Content uploaded by Vijay Bhaskar Semwal
Author content
All content in this area was uploaded by Vijay Bhaskar Semwal on Dec 22, 2019
Content may be subject to copyright.
Clinical Human Gait Classification: Extreme Learning Machine Approach
1
Prithvi Patil,
2
K.Shusheel Kumar,
3
Neha Gaud and
4
Vijay Bhaskar Semwal*
1
National Institute of Technology, Srinagar
2
B.I.E.T. Jhansi
3
Vikram University Ujjain
4
Maulana Azad National Institute of Technology, Bhopal
prithvipatil1357@gmail.com,sus.iiita.932@gmail.com, gaud28neha@gmail.com,vsemwal@gmail.com*
Abstract:
This study reports a novel approach for biometric
gait pattern classification using Extreme Learning Machine
(ELM) algorithm. Clinical gait analysis can be used for early
detection of gait abnormality in brain or neurological disorder
subjects. In many cases gait abnormality cannot be detected
through visual observation alone, but becomes apparent only in
a quantitative analysis of subject’s gait. This can also help us
understand the neuro-muscular mechanics associated with brain
disorders. Human gait is also of profound interest to the
research community in the field of biometric identification and
bipedal robot locomotion due to its uniqueness and efficiency.
This paper explores multi-class gait classification using four
machine learning methods (KNN, SVM, ELM, MLP) and
evaluates their performance for multi class gait classification.
The proposed method achieves very good results. TheELM is
used first time in to analyses the neuromuscular of patients
suffering from multiple sclerosis and stroke
Keywords: Extreme Learning Machine (ELM), Biometric
identification, Gait classification, Bipedal walk, Multiple
sclerosis, Stroke gait, Clinical gait analysis, Cerebral palsy,
Crouch gait, Anthropomorphic robot.
I. INTRODUCTION
Human gait is a complex locomotion for forward
propulsion of center of gravity of the human body. It is achieved
through a process of continuous learning as a child interacts
with his environment [1]. It is highly nonlinear and complex
due to varying configuration during different sub phases of
human gait [2]. The human walk is designed as 8 discrete sub
phases [3] with continuous behavior and researchers have
started considering it as hybrid system [4]. Fig.1 shows the 8
different phases of bipedal walk. It is inherently unstable and
nonlinear due to high degrees of non-linearity, high
dimensionality, under actuation (in swing phase) and over
actuation (in stance phase) [5]. Human gait is also unobtrusive
compared to other biometric identifiers such as fingerprint [6].
Many Researchers have used human gait for purposes such as
biometric identification, understanding the problem of elderly
people, to understand the biomechanics of human walk,
artificial limb development and bipedal walk [7] [8].
As on today, due to its inherent complexities, bipedal
robots are not efficient to work outside laboratory environment
i.e. unstructured environments designed for humans and are
controlled as fully-actuated system, which is not energy
efficient [9]. Among various types of humanoid locomotion,
bipedal walking is the most natural, energy efficient and
interesting, since bipedal robots have more potential to move in
rugged terrains or complex environments, where wheeled or
tracked robots cannot operate. On the other hand, bipedal robots
are less stable and prone to falling down. To meet this
challenge, over the years, many solutions have been proposed
[10].
Gait can be studied as non-linear time series of
kinematic trajectories [11],[12]. The analysis involves the
investigating sensory motor interaction and understanding the
biomechanics of locomotors system [13]. Gait classification
allocates walking patterns into groups that can be identified and
differentiated from one another based on a set of defined
variables or features. We believe the problem being addressed
so far using conventional mechanics based model and
automated control theory can effectively be addressed using
data driven computational theory. Throughout this work we
tried to validate this hypothesis using machine learning
techniques including ELM for patients suffering from multiple
sclerosis and stroke. The ELM is used first time in to analyses
the neuromuscular of patients suffering from multiple sclerosis
and stroke
The paper is organized as follows. Section II is about
methodology. Section III introduces ELM and briefly describes
other machine learning approaches used in this study. Section
IV discusses performance evaluation and results achieved by
different machine learning algorithms. Section V is about
Conclusion and Future work.
II. METHODOLOGY
A.
The Proposed System:
Proposed system consists of eight phases which is shown in
the Fig.1. They are gait data collection, gait data detection,
trajectories smoothing, feature extraction, feature selection and
classification using ML algorithms [14]. Figure 2 depicts a
systematic approach for proposed gait recognition process.
1st International Conference on Advances in Science, Engineering and Robotics Technology 2019 (ICASERT 2019)
978-1-7281-3445-1/19/$31.00 c
2019 IEEE
Fig.1. Breakdown of Human Gait into different Discrete Sub Phases
Fig.2. Proposed Architecture of the System
The Gait is divided into 2 phases, namely stance phase and
swing phase. The swing phase is critical since only one leg
needs to carry the entire body load. We consider it as discrete
phase and the configuration during this phase will be over
actuation. The swing phase can be considered as continuous
phase as well. The configuration is under actuation during this
phase.The Percentage-wise distribution of gait time series as
following:
Stance phase:
1. Initial Contact– IC[0-2%]
2. Loading Response– LR[2-10%]
3. Mid Stance– MS[10-30%]
4. Terminal Stance– TS[30-50%]
Swing phase:
5. Pre Swing– PS[50-60%]
6. Initial Swing– IS[60-73%]
7. Mid Swing - MS[73-87%]
8. Terminal Swing – TS[87-100%]
B. Data Collection:
For class Multiple sclerosis and Stroke gait, data was collected
and recorded using accelerometer for capturing acceleration in
all the three planes for the 6 joints (left hip, right hip, left knee,
right knee, left ankle, right ankle).This was later converted into
joint angles using inverse kinematic solution [17].
C. Data Smoothing:
To remove unwanted noise from the data collected through
accelerometers, we have cubic spline interpolation.
D. Dimension Reduction:
After data smoothing, we employed PCA (Principle
Component Analysis) and IFA(Incremental Feature Analysis)
to compute covariance between different features
incrementally.
E. Feature selection:
After analyzing covariance between different set of features,
we then proceed to select set of 6 features out of 10, which
have highest covariance. This not only reduces the chance of
over fitting, but also leads to improved generalization and
classification accuracy as discussed in [15],[16].Table-1 below
shows all the 10 kinematic features. F1-F6 represents the
selected features.
Table 1:
List of 10
kinematic features
III. ALGORITHMS:
A. Extreme Learning Machines(ELM)
ELM was proposed by Huang [18], [19]. It is different from
MLP and SVM. MLP used to consider many hidden layer
neuron and considered whole networks as a black box. ELM
theories show that hidden neurons do not need tuning because
its hidden nodes parameters (c
i
,a
i
) are randomly assigned [20].
For N arbitrary distinct samples
(
,
)∈
×
, the
single ELM classifier with N hidden nodes becomes a linear
system as,
∑β
G(x
;c
;a
)=t
,k=1,…….N
. (1)
where c
i
€ Rn and a
i
€ R are the learning parameters of hidden
nodes and randomly assigned weight β
i
connecting the ith
hidden node to the output node, xk are the training examples,
t
k
is the target output for k = 1, ..., N, and G (x
k
; c
i
; a
i
) is the
Feature Category Feature Name
F1 Left Ankle
F2 Right Ankle
F3 Left Knee
F4 Right Knee
F5 Left Hip
F6 Right Hip
F7 CoP*
F8 CoG*
F9 GRF*
F10 Velocity
output of the i
th
hidden node with respect to the input xk. The
output weights can be described in matrix form as
β=β
….β
(2)
Equation (1) can be rewritten as:
Hβ=T (3)
Where H (c
1
…….c
N
, a1 ……, x
1
, ….., x
N
) =
G(x
;c
,
a
)…G(x
;c
,
a
)
⋮⋱⋮
G(x
;c
,
a
)…G(x
;c
,
a
)
(4)
T=t
…t
(5)
The weights β can be obtained using equation 6 by
taking the least-square solution:
β
=H
T (6)
Here, H† is the Moore-Penrose generalized inverse [21] and
H is the output matrix of the hidden layer. The ELM used to
implement the multi-class classification using a network
architecture which has output nodes equal to the number of
pattern in the classes. The network output can be written as O
= (O
1
; O
2
; …O
n
)
T
. For each training example say x, the target
output tg is coded into n bits: (tg
1
; --- tg
n
)
T
.For a pattern of
class k, only the target output t
k
is “1” and the rest is “-1”
B. MLP, SVM and KNN Classifiers:
Multi-layer perceptron (MLP) is the simplest neural network
architecture which is used most frequently. It has configuration
flexibility, good representational capabilities and
programmable algorithms. The Complexity of neural network
depends of no. of hidden layer and no. of neurons. The best
model of neural network is derived by trial and error. The
major challenge of neural network is to decide the no. of
hidden layer and no. of neurons, as there is no as such
mechanism of optimal selection of neurons and hidden layers.
Most MLP are based on gradient descent techniques that
require iterative tuning which makes MLP extremely slow and
the algorithm can also converge to local minima and provide
sub optimal results.
Fig.3. SLFNarchitecture of ELM
The objective of next classifiers SVM is to find the
optimal decision boundary with maximum margin between
classes in higher dimension feature space. It used to map the
lower dimension data into higher dimension feature space
using kernel function. The kernel function used to take linear
time for this transformation. It converts the nonlinearly
separable data into linearly separable data in high dimension
feature space. Using SVM, an optimal separating hyper plane
in the higher dimensional feature space can be computed by
using kernel functions in the input space. In this study, we
have used a Support Vector Machine classifier based on
LibSVM. It uses a RBF(Radial Basis Function) kernel.
KNN used to match the similarity measure to classify the
new class. It is simple algorithm. It is a simple non-parametric
technique for classification. When the new testing data is fed,
KNN computes the distance between the query data and
training samples. Based on the defined threshold, number-k, k
samples with least distances are selected and the class with
more samples in the boundary is the result.
IV. RESULTS AND DISCUSSIONS
Our experiments were conducted on a Windows 10 machine
with a 2.50-GHz 7200U i5 CPU and 8.0-GB RAM. To find the
best parameters for each of the classifiers, a grid search was
run exhaustively with 10-fold cross validation to reduce
variation associated with randomness of data shuffling and
arrive at mean scores. KNN, SVM and MLP were
implemented using scikit-learn python module. The best set of
parameters was found to be as under. Table 2 represents the
best set of parameters for MLP.
Table2: Set of Parameter for
MLP
Hidden Layer Neurons 50
Activation function ‘relu’
solver ‘Lbfgs’
alpha 0.0001
momentum 0.9
iterations 1000
‘relu’ is an activation function such that,f(x)=max(0,x).‘lbfgs’
is an optimizer in the family of quasi-Newton methods.‘alpha’
is the L2 penalty parameter.‘multiquadric’ is an Radial basis
activation function suchthat,f(x)=√1+( x)2. table 3,4 and 5
represents the best set of parameters for KNN, SVM, ELM.
Table 3:Set of Parameter for
KNN
Neighbors 4
weights ‘uniform’
metric ‘manhattan’
Table 4:Set of Parameter for
SVM
Parameter C 32768
kernel ‘rbf’
gamma 0.001953125
Table 5: Set of Parameter for
ELM
Hidden neurons 370
Activation function ‘multiquadric’
Rbf-width 0.4
To evaluate results we computed confusion matrix,
precision, recall, f1 score for each of the four classifiers and
calculated overall classification accuracy. Overall accuracy is
simply number of correct predictions divided by total number
of predictions. F1 score is the harmonic mean of precision and
recall. Additionally, we also recorded time taken for each
classifier to train and predict labels on testing set. All
performance measures were generated using scikit-learn
python module. In order to eliminate high variation from
results we have employed a 10-fold cross validation.
The data set consists of 5 classes namely,(Normal,
Crouch1,Crouch2) obtained from OpenSim [22], Multiple
sclerosis and Stroke. Data for classes Multiple sclerosis and
Stroke were collected clinically capturing the same kinematic
features as in the first three classes. Subjects didn’t suffer from
any other medical condition besides the said diseases. There
are, a total of 945 samples across all 5 classes with their
individual numbers given under ‘support’ column in the
classification report tables. Crouch gait is a common
movement abnormality among children with cerebral palsy,
decreases walking efficiency due to the increased knee and hip
flexion during the stance phase of gait [23][24].Crouch gait is
of four types, out of which we have included type 1 and type 2
for classification. Table 6 and 7 represents the confusion
matrix and classification report for ELM.
Table 6:
ELM Confusion Matrix
Normal Crouch1 Crouch2 MS Stroke
Normal 202 0 0 0 0
Crouch1 0
54 19 0 0
Crouch2 0 25 45 0 0
MS 0 0 1 299 0
Stroke 0 9 7 0 284
Table 7: Classification Report using ELM
Precision Recall F1-
score
Support
Normal 1.00 1.00 1.00 202
Crouch1 0.61 0.74 0.67 73
Crouch2 0.62 0.64 0.63 70
MS 1.00 1.00 1.00 300
Stroke 1.00 0.95 0.97 300
The Overall classification accuracy using ELM is 93.54%.
Table 8 and 9 represents the confusion matrix and
classification report for MLP.
Table 8:
MLP Confusion matrix
Table 9: Classification Report using MLP
Precision Recall F1-
score
Support
Normal 0.99 0.98 0.98 202
Crouch1 0.37 0.36 0.36 73
Crouch2 0.34 0.37 0.36 70
MS 1.00 1.00 1.00 300
Stroke 1.00 1.00 1.00 300
The Overall classification accuracy is 89.84% achieved using
MLP. Table 10 and 11 represents the confusion matrix and
classification report for KNN..
Table 10:
KNN Confusion matrix
Normal Crouch1 Crouch2 MS Stroke
Normal 191 5 6 0 0
Crouch1 20
20 33 0 0
Crouch2 19 50 1 0 0
MS 0 0 0 300 0
Stroke 0 0 0 0 300
Table 11: Classification Report using KNN
Precision Recall F1-
score
Support
Normal 0.83 0.95 0.88 202
Crouch1 0.27 0.27 0.27 73
Crouch2 0.03 0.01 0.02 70
MS 1.00 1.00 1.00 300
Stroke 1.00 1.00 1.00 300
The Overall classification accuracy=85.92% using KNN.
Table 12 and 13 represents the confusion matrix and
classification report for SVM.
Normal Crouch1 Crouch2 MS Stroke
Normal 197 2 3 0 0
Crouch1 0
26 47 0 0
Crouch2 2 42 26 0 0
MS 0 0 0 300 0
Stroke 0 0 0 0 300
Table 12:
SVM
Confusion matrix:
Table13: Classification Report using SVM
Precision Recall F1-
score
Support
Normal 1.00 1.00 1.00 202
Crouch1 0.78 0.73 0.75 73
Crouch2 0.73 0.79 0.76 70
MS 1.00 1.00 1.00 300
Stroke 1.00 1.00 1.00 300
We have achieved the overall classification accuracy 96.29%
using SVM.
Comparative analysis
We have used three figures as shown below to bring out a
comparative analysis of the four classifiers. From figure-4 we
can see that although numerically the poorest performing
classifier-KNN, achieves a good overall accuracy of 85.92%,
but it’s classification is very poor in the case of Crouch 1 and
Crouch2.MLP too, fails to have a decent f1 score in the same
two classes, while SVM and ELM have evidently performed
well across all 5 classes. The overall accuracy comparison is
provided by figure-5.Interestingly, although KNN is the fastest
of all four algorithms, it’s also the least accurate, while ELM,
despite having a far more complex Neural network of 370
neurons against MLP’s 50, predicts labels in less than 1/10
th
the time.
Fig.4. Comparison across classes
Fig.5. Different Classifier Classification Accuracy
Fig.6. Different Classifier Classification time in
seconds
VI. CONCLUSION
This paper presented the performance of ELM based classifiers
with other machine learning classifiers named SVM, MLP and
KNN for multi class gait classification. The paper has used
kinematic features. The results show that ELM performs very
good classification accuracy compare to other classifiers. The
ELM has achieved the 93.54% overall classification accuracy.
As the performance of machine learning algorithms dependent
on the selected feature and parameter used, to improve the
performance the paper has implemented increment feature
analysis technique for kinematics features selection and
various parameters is adjusted. In the future one could
experiment with dynamic and kinematic features as this study
was limited to kinematic gait characteristics alone.
References:
[1] Semwal, Vijay Bhaskar, et al. "Biped model based on human
Gait pattern parameters for sagittal plane
movement." IEEE,International Conference onControl,
Automation, Robotics and Embedded Systems (CARE), 2013.
Normal Crouch1 Crouch2 MS Stroke
Normal 202 0 0 0 0
Crouch1 0
53 20 0 0
Crouch2 0 15 55 0 0
MS 0 0 0 300 0
Stroke 0 0 0 0 300
[2] Kale, A.;Sundaresan, et.al. "Identification of humans using
gait." Image Processing, IEEE Transactions on 13, no. 9 (2004):
1163-1173.
[3] Kiichi_Katoh,, Makoto Katoh.: Using FS in Tuning a Biped
Locomotion Surrogate Robot for Walking Training. In: 2011 4th
International Conference on Biomedical Engineering and
Informatics (BMEI).
[4] Nandi, Gora Chand, et al. "Modeling bipedal locomotion
trajectories using hybrid automata." IEEE TENCON,2016.
[5] Mitsuru Yoneyama, Yosuke Kurihara, KajiroWatanabe,:
Accelerometry-Based Gait Analysis and Its Application to
Parkinson’s Disease Assessment—Part 2: Ieee Transactions On
Neural Systems And Rehabilitation Engineering, Vol. 21, No. 6,
November 2013
[6] Venkataraman, S. T. "A simple legged locomotion gait
model." Robotics and Autonomous Systems 22.1 (1997): 75-85.
[7] Semwal, Vijay Bhaskarand Gora Chand Nandi. "Biometric gait
identification based on a multilayer perceptron." Robotics and
Autonomous Systems 65 (2015): 65-75.
[8] Watelain, Eric, Franck Barbier, Paul Allard, André Thevenon,
and Jean-Claude Angué. "Gait pattern classification of healthy
elderly men based on biomechanical data." Archives of physical
medicine and rehabilitation 81, no. 5 (2000): 579-586.
[9] Huang, Guang-Bin, Lei Chen, and Chee Kheong Siew.
"Universal approximation using incremental constructive
feedforward networks with random hidden nodes." IEEE Trans.
Neural Networks 17.4 (2006): 879-892.
[10] Khusainov, R., A. Klimchik, and E. Magid. "Comparison of
kinematic and dynamic leg trajectory optimization techniques for
biped robot locomotion." Journal of Physics: Conference Series.
Vol. 803. No. 1. IOP Publishing, 2017.
[11] Semwal, Vijay Bhaskar, et al. "Design of Vector Field for
Different Subphases of Gait and Regeneration of Gait
Pattern." IEEE Transactions on Automation Science and
Engineering (2016).
[12] Khusainov, Ramil, et al. "Toward a human-like locomotion:
modelling dynamically stable locomotion of an anthropomorphic
robot in simulink environment." Informatics in Control,
Automation and Robotics (ICINCO), 2015 12th International
Conference on. Vol. 2. IEEE, 2015.
[13] Semwal, Vijay Bhaskar, and Gora Chand Nandi. "Generation of
joint trajectories using hybrid automate-based model: a rocking
block-based approach." IEEE Sensors Journal 16.14 (2016):
5805-5816.
[14] Huang, Guang-Bin, et al. "Extreme learning machine for
regression and multiclass classification." IEEE Transactions on
Systems, Man, and Cybernetics, Part B (Cybernetics) 42.2
(2012): 513-529.
[15] Semwal, Vijay Bhaskar, et al. "An optimized feature selection
technique based on incremental feature analysis for bio-metric
gait data classification." Multimedia Tools and
Applications (2016): 1-19.
[16] Vijay Bhaskar Semwal and G.C. Nandi, Robust and more
accurate feature and classification using deep neural network,
Neural Computing and Application, Springer.
[17] Semwal, Vijay Bhaskar, and Gora Chand Nandi. "Toward
developing a computational model for bipedal push recovery–a
brief." IEEE Sensors Journal 15.4 (2015): 2021-2022.
[18] Huang, Guang-Bin, Qin-Yu Zhu, and Chee-Kheong Siew.
"Extreme learning machine: a new learning scheme of
feedforward neural networks." Neural Networks, 2004.
Proceedings. 2004 IEEE International Joint Conference on. Vol.
2. IEEE, 2004.
[19] Huang, Guang-Bin, Qin-Yu Zhu, and Chee-Kheong Siew.
"Extreme learning machine: theory and
applications." Neurocomputing 70.1 (2006): 489-501
[20] Huang, Guang-Bin. "An insight into extreme learning machines:
random neurons, random features and kernels." Cognitive
Computation 6.3 (2014): 376-390.
[21] Rao, Calyampudi Radhakrishna, and Sujit Kumar Mitra.
"Generalized inverse of matrices and its applications." (1971):
601-620.
[22] Delp, Scott L., et al. "OpenSim: open-source software to create
and analyze dynamic simulations of movement." IEEE
transactions on biomedical engineering 54.11 (2007): 1940-1950.
[23] Wren, Tishya AL, Susan Rethlefsen, and Robert M. Kay.
"Prevalence of specific gait abnormalities in children with
cerebral palsy: influence of cerebral palsy subtype, age, and
previous surgery." Journal of Pediatric Orthopaedics 25.1 (2005):
79-83.
[24] Semwal, Vijay Bhaskar, Neha Gaud, and G. C. Nandi. "Human
gait state prediction using cellular automata and classification
using ELM." Machine Intelligence and Signal Analysis. Springer,
Singapore, 2019. 135-145.