 0.688
0.563 0.540
0.731 0.630
0.710 0.746
mean
0.632 0.661
0.602 0.565
0.684 0.611
0.648 0.711
std 0.012 0.023 0.033 0.027 0.039 0.013 0.043 0.026
81
(C)
wsize p AnkleX AnkleY AnkleZ
Left Right
Left Right
Left Right
5 20
0.745 0.747
0.910 0.861
0.843 0.841
10 10
0.745 0.748
0.913 0.861
0.840 0.835
15 6
0.742 0.745
0.916 0.864
0.838 0.835
20 5
0.728 0.764
0.924 0.881
0.856 0.839
mean
0.740 0.751
0.916 0.867
0.844 0.838
std 0.008 0.009 0.006 0.009 0.008 0.003
wsize p KneeX KneeY KneeZ
Left Right
Left Right
Left Right
5 20
0.572 0.644
0.886 0.922
0.643 0.569
10 10
0.571 0.645
0.885 0.923
0.662 0.585
15 6
0.571 0.646
0.885 0.926
0.662 0.567
20 5
0.570 0.647
0.882 0.919
0.704 0.570
mean
0.571 0.645
0.885 0.923
0.668 0.573
std 0.001 0.001 0.002 0.003 0.026 0.008
wsize p HipX HipY HipZ
Left Right
Left Right
Left Right
5 20
0.567 0.531
0.792 0.735
0.643 0.628
10 10
0.575 0.529
0.794 0.728
0.645 0.624
15 6
0.557 0.555
0.794 0.724
0.660 0.639
20 5
0.563 0.558
0.791 0.713
0.647 0.644
mean
0.565 0.543
0.793 0.725
0.649 0.634
std 0.007 0.015 0.002 0.009 0.008 0.009
Discussion:
For GRFs, Figure 7.8 (A) and high DS presented in Table 7.3 (A) show a high correlation
between patients and the reference patterns. The most obvious discrepancy for muscles activity
is in BF (DS is 0.602 for left and 0.565 for right), Table 7.3 (B),due to the decreased activity
during the first and the final gait phase, as well as, the increased activity during the rest gait
phase, Figure 7.8 (B). Another discrepancy represented by left VL (DS is 0.569) resulted from
the lower activity during the initial contact to mid stance phase. The Hip joint angles in 3D
exhibit the smaller DS, larger differences, Table 7.3 (C) and Figure 7.8 (C).
82
Case 3: congenital dislocation of left hip (CD)
A 22-year old female patient suffers from congenital dislocation of left hip. She has been
treated with an open reduction and cast program in the first and second position. Her right limb
exhibits lower length than left one by approximately 2 cm. Currently, she wears a 1 cm insole on
her right shoe for leveling the pelvis. In this case study, the effect of wearing an insole in the
right shoe has been studied through the analysis of two trials, without insole and with insole,
compared to the reference. This case study serves as a practical example for the assessment of
the improvements in response to an intervention.
Results:
(A)
0 20 40 60 80 100-5
0
5
Gait Cycle (%)
FX (
N/k
g)
0 20 40 60 80 100-2
-1
0
1
Gait Cycle (%)
FY
(N
/kg
)
0 20 40 60 80 1000
5
10
15
Gait Cycle (%)
FZ
(N
/kg
)
Without insoleNormal
Right
Left
0 20 40 60 80 100-5
0
5
Gait Cycle (%)
FX (
N/k
g)
0 20 40 60 80 100-2
-1
0
1
Gait Cycle (%)
FY
(N
/kg
)
0 20 40 60 80 1000
5
10
15
Gait Cycle (%)
FZ
(N
/kg
)
With right insole Normal
Right
Left
83
(B)
0 50 1000
2
4
So
l
Without insole
0 50 1000
2
4
TA
0 50 1000
2
4
GL
0 50 1000
2
4
VL
0 50 1000
2
4
RF
0 50 1000
2
4
BF
0 50 1000
2
4
Gait Cycle (%)
GM
ed
0 50 1000
2
4
Gait Cycle (%)
ES
Normal
Right
Left
0 50 1000
2
4
So
l
With right insole
0 50 1000
2
4
TA
0 50 1000
2
4
GL
0 50 1000
2
4
VL
0 50 1000
2
4
RF
0 50 1000
2
4
BF
0 50 1000
2
4
Gait Cycle (%)
GM
ed
0 50 1000
2
4
Gait Cycle (%)
ES
Normal
Right
Left
84
(C)
Figure 7.9: A comparison between CD patient and the able-bodied reference (normal) for (A)
GRFs, (B) EMG, and (C) ankle, knee, and hip joint angles in 3D, within the full gait cycle.
0 50 100-10
0
10
20
30
An
kle
X (
de
g)
Gait Cycle (%)
0 50 100-20
-10
0
10
20
An
kle
Y (
de
g)
Gait Cycle (%)
0 50 100-15
-10
-5
0
5
An
kle
Z (
de
g)
Gait Cycle (%)
0 50 100-10
0
10
20
Kn
ee
X (
de
g)
Gait Cycle (%)
0 50 100-20
0
20
40
60
80
Kn
ee
Y (
de
g)
Gait Cycle (%)
Without insole
0 50 100-10
-5
0
5
10
Kn
ee
Z (
de
g)
Gait Cycle (%)
0 50 100
-20
-10
0
10
Hip
X (
de
g)
Gait Cycle (%)
0 50 100
-20
-10
0
10
20
Hip
Y (
de
g)
Gait Cycle (%)
0 50 100-15
-10
-5
0
5
10
Hip
Z (
de
g)
Gait Cycle (%)
Normal
Right
Left
0 50 100-10
0
10
20
30
An
kle
X (
de
g)
Gait Cycle (%)
0 50 100-20
-10
0
10
20
An
kle
Y (
de
g)
Gait Cycle (%)
With right insole
0 50 100-15
-10
-5
0
5A
nk
leZ
(d
eg
)
Gait Cycle (%)
0 50 100-10
0
10
20
Kn
ee
X (
de
g)
Gait Cycle (%)
0 50 100-50
0
50
100
Kn
ee
Y (
de
g)
Gait Cycle (%)
0 50 100-10
-5
0
5
10
Kn
ee
Z (
de
g)
Gait Cycle (%)
0 50 100
-20
-10
0
10
Hip
X (
de
g)
Gait Cycle (%)
0 50 100
-20
-10
0
10
20
Hip
Y (
de
g)
Gait Cycle (%)
0 50 100-20
-10
0
10
Hip
Z (
de
g)
Gait Cycle (%)
Normal
Right
Left
85
Table 7.4: The degree of similarity DS(G,H) between CD patient and the able-bodied reference
for (A) GRFs, (B) EMG, and (C) ankle, knee, and hip joint angles in 3D. DS is
calculated using granulation window size (wsize) 10. DS are reported for two trials,
without insoles and with 1 cm insole on right shoe.
(A)
Insole FX FY FZ
Left Right
Left Right
Left Right
without
0.921 0.925
0.891 0.831
0.920 0.857
with 0.920 0.904 0.922 0.833 0.925 0.835
(B)
Insole Sol TA GL VL
Left Right
Left Right
Left Right
Left Right
without
0.540 0.576
0.691 0.700
0.602 0.584
0.810 0.720
with 0.501 0.503 0.642 0.613 0.528 0.437 0.457 0.399
Insole RF BF Gmed ES
Left Right
Left Right
Left Right
Left Right
without
0.528 0.537
0.608 0.406
0.854 0.519
0.440 0.531
with 0.675 0.545 0.474 0.328 0.443 0.385 0.417 0.428
(C)
Insole AnkleX AnkleY AnkleZ
Left Right
Left Right
Left Right
without
0.656 0.668
0.797 0.691
0.747 0.504
with 0.656 0.649 0.816 0.748 0.754 0.522
Insole KneeX KneeY KneeZ
Left Right
Left Right
Left Right
without
0.677 0.639
0.681 0.757
0.497 0.346
with 0.681 0.636 0.697 0.819 0.550 0.424
Insole HipX HipY HipZ
Left Right
Left Right
Left Right
without
0.708 0.572
0.810 0.708
0.603 0.651
with 0.695 0.520 0.788 0.757 0.611 0.597
86
Discussion:
In general, Table 7.4 shows less DS for right leg than left within the two trials. There is
no significant effect of the insole on the GRFs. Although FZ of right leg with insole is slightly
lower than the same force without insole at the beginning of gait cycle as shown in Figure 7.9
(A). This change results from the reduced force of heel strike due to the insole.
EMG patterns for all muscles indicate a similar shape with and without insole, Figure 7.9
(B). However, there are substantial differences in the magnitude. All muscles show a lower
magnitude with insole than without insole for both right and left, and thus, lower DS except RF.
This change may also be explained as a result of magnitude normalization; however, the original
data before normalization preserve the same ratio between EMG magnitudes of the two trials.
The DS of joint angles in Table 7.4 (C) demonstrate a significant effect of the insole on
the right side in the sagittal plane; DS has been increased of 5% after wearing the insole. The
minimum right hip angle has been reduced by nearly 5 degrees in the three planes as shown in
Figure 7.9 (C).
7.5 Conclusion
An investigation for the application of fuzzy granular computing algorithm for the
analyses of human dynamic behavior in 3D space has been made in this chapter. Using the
analysis of three case studies, the results show a high efficiency of the proposed methodology for
the assessment of functional impairments, as well as, the improvements in response to an
intervention during treatment process.
87
Chapter 8: Research Outcomes, Recommendations for Future Works, and
List of Publications
8.1 Research outcomes
A significant contribution has been made by this research in the area of human motion
analysis. An innovative technology has been introduced and proved to be very efficient. As a
result of this intensive work, an enormous number of original papers has been published in
international scientific peer reviewed journals and conference proceeding.
The main outcomes and contributions of this research are summarized as follows:
A novel device called LIMA smart goniometer has been developed for cost-effective and
reliable measurement of joint angles in 3D.
An innovative computational intelligence system has been designed and implemented
based on the algorithms of fuzzy granular computing and multi-sensor data fusion
The application of fuzzy granular computing for the analysis of human dynamic behavior
in 3D space has been investigated.
An objective and quantitative assessment of functional gait impairment has been
provided.
A novel system has been presented that combines both wearable sensors system for data
acquisition and fuzzy granular computing for data analysis, thus, providing a potential
basis for a flexible, efficient and cost-effective systems for clinical gait assessment.
88
8.2 Recommendations for future works
Development of a wireless wearable sensors system (LIMA smart goniometer v2.0)
Development of a 9DOF IMU with fuzzy-based fusion algorithm.
Expansion the experimental database so that it includes a large number of patients with
various gait disorders.
Strong collaboration with medical doctors and therapists will enrich the work with
clinical interpretation of the experimental results.
8.3 List of publications
Peer reviewed original papers has been published on:
A. Journals:
1. Murad Alaqtash, Huiying Yu, R. Brower, A. Abdelgawad, T. Sarkodie-Gyan,
“Application of wearable sensors for human gait analysis using fuzzy computational
algorithm”, Engineering Applications of Artificial Intelligence 24(6), Sep. 2011, 1018-
1025.
2. T. Sarkodie-Gyan, Huiying Yu, Murad Alaqtash, A. Abdelgawad, E. Spier, R. Brower,
“Measurement of functional impairments in human locomotion using pattern analysis”,
Measurement 44(1), Jan. 2011, 181-191.
3. Huiying Yu, Murad Alaqtash, Eric Spier, T. Sarkodie-Gyan, “Analysis of muscle
activity during gait cycle using fuzzy rule-based reasoning”, Measurement 43(9), Nov.
2010, 1106-1114.
B. Conference proceedings:
1. Murad Alaqtash, Thompson Sarkodie-Gyan, and Vladik Kreinovich, “Assessment of
Functional Impairment in Human Locomotion: A Fuzzy-Motivated Approach”, to appear
89
31st Annual North American Fuzzy Information Processing Society conference (NAFIPS
2012), Berkeley, CA, USA, Aug. 6-8, 2012.
2. Murad Alaqtash, Thompson Sarkodie-Gyan, Huiying Yu, Amr Abdelgawad, “An
automated gait assessment tool based on multi-sensors fusion and fuzzy information
granules”, to appear 7th
World Congress for NeuroRehabilitation (WCNR12), Melbourne,
Australia, May 16-19, 2012
3. Murad Alaqtash, Thompson Sarkodie-Gyan, Huiying Yu, Olac Fuentes, Richard
Brower, and Amr Abdelgawad, “Automatic Classification of Pathological Gait Patterns
using Ground Reaction Forces and Machine Learning Algorithms”, Proc. 33rd Ann. Int.
Conf. IEEE Engineering in Medicine and Biology Society (EMBC’11), Boston, MA, Aug.
30-Sep. 3, 2011, 453-457.
4. Murad Alaqtash, Huiying Yu, R. Brower, A. Abdelgawad, E. Spier, T. Sarkodie-Gyan,
“Application of Wearable Miniature Non-invasive Sensory System in Human
Locomotion using Soft Computing Algorithm”, Intelligent Robotics and Applications
(ICIRA’10), Lecture Notes in Computer Science 6424, 2010, 288-299.
5. Thompson Sarkodie-Gyan, Murad Alaqtash, Huiying Yu, Chad MacDonald, Eric Spier,
Richard Brower, “Determination of Functional Impairments using the methods of
computational intelligence”, Proc. 6th World Congress for Neurorehabilitation (WCNR
2010), Vienna, Austria, Mar. 21-25, 2010.
6. T. Sarkodie-Gyan, Huiying Yu, Murad Alaqtash, Eric Spier, and Richard Brower,
“Recognition and Decision-making algorithm in Human Locomotion based on the
Principles of Fuzzy Reasoning”, Proc. IEEE Int. Conf. on Robotics and Biomimetics,
Guilin, China, Dec. 19-23, 2009, 529-534.
90
References
[1] Alaqtash M., Sarkodie-Gyan T., and Kreinovich V., Aug. 2012, Assessment of
Functional Impairment in Human Locomotion: A Fuzzy-Motivated Approach, to appear 31st
Annual North American Fuzzy Information Processing Society conference (NAFIPS 2012),
Berkeley, CA, USA.
[2] Alaqtash M., Sarkodie-Gyan T., Yu H., Fuentes O., Brower R., and Abdelgawad A., Sep.
2011b, Automatic Classification of Pathological Gait Patterns using Ground Reaction Forces and
Machine Learning Algorithms, Proc. 33rd Ann. Int. Conf. IEEE Engineering in Medicine and
Biology Society (EMBC’11), Boston, MA.
[3] Alaqtash M., Yu H., Brower R., Abdelgawad A., and Sarkodie-Gyan T., Sep. 2011,
Application of wearable sensors for human gait analysis using fuzzy computational algorithm,
Engineering Applications of Artificial Intelligence 24(6), 1018-1025.
[4] Allison G.T., Marshal R.N., and Singer K.P., 1993, EMG Signal Amplitude
Normalization Technique in Stretch-shortening Cycle Movements, Electromyography and
Kinesiology 3(4), 236-244.
[5] Arsenault A.B., Winter D.A., Marteniuk R.G., and Hayes K.C., 2001, How many strides
are required for the analysis of electromyographic data in gait?, Scand. J. Rehabil. Med. 18, 133-
135.
[6] Ayyappa E., 1997, Normal Human Locomotion, Part 1: Basic Concepts and
Terminology, Journal of Prosthetics and Orthotics 9, 10-17.
[7] Baker R., 2006, Gait analysis methods in rehabilitation, Journal of NeuroEngineering and
Rehabilitation 3(4).
91
[8] Bargiela A., and Pedrycz W., Granular Computing: An Introduction, Kluwer Academic
Publisher, Dordrecht, 2003.
[9] Begg R., Palaniswami M., and Owen B., May 2005b, Support vector machines for
automated gait classification, IEEE Trans. Biomedical Engineering 52(5), 828-838.
[10] Begg R., and Kamruzzaman J., 2005a. A machine learning approach for automated
recognition of movement patterns using basic, kinetic and kinematic gait data. Journal of
Biomechanics 38, 401–408.
[11] Bonato P., 2005. Advances in wearable technology and applications in physical medicine
and rehabilitation, Journal of NeuroEngineering and Rehabilitation 2, 1-4.
[12] Brand R.A., 1987, Can Biomechanics contribute to clinical orthopaedic assessments,
Iowa Orthopaedic Journal 9, 61-64.
[13] Brand R.A., and Crowninshield R.D., 1981, Comment on criteria for patient evaluation
tools, Journal of Biomechanics 14, 655.
[14] Chau T., 2001a. A review of analytical techniques for gait data. Part 1: fuzzy, statistical
and fractal methods, Gait and Posture 13, 49–66.
[15] Chau T., 2001b. A review of analytical techniques for gait data. Part 2: neural network
and wavelet methods, Gait and Posture 13, 102–120.
[16] Chawla N.V., Bowyer K.W., Hall L.O., and Kegelmeyer W. P., 2002, SMOTE: Synthetic
Minority Over-sampling Technique, Journal of Artificial Intelligence Research 16, 321-357.
[17] Cheng P., and Oelmann B., Feb. 2010, Joint-Angle Measurement Using Accelerometers
and Gyroscopes—A Survey, IEEE Trans. Instrumentation and Measurements 59(2).
[18] Cram J.R., Kasman G.S., and Holtz J., 1998, Introduction to Surface Electromyography,
Aspen Publishers, Maryland.
92
[19] De Luca C.J., 2006, Electromyography Encyclopedia of Medical Devices and
Instrumentation, John Wiley Publisher, 98-109.
[20] Dejnabadi H., Jolles B.M., and Aminian K., Aug. 2005, A new approach to accurate
measurement of uniaxial joint angles based on a combination of accelerometers and gyroscopes,
IEEE Trans. Biomed. Eng. 52(8), 1478–1484.
[21] Dejnabadi H., Jolles B.M., Casanova E., Fua P., and Aminian K., Jul. 2006, Estimation
and visualization of sagittal kinematics of lower limbs orientation using body-fixed sensors,
IEEE Trans. Biomed. Eng. 53(7), 1385–1393.
[22] Dobson F., Morris M., Baker R., and Graham H., 2007, Gait classification in children
with cerebral palsy: A systematic review, Gait & Posture 25, 140-152.
[23] Dong W., Chen I., Lim K. Y., and Goh Y. K., Apr. 2007, Measuring uniaxial joint angles
with a minimal accelerometer configuration, in Proc. I-CREATe, 88–91.
[24] Duda R.O., Hart P.E., and Stork D.G., Pattern Classification. 2nd ed., New York: Wiley,
2001.
[25] Gavrila D.M., and Davis L.S., 1996, 3-D model-based tracking of humans in action: a
multi-view approach, in Proc. IEEE Computer Vision and Pattern Recognition, San Francisco,
73-79.
[26] Ghassemi F., Tafazoli S., Lawrence P. D., and Hashtrudi-Zaad K., Jan. 2008, Design and
calibration of an integration-free accelerometer-based jointangle sensor, IEEE Trans. Instrum.
Meas. 57(1), 150–159.
[27] Ghassemi F., Tafazoli S., Lawrence P.D., and Hashtrudi-Zaad K., May 2002, An
accelerometer-based joint angle sensor for heavy-duty manipulators, in Proc. 19th Annu. Conf.
IEEE Robot. Autom. 2, 1771–1776.
93
[28] Giansanti D., Maccioni G., and Macellari V., 2005, The development and test of a device
for the reconstruction of 3-D position and orientation by means of a kinematic sensor assembly
with rate gyroscopes and accelerometers, IEEE Transactions on Biomedical Engineering 52(7),
1271-1277.
[29] Gok H., Ergin S., and Yavuzer G., 2002, Kinetic and kinematic characteristics of gait in
patients with medial knee arthrosis, Acta Orthopaedica 73(6), 647-652.
[30] Granata K.P., Padua D.A., and Abel M.F., 2004, Repeatability of surface EMG during
gait in children, Gait & Posture 22, 346-350.
[31] Hanson M.A., Powell H.C., Barth A.T., Lach J., and Brandt-Pearce M., 3-5 June 2009,
Neural Network Gait Classification for On-Body Inertial Sensors, 6th Int. Workshop on
Wearable and Implantable Body Sensor Networks, 181-186.
[32] Haykin S., 1999, Neural Networks; a Comprehensive Foundation, Englewood Cliffs, NJ,
Prentice-Hall.
[33] Holt C.A., Hayes N.J., VanDeursen R.W.M., and O’Callaghan P.O., 2000, Three-
dimensional analysis of the tibiofemoral joint using external marker clusters and the JCS
approach - Comparison of normal and osteoarthritic knee function, Computer Methods in
Biomechanics and Biomedical Engineering 3, Gordon and Breach, London.
[34] Ismail A.R., and Asfour S.S., 1999, Discrete wavelet transform: a tool in smoothing
kinematic data, Journal of Biomechanics 32, 317–21.
[35] Jasiewicz J.M., Allum J.H., Middleton J.W., Barriskill A., Condie P., et al., 2006, Gait
event detection using linear accelerometers or angular velocity transducers in able-bodied and
spinal-cord injured individuals, Gait & Posture 24(4), 502-509.
94
[36] Jonesa L., Beynonb M., Holta C., and Royc S., 2006, An application of the Dempster–
Shafer theory of evidence to the classification of knee function and detection of improvement
due to total knee replacement surgery, Journal of Biomechanics 39, 2512–2520.
[37] Karaulovaa I.A., Hallb P.M., and Marshall A.D., 2002, Tracking people in three
dimensions using a hierarchical model of dynamics, Image and Vision Computing 20, 91–700.
[38] Kavanagh J.J., and Menz H.B., 2008, Accelerometry: A technique for quantifying
movement patterns during walking, Gait & Posture 28, 1-15.
[39] Kikuchi Y., Nakamura F., Wakiwaka H., and Yamada H., Sep. 1997, Index phase output
characteristics of magnetic rotary encoder using a magnetoresistive element, IEEE Trans. Magn.
33(5), 3370–3372.
[40] Kohle M., and Merkl D., 2000, Analyzing Human Gait Patterns for Malfunction
Detection, Proc. ACM symposium on Applied computing 1, Como, Italy, 41 – 45.
[41] Kojima T., Kikuchi Y., Seki S., and Wakiwaka H., Mar. 2004, Study on high accuracy
optical encoder with 30 bits, in Proc. 8th Workshop IEEE Adv. Motion Control, 493–498.
[42] Kurata S., Makikawa M., Kobayashi H., Takahashi A., and Tokue R., Oct. 1998, Joint
motion monitoring by accelerometers set at both near sides around the joint, in Proc. 20th Annu.
Conf. IEEE Eng. Med. Biol. Soc. 4, 1936–1939.
[43] Lafuente R., Belda J.M., Sanches-Lacuesta J., Soler C., and Prat J., Nov. 1997, Design
and test of neural networks and statistical classifiers in computer aided movement analysis: a
case study on gait analysis, Clinical Biomechincs 13(3), 216-229.
[44] Lai D.T.H., Levinger P., Begg R.K., Gilleard W.L., and Palaniswami M., Sep. 2009,
Automatic Recognition of Gait Patterns Exhibiting Patellofemoral Pain Syndrome Using a
95
Support Vector Machine Approach, IEEE Transactions on Information Technology in
Biomedicine 13(5), 810-817.
[45] Lau H.Y., Tong K.Y., and Zhu H., 2008, Support vector machine for classification of
walking conditions using miniature kinematic sensors, Med. Biol. Eng. Comput. 46, 563-573.
[46] Luinge H.J., and Veltink P.H., Mar. 2004, Inclination measurement of human movement
using a 3-D accelerometer with autocalibration, IEEE Trans. Neural Syst. Rehabil. Eng. 12(1),
112-121.
[47] Luinge H.J., and Veltink P.H., 2005, Measuring orientation of human body segments
using miniature gyroscopes and accelerometers, Medical and Biological Engineering and
Computing 43(2), 273-282.
[48] Mayagoitaia R.E., Nene A.V., and Veltink P.H., 2002, Accelerometer and rate gyroscope
measurement of kinematics: an inexpensive alternative to optimal motion analysis systems,
Biomechanics 35(4), 537-542.
[49] Mezghani N., Husse S., Boivin K., Turcot K., Aissaoui R., Hagemeister N., and de Guise
J., March 2008, Automatic classification of asymptomatic and osteoarthritis knee gait patterns
using kinematic data features and the nearest neighbor classifier, IEEE Trans. Biomedical
Engineering 55(3), 1230-1232.
[50] Mitchell T. M., 1997, Machine Learning. 1st ed., McGraw-Hill.
[51] O’Malley M.J., Abel M.F., Damiano D.L., and Vaughan C.L., 1997, Fuzzy clustering of
children with cerebral palsy based on temporal-distance gait parameters, IEEE Trans Rehabil
Eng 5, 300-309.
[52] Ohtaki Y., Sagawa K., and Inooka H., Dec. 2001, A method for gait analysis in a daily
living environment by body-mounted instruments, JSME Int. J. Ser. C. 44(4), 1125–1132.
96
[53] Pappas I.P.I., Popovic M.R., Keller T., Dietz V., and Morari M., 2001, A reliable gait
phase detection system, IEEE Transactions on Neural Systems and Rehabilitation Engineering
9(2), 113-125.
[54] Parkka J., Ermes M., Korpipaa P., Mantyjarvi J., Peltola J., and Korhonen I., 2006,
Activity classification using realistic data from wearable sensors, IEEE Transactions on
Information Technology in Biomedicine 10(1), 119-128.
[55] Pasparkis D., and Darras N., 2009, Normal walking: principles, basic concepts,
terminology, 3-dimensional clinical gait analysis, Hellenic Orthopaedic Association 60(4), 183-
194.
[56] Pedrycz W., and Gacek A., 2002, Temporal granulation and its application to signal
analysis, Information Sciences 143, 47-71.
[57] Perry J., 1992, Gait Analysis: Normal and Pathological Function, SLACK Inc.,
Thorofare, NJ, 11-15.
[58] Preece S.J., Goulermas J.Y., Kenney L., Howard D., Meijer K., and Crompton R., 2009,
Activity identification using body-mounted sensors—a review of classification techniques,
Physiol. Meas. 30, 1-33.
[59] Rehbinder H., and Hu X., 2004, Drift-free attitude estimation for accelerated rigid bodies,
Automatica 40, 653–659.
[60] Ricamato A.L., and Hidler J.M., 2005, Quantification of the dynamic properties of EMG
patterns during gait, J. Electromyography and Kinesiology 15, 384–392.
[61] Russell S., and Norvig P., 2003, Artificial Intelligence a modern approach, 2nd ed., New
Jersey: Prentice-Hall.
97
[62] Salarian A., Russmann H., Vingerhoets F.J.G., Burkhard P.R., and Aminian K., Dec.
2007, Ambulatory Monitoring of Physical Activities in Patients With Parkinson's Disease, IEEE
Transactions on Biomedical Engineering 54(12), 2296-2299.
[63] Sarkodie-Gyan T., Yu H., Alaqtash M., Abdelgawad A., Spier E., Brower R., Jan. 2011,
Measurement of functional impairments in human locomotion using pattern analysis,
Measurement 44(1), 181-191.
[64] Selles R.W., Formanoy M.A.G., Bussmann J.B.J., Janssens P.J., and Stam H.J., 2005,
Automated Estimation of Initial and Terminal Contact Timing Using Accelerometers;
Development and Validation in Transtibial Amputees and Controls, IEEE Transactions on
Neural Systems and Rehabilitation Engineering 13(1), 81-88.
[65] Simon S. R., Dec. 2004, Quantification of human motion: gait analysis--benefits and
limitations to its application to clinical problems, Journal of Biomechanics 37(12), 1869-1880.
[66] Stucki G., Ewert T., and Cieza A., 2002,Value and application of the ICF in rehabilitation
medicine, Disability and Rehabilitation 24, 932–938.
[67] Takahashi T., Ishida K., Hirose D., Nagano Y., Okumiya K., Nishinaga M., Doi Y., and
Yamamoto H., 2004,Vertical ground reaction force shape is associated with gait parameters,
timed up and go, and functional reach in elderly females, J. Rehabil. Med. 36, 42-45.
[68] Takeda R., Tadano S., Todoh M., Morikawa M., Nakayasu M., and Yoshinari S., 2009,
Gait analysis using gravitational acceleration measured by wearable sensors, Journal of
Biomechanics 42, 223-233
[69] Tao L., Yoshio I., Kyoko S., Haruhiko K., 2009, Development of a wearable sensor
system for quantitative gait analysis, Measurement 42(7), 978-988.
98
[70] Tong K., and Granat M.H., 1999, A practical gait analysis system using gyroscopes,
Medical Engineering & Physics 21, 87–94.
[71] Wang L., Tan T., Hu W., and Ning H., Sept. 2003, Automatic gait recognition based on
statistical shape analysis, IEEE Trans. Image Processing 12(9), 1120-1131.
[72] Willemsen A.T.M., Frigo C., and Boom H.B.K., Dec. 1991, Lower extremity angle
measurement with accelerometers-Error and sensitivity analysis, IEEE Trans. Biomed. Eng.
38(12), 1186–1193.
[73] Williamson R., and Andrews B.J., May 2001, Detecting absolute human knee angle and
angular velocity using accelerometers and rate gyroscopes, Med. Biol. Eng. Comput. 39(3), 294–
302.
[74] Winter D.A., 1991, Biomechanics and Motor Control of Human Gait: Normal, Elderly
and Pathological, Waterloo Biomechnanics Press, Waterloo, Ontario.
[75] Winter D.A., 2009, Biomechanics and Motor control of Human Movement, fourth ed.,
John Wiley & sons, Hoboken, New Jersey.
[76] World Health Organization, 2001, International classification of functioning, disability
and health, Geneva.
[77] World Health Organization, 2006, Neurological Disorders: Public Health Challenges.
[78] Wu W.L., Su F.C., Cheng Y.M., and Chou Y.L., 2001, Potential of the genetic algorithm
neural network in the assessment of gait patterns in ankle arthrodesis, Ann. Biomed. Eng. 29(1),
83-91.
[79] Yao Y.Y., 2000, Granular computing: basic issues and prossible solutions, Proceedings
of the 5th Joint Conferences on Information Sciences, New Jersey, USA, 186-189.
99
[80] Yu F., and Pedrycz W., 2009, The design of fuzzy information granules: Tradeoffs
between specificity and experimental evidence, Applied Soft Computing 9(1), 264-273.
[81] Yu F.S.H., Chen F., and Dong K.Q., 2005, A granulation-based method for finding
similarity between time series, Proceedings of 2005 IEEE International Conference on Granular
Computing Beijing, China.
[82] Yu H., Alaqtash M., Spier E., Sarkodie-Gyan T., 2010, Analysis of muscle activity
during gait cycle using fuzzy rule-based reasoning, Measurement (43)9, 1106-1114.
[83] Zadeh L.A., 1997, Towards a theory of fuzzy information granulation and its centrality in
human reasoning and fuzzy logic, Fuzzy Sets and Systems, 90(2), 111-127.
[84] Zadeh L.A., 1998, Some reflections on soft computing, granular computing and their
roles in the conception, design and utilization of information/intelligent systems, Soft Computing
2(1), 23-25.
[85] Zadeh L.A., 1994, Soft Computing and Fuzzy Logic, IEEE Softw. 11(6), 48-56.
[86] Zhu R., and Zhou Z., Jun. 2004, A real-time articulated human motion tracking using tri-
axis inertial/magnetic sensors package, IEEE Trans. Neural Syst. Rehabil. Eng. 12(2), 295–302.
100
Vita
Mr. Murad Alaqtash earned his Bachelor’s and Master’s degrees in computer engineering
from Jordan University of Science and Technology, Jordan, in 2003 and 2007, respectively. He
worked as a computer engineer at the ministry of education, Jordan, for three years. After he
received his Master’s degrees, he worked as full time lecturer at the department of electrical and
engineering, Tafila Technical University, Jordan, before traveling to USA in August 2008 and
attending The University of Texas at El Paso toward his doctoral degree in Computer
Engineering.
Mr. Alaqtash worked as an assistant instructor for department of Electrical and Computer
Engineering and as a research assistant for Research Institute for Manufacturing and Engineering
Systems. While pursuing his doctoral degree, he served as a research and development engineer
at the laboratory for Human Motion Analysis and Neurorehabilitation and the laboratory for
Industrial Metrology and Automation. His main research interests include soft computing
algorithms, intelligent systems, wearable biosensors, and their applications in human motion
analysis. He has presented his research at international conferences and published original
papers related to his field in international scientific journals and conference proceedings.
Mr. Alaqtash has participated in several institutional and professional societies as the
president of Neuroscience Research Association, vice president of the UTEP Student Chapter of
Engineering World Health, and a member of the Society for Robotics and BioCybernetics.
Mr. Alaqtash has been recognized as an outstanding graduate student in the department of
Electrical and Computer Engineering and awarded by Dodson dissertation fellowship from the
Graduate School.
101
After he receives his PhD degree in computer engineering from University of Texas at El
Paso, Mr. Alaqtash is planning to serve his home country, Jordan, where he will be a faculty
member at Tafila Technical University.
Permanent address: 611 Rubin Dr. Apt. C28
El Paso, TX 79912
This dissertation was typed by Murad Mohammad Suleiman Alaqtash.