the systematic review on gait analysis: trends and
TRANSCRIPT
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1636
The Systematic Review On Gait Analysis:
Trends And Developments
Nithyakani P1, M Ferni Ukrit2
Assistant Professor, School of Computing, SRMIST. [email protected] [email protected]
ABSTRACT -The research on gait analysis has a drastic growth in recent years for various
applications like authentication, video surveillance, health monitoring system and animation. Gait
recognition is an emerging intelligent system because of its recognition at a distance and it works
well with low resolution videos. This paper provide the various gait features extracted in model
based approach and model free approach. This paper describes the survey of gait recognition
techniques which include the four major steps such as data acquisition, preprocessing, feature
extraction and classification. Descriptions about wearable sensor, Vision base and pathological
database is summarized. In recent years, performance of human recognition has been impressively
enhanced by deep architectures. This paper offers an up to- date survey of deep architectures on gait
recognition, mainly focusing on the performance of convolutional neural network combined with
other architectures. Furthermore, general challenges faced in gait recognition and directions for
future research are discussed.
1. INTRODUCTION
Human analysis and identification has been an attractive technology in many fields.In the digital
imaging, a human can be analyzed by his/her unique feature of face, iris, sweat, hair, periocular region
of eyes,Gait, smell, finger and palm[18]. Gait defines the style of walking. Human Gait analysis is a
study of human movement to analyze the style of walking and its related features which are applied in
the field of biometrics, surveillance, diagnosis of gait disease, therapy and rehabilitation, etc. Gait
recognition can be performed through involvement and non-involvement of a person to the
recognition system. In human involvement approach, person is directly contacted to the recognition
system with the help of different sensors, accelerometer or monitoring devices.Non-involvement
approach uses camera at a distance to identify a person identity with or without their knowledge based
on its application. Gait analysis is prevalent than other features since thefeatures can be extraction
from low resolution and without the cooperation of any individual.
“On the Gait of Animals” is the first book to report the gait analysis written by Aristotle in 350BC[1].
In 1680, Giovanni Alfonso Borelli made the extensive study on the animal locomotion mechanism. In
1890, Christian Wilhelm Braune and Otto Fischer performed the investigation of gait activity over
human and published a series of papers. They have extended their research on the center of gravity,
moments of inertia of the human body and its segments. In 1880s Eadweard Muybridge and Étienne-
Jules Marey developed the animated photography to capture the several movement phases in a
photographic surface. Muybridge used multiple cameras to record the locomotion and presented
through his public lectures and demonstrations. In 1970s, video camera systems were available to
study the several human pathological conditions. Human individuals were recognized with their gait
style in 1970s. Later gait analysis was impressed by the forensic and animation experts [3-5].
Advancements in technologies of artificial intelligence and computing,Gait recognition has become
the most popular data driven and artificial surveillance tool which can recognize the human at a
distance. This survey aims to highlight the technologies developed for gait analysis. Section II
provides the application of gait recognition in various fields. Section III describes about the human
gait features considered for model based and appearance based recognition. Section IV focuses on the
role of machine learning algorithms in recognizing human. Section V presents the description and
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1637
comparison of recently implemented deep architecture which gives the high accuracy. Section VI
discusses about future directions in the field of human gait analysis andSection VII concludes with the
challenges and the future work.
2. APPLICATIONS OF GAIT RECOGNITION
2.1. Application-Based Systems
Authentication is necessary to avoid violation in secured protected areas, home, public security,etc,.
Authentication processes[3] are carried out in three methodologies such as knowledge based
(credentials, PIN, MPIN), object or token based (E.g bank card, identity card etc.,) and biometric
based. Knowledge based authentication process faces memorability issues specifically robust
passwords. Security breach arises when the password is reused in many services. Object based
method usually combined with knowledge based method for accessing the services provided with
physical Keys. Biometric authentication is used to identify a person by his/her unique features like
fingerprints, face, and periocular region of eye, retina, palm, sweet, body odor and gait. These features
are unique in all subjects and also remain stable for a specific period. Behavioural features are
difficult to imitate. Gait recognition method includes computer vision techniques, signal analysis,
motion estimation, processing the video data, and ubiquitous applications.
2.2. Pattern Recognition
Biometrics is divided into physical and behavioral features. Physical features are the pattern of height,
structure of face, retina, eye etc. Physical features has universality, unique features for a long term
period. Behavioural features includes the way of walking, speaking, singing, movement etc,.Though
behavioral features varies on emotional factor, difficult to forge or imitate the individuality. A set of
biometric images are recorded for verifying the similarity of captured image. Due to the recognition
accuracy in various capturing modes, gait recognition has the highest popularity among the other
biometric approach.
2.3. Life And Medical Sciences
Gait recognition helps to identify the quality of walking of an individual with several common gait
disease such as parkinson’s disease , multiple sclerosis, cardiopathies, post-stroke Hemiplegic gait,
spastic diplegic, neuropathic, myopathic, choreiform, ataxic (cerebellar) and sensory [9-10]. Study of
the gait sequences helps for early diagnoses of their musculoskeletal and neurological complications
to provide the best treatment[62]. Kinematic and kinetic data are extracted from gait analysis helps
the clinicians to interpret, assess, and evaluate the individual gait disorder. Data collection is
performed by videotaping the individual before the examination, examining physically to analyze the
muscle strength and tone, motion of joint, abnormalities in bone, stubborn muscular contractions,
knee and ankle width, distance between right and left sacroiliac spine, etc.
2.4. Animation
Animation of human gait is focused by the fields of biomechanics, signal processing, kinematics,
dynamics and robotics[6-7,11]. Interest in human gait animation is growing rapidly in the Computer
Animation Community despite of the complex techniques. The first computer animation approach for
human was developed in 1970s which depicts the early sequential pictorial representation of series.
Later Humanoid animation(H-Anim) standard was approved in 1990s and greatly influenced by the
research experts of gaming,movies, virtual avatars, graphics, ergonomics and simulations industry.
The challenge in human animation is to simulate the natural motion style into the virtual
characteristics. Animation is performed by the motion sequence extraction of skeleton representation
of human connected hinges points like shoulder, neck , elbow, waist, knee and ankle . Higher quality
is obtained by modelling the skeleton with surfaces. Methods of human gait animation are knowledge
based kinematic animation, dynamic animation[61] and interactive synthetic walking motion data.
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1638
3. Gait Recognition Features
Gait is a synchronized movement of bones, muscles and nerves of hip, knee and ankle joints rotations.
The cyclic nature of the walking style provides the periodic leg movement. Gait of a person is
generally represented as a walk cycle describes in the Figure 1which has two phases: stance phase and
swing phase ; gait cycle has eight stages[2] as (i) initial contact of the foot or toe. (ii) Loading
response as placing the entire foot over the floor with the limb segments reaction. (iii) Mid-stance as
the intermediate support. (iv) Terminal stance is denoted as heel off. V) Pre-swing is the initialization
of knee movement. VI) Initial Swing .Parameters extracted from gait sequence are step length, stride
length, cadence, speed, dynamic base, progression line, foot angle, hip angle and squat performance.
The measurements extracted were analyzed, interpreted to recognize the individual through the
techniques like temporal or spatial, kinematics, Marker less gait capture, pressure analyzer, kinetics
and dynamic electromyography.
Fig. 1.Asingle Gait cycle is also termed as stride. Stance phase begins with the heel strike and toe off
the same foot. Average seconds of stance phase is 0.64s. Swing phasebegins with another foot heel
strike and toe off. Average seconds of swing phase is 0.40s.A stride can also be represented with 8
stages: Initial contact, loading response, mid stance, terminal stance, pre-swing, initial swing, mid
swing, terminal swing.
Human gait recognition system depends on the video has two major categories as model free gait
recognition and model based gait recognition. Model based approaches extract the structural pattern
of model from human body using various methodologies such as pendulum, leg inclination, stride,
trajectories, joint angles, 2D and 3D temporal models [19-22] with their length, width and position
dimensions. It helps to track the individual with set of mathematical parameters and its quantifiable
relationships. Model based approach used in handling self-occlusions. Issues faced in Model based
approach were viewpoint invariant, physiological changes such as age, weight, height , walking
speed,etc., and external factors like clothes , foot wear, load , etc., and environmental factors.In
model based approach kinematics track joints of an individual are head, shoulder, spine, foot, knee,
ankle, thumb, neck, hand,etc shown in the Figure 2a.It also defines the dynamics or kinematics of the
body movements in the terms of gait cycle time, stride length, Fourier magnitude, footstep , ground
reaction force, pressure mat features, etc, Model based gait recognition utilizes spatio temporal
volume, linear regression analysis, trignometric-polynomial interpolant functions, biped locomotion
feature , Principal component Analysis, canonical analysis, 1NN, KNN, Hidden markov Models.
Accuracy of the model based approach depends on the silhouette quality.
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1639
Fig. 2.Gait measurements andtypes of wearable sensors
a) Measurements of human body used for feature extractionb) wearable sensors used to extract gait
features
The other categories of gait recognition approaches were image and video based approach, wearable
sensors, floor sensor and radar sensor[26]. Image and video based approach which involves remote
capturing of human movement. Image and video processing capture the gait sequence of an individual
with the help of several optic sensors like digital camera ,RGB, Low RGB-D Microsoft Kinect, laser
range scanners, infrared sensors and time to flight cameras. Table 1 presents the description about
various gait datasets collected using camera and wearable sensors. Wearable sensors which are
attached to the human body measures movement time, calculate joint or inclination angle, walking
speed, step or stride length, segment position , acceleration ,etc. Wearble sensors used for human
recognition are accelerometer, gyroscope, force sensors, strain gauges, inclinometers, goniometer,
Magnetoresistive sensor, electromagnetic tracking system, fabric sensor, and Electromyography
sensor (Figure 2b). Sensors can be combined with the respect to application of gait analysis. Floor
sensor based approach measures pressure , step length ,etc. with the help of pressure under floor mat
and floor accelerometer .In gait recognition, various factors can modify the gait pattern permanently
or transiently. External factors are terrain, footwear, clothing and cargo; Internal factors are sex,
weight, height, age, physique, etc.
3.1. Model Based Features
Model based approach model[19]generate the structure for human body and the arbitrary 2D views
from the 3D model using the techniques like pendulum, cadence , stride, 2d Stick , kinematic features,
angular trajectories etc. 3D model uses more than two calibrated cameras to capture the gait
sequences and to avoid the problems of variations in surfaces, self-occlusion and fixed viewpoint.
Structural method uses geometrical patterns of time variations and motion variations of gait sequence.
This method has overcome the problem of variation in light, background, view and segmentation. The
features extracted in model based approaches were classified using K- Nearest Neighbor, Back
Propagation Neural network, Multilayer Perceptron, Principal Component Analysis, Support Vector
machine and Linear Discriminant Analysis.
3.2. Model Free Features
Model free approach extract the gait features for classification and analysis from the captured video or
image without the person direct involvement with the system. The general process involves person
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1640
detection, background subtraction, silhouette extraction, feature extraction and classification. Model
free approach is preferred for its low computational cost and better performance in low quality video.
Background subtraction is a process of classifying the pixel as foreground and background. Basic
background techniques are frame differencing and filtering which uses the concept of min, max,
mean, median and histogram filter. Gait image descriptors are body length and height, GEI, gradient
histogram energy image , Fourier descriptors, normal vectors, Principal Component Analysis, k-
means, HMM, standard deviation and frequency domain features , Gabor filters Multi-linear
Discriminant Analysis, Chrono-Gait Image, Linear Discriminant Analysis , Microsoft Kinect distance
method.
TABLE 1Datasets available for gait biometric analysis
METHOD DATASET DEVICE SUBJECTS DATA FORMAT
WEARABLE
SENSOR
OU-ISIR Inertial
Sensor
tri-axial
accelerometer
and gyroscope,
smartphone
744
389 – Males
355 – females,
Age- 2 to 87
ZJU-gaitacc Wii Remote -
5
175 Subjects
22-session 0
153- session 1
153- session 2
BWRMultiDevice
3 mobile
devices with
accelerometer
25 subjects
(session 1& 2)
CASIA D – FOOT
Foot pressure
measurement
plate + foot
scanner
88 subjects
VISION
BASED
OU-ISIR
Biometric
Database
camera 62 528,
Single view
OU-ISIR
Multiview
Database
7 cameras 10,307,
Multi view
CASIA A –
STANDARD
DATASET
camera
20 person ;
each with 12
sequence in
three
directions
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1641
CASIA B –
Multiview Gait
dataset
Captured in 11
views 124 subjects
CASIA C –
INFRARED
Infrared
thermal
camera
153 subjects
Tum Gaid Dataset
RGB Video ,
audio and
depth
305 subjects
DBMHI-DB2 Camera (front
and back) 100 subjects
Pathological
Dataset
Gait in
Neurodegenerative
Disease Database
force-sensitive
resistors
64 Subjects
(Parkinson
Disease -15 ,
Huntington
Disease – 20,
amyotrophic
lateral
sclerosis- 13,
Healthy – 16)
Daphnet Freezing
of Gait Data Set
Accelerator
Sensor
10 Parkinson
disease
Patient(3
measurements-
ankle, knee
and hip)
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1642
4. MACHINE LEARNING IN GAIT RECOGNITION
Machine learning concepts are required to build the intelligent system to learn, recognize, analyze the
gait characteristics with large number of image sequence collected to create the gait dataset. Figure
3gives the general process of gait analysis using machine learning algorithms. Artificial intelligence
techniques like data mining, machine learning, deep learning and automation are considered for the
innovation in gait analysis fields. In machine learning paradigm, learning models are like supervised
learning, unsupervised learning, reinforcement learning used in various design perspective to capture
the gait sequence, finding the gait pattern from the characteristic features extracted with respect to the
implementing field.
Machine learning approach follows the sequential procedure of data collection, feature extraction,
learning algorithm and the evaluation methodology. Feature selection is a process used to identify the
unique characteristics and eliminate the unnecessary information to reduce the problem of over fitting
and training time thereby increasing the accuracy of the model. Feature selection techniques which are
commonly used are univariate selection, feature importance, correlation matrix with heatmap and
filter methods. Filter and wrapper feature selection method is specially used for gait sequence images.
Wrapper methods is quite expensive since the learning algorithm repeats the random resampling
technique usage for obtaining the feature subset accurately. Filter method use heuristic search
procedure to obtain the undesirable feature subset. Without the help of learning algorithm, undesirable
features were removed in terms of correlation between the examined data. Popular filter method is
recursive feature elimination algorithm which removes the low weighted features. Feature subset were
fed into evaluation procedure which determines the quality by comparing with the previous best value
given by predefined stopping criterion and validated [22].
Model selection is the technique of selecting the appropriate model from the set of models which has
high classification or prediction accuracy. Model is selected with the help of statistical technique,
exploratory data analysis, scientific techniques and specification of a model. It is recommended that
learning approach should be selected along with model selection [23, 24]. Generally, dataset is divide
into three portions namely training data, cross-validation data and testing data. Training data is
utilized to train the model with the input and known outcome. Cross validation data helps for frequent
evaluation to improve the effectiveness of the model. Testing data is used for validating the specific
model. Selected model is estimated to perform well with the training data and computed with similar
learning model with various hyper parameters, complexities. Validation dataset is opted using the
methods: holdout method, k-fold cross validation, leave-one cross validation, random subsampling
cross validation and bootstrapping.
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1643
Fig. 3. Basic machine learning approaches used for Gait classification, prediction and analysis
4.1. Classification Network
Classification Network has statistical and derivative phase. Statistical phase extracts the statistical
measurements from the non-linear transformed input. Derivative phase improves the accuracy with
mathematical model. Classification networks are classified into shallow network and deep neural
network. Shallow neural network usessolitary hidden layer whereas deep neural network uses more
than one hidden layer. Shallow networks used for gait recognition are K-Neighbor, Hidden Markov
Model, Bayesian Network, Support Vector machine and BPNN. Deep neural network increases the
accuracy rate when compared to shallow network and has become more interesting approaches in
extracting features in recent years. Wang et al[33] used nearest neighbour approach to generate the
silhouette signal to identify the person and achieved 77.4% classification rate with CASIA B dataset.
Jeevan et al [34] recognizes the gait of an individual with Gait Pal and Pal Entrophy using Principal
Component Analysis with Support Vector machines. Implementing with treadmill dataset and CASIA
A, B and C datasets, achieved 92.3% accuracy. Xiao et al[35] recognizes the individual by their gait
sequence using Zernike moments and Back Propagation Neural Network (BPNN)
4.2. Deep Learning Approaches
In the family of machine learning algorithms, deep learning has become the most important methods
due to its wide applications. Deep architectures represent the high level abstractions using multiple
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1644
stacked learning neural network layers and separate network segment is used to learn separate data
representation. The energetic factors of deep neural network are feature learning, hierarchical and
distributed representations, computational speed with large dataset varies in the automation process
and analysis of depth. Multiple neural network layers are convolutional layer, sampling pooling
layers, activation layer and fully connected layer. The multiple arrangement of these layer and
dimensions varies corresponding to the application and architecture.
5. DEEP ARCHITECTURES
In deep learning approaches, there is no need to create a separate feature set since the discriminative
features were found while training the modelDeep architectures which are used widely for gait
recognition is convolutional neural network, multiple stacked auto encoder and deep Boltzmann
machine and Restricted Boltzmann Machine.
5.1. Deep Boltzmann Machine
A Deep Boltzmann Machine uses hidden random variables for multiple layers of symmetrically
couple stochastic Markov Random Field[44]. It has visible unit set {0,1}D and hidden layer series
units h(1) {0,1}F1, h(2) {0,1}F2 ,…… h(L) {0,1}FL . Units in adjacent layers have connection between
them. For instance, consider three hidden layers for DBM as shown in Figure 4. The probability of a
visible vector assigned by Deep Boltzmann Machine is given as (1).
𝑃(𝑣; 𝜃) = 1
(𝜃∑ 𝑒𝑥𝑝
ℎ
(∑ 𝑊𝑖𝑗(1)
𝑢𝑖
𝑖𝑗
ℎ𝑗(1)
+ ∑ 𝑊𝑗𝑙(2)
𝑗𝑙
ℎ𝑗(1)
ℎ𝑙(2)
+ ∑ 𝑊𝑙𝑚(3)
ℎ𝑙(2)
𝑙𝑚
ℎ𝑚(3)
)
(1)
In equation (1) h is a set of hidden units represented as {h(1) , h(2) , h(3)} and the model parameter =
{W(1) , W(2) , W(3)}. This represents symmetric interactions between the visible units and hidden units.
Restricted Boltzmann Machine model is recovered by assigning W(2) and W(3) as 0. Initially the
weights in hidden units are asymmetric as the weights are twice as the top layer. Later to maintain the
symmetry, weights are modified.
Fig. 4.a) Deep Boltzmann machine has undirected connection. b) DBM is pretrained with RBM stack.
The weights of first and last layers of RBM is modified as W(1) and W(3) to maintain the symmetric
pattern [44].
Yu et al[43] proposed the Convolutional Restricted Boltzmann Machine to extract the gait features to
identify the individuals. CRBM was experimented on CASIA B gait dataset with 11 different angles
and 3 variations in outfit obtained the accuracy of 96.5%.
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1645
5.2. Deep Stacked Autoencoder
Fig. 5. a) single autoencoder with same dimensionality of input layer (x1, x2, , x6) and output layer ,
lower dimensionality in feature extraction layer. b) Stacked autoencoder enables the supervised
classification in the final output layer [45].
Deep Stacked Autoencoder has more layers to extract the discriminative and representative features. It
is formed by stacking the autoencoder network one above the other and implemented multilayer
perceptron. A single autoencoder network has encoding layer for encoding the input and decoder layer
for reconstructing the input. In decoding layer, number of neurons and dimension of input are
considered to be in the same count. It recovers the input x with high accuracy using the two stage
approximation model. The code h used for recovering is
𝑓𝑑𝑒𝑐 (𝑓𝑒𝑛𝑐(𝑥)) = 𝑓𝑑𝑒𝑐(ℎ) = 𝑥 ≈ 𝑥
Where fenc is encoding function and fdec is the decoding function computed by their respective coding
layer. The main function of stacked autoencoder to train the unsupervised input layer by layer. Figure
5 shows how the single autoencoder turns to stacked autoencoder to perform the classification. The
final output layer produces the supervised classification by fine tuning the pre-trained network with
Back Propagation[45]. Sun et al [46] proposed the deep network by combining the convolutinal neural
network and stacked autoencoder to recognize the human with the gait subtle features. Micro-motion
dataset is a radar echo dataset collected using the Radio Frequency device is used to implement the
proposed network and achieved 96% as highest average precision.
5.3. Recurrent Neural Network
Recurrent Neural Network (RNN) is designed to solve classification problems. RNN deals with
sequential data with more interdependencies especially well-suited for clinical data.In RNN, input
sequence is termed as X1, X2,…., Xnand output sequence is termed as Y1, Y2,…., Yn. Output Vector Yi
depends on the past values of X and it is given by
𝑌𝑖 = 𝑓𝑖(𝑋1, … . . 𝑋𝑛)
Hidden states are utilized to learn parameters from the sequential data given in the previous and
current recurrence. The hidden state, hi and recurrent layer output, yi is calculated as
ℎ𝑖 = tanh(𝑊ℎℎℎ𝑖−1 + 𝑊𝑥ℎ𝑥𝑖 + 𝑏ℎ )
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1646
Where W denote weight, b denote biases and i denote the frame index.To solve the vanishing
gradient problem, Long Short Term Memory (LSTM) and Gated Recurrent units(GRU) is proposed.
These networks maintain the long term inter relations and nonlinear dynamics effectively.To control
the number of parameters in the network, same weight is determined for all layers.
Hidden state is updated with forget gate value fi, input gate value gi and output gate value oi..
Learning variables are given as follows:
𝑓𝑖 = 𝜎(𝑊𝑥𝑓𝑥𝑖 + 𝑊ℎ𝑓ℎ𝑖−1 + 𝑏𝑓)
𝑔𝑖 = 𝜎(𝑊𝑥𝑖𝑥𝑖 + 𝑊ℎ𝑖ℎ𝑖−1 + 𝑏𝑖)
𝐶𝑖 = 𝑓𝑡 ⃘ 𝐶𝑖−1 + 𝑖𝑡 ⃘ tanh(𝑊𝑥𝐶𝑥𝑖 + 𝑊ℎ𝐶ℎ𝑖−1 + 𝑏𝐶)
𝑜𝑖 = 𝜎(𝑊𝑥𝑖𝑥𝑖 + 𝑊ℎ𝑖ℎ𝑖−1 + 𝑏𝑖)
ℎ𝑖 = 𝑖𝑡 ⃘ tanh(𝐶𝑖)
Where W denote weights and b denote biases. Prado et al [65] used RNN to classify the cerebral
palsy disease in adults and children. Data were collected from 28 adults and 7children with cerebral
palsy using the pressure inertial foot sensor. They have proposed RNN with instrumented foot sensor
which does not require any preporocessing technique, thereby processing time is reduced to less than
1 second.
5.4. Convoultional Neural Network
Convolutional neural network (CNN) is multilayer perceptron in a regularized version effectively
implemented on Graphics Processing Unit (GPU) or General Purpose computing on Graphics
Processing Unit (GPGPU). GPU which is mainly developed for video game application is used to
increase the speed of deep Neural Network processing. Convolutional Neural network has
convolution layer, Pooling Layer, Activation function, Loss function, regularization, optimization
process. Figure6 demonstrates the general outline of convolutional neural network.First parameter of
the CNN is the input data in the form of image or signal. Second parameter is the filter. Fully
connected layer gives the output in the form of vector from which feature map is obtained. In the form
of multidimensional array, each input and filter are saved independently in the network. Local
information of the image is identified by convolutional layer.
Fig. 6. Schematic diagram of convolutional Neural network
Convolution layer is categorized into titled convolution, transposed convolution, dilated convolution,
network in network and inception module. Pooling layer is categorized into average pooling, max
pooling, Lp pooling , mixed pooling , stochastic pooling , spectral pooling , spatial pyramid pooling
and multi-scale order less pooling. Activation function used are Rectified Linear Unit (ReLu), Leaky
ReLU, parametric ReLU, Randomized ReLU, Exponential Linear Unit, Maxout and Probout.
Methods used for convolution neural network are as follows : LeNet 5 is a 8 Layered architecture
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1647
used for image recognition. AlexNet uses GPU to train the high resolution large dataset with 1000
different classes. Alexnet does faster processing compared to LeNet5 due to its non saturating
neurons. Overfitting problem is reducing by using dropout method. Zeiler and Fergus proposed ZFNet
as multi-layered deconvolutional neural network with 8 layers. ZfNet uses GPU to classify large
image dataset and dropout to reduce the parameters.Network – In-Network ( NIN) is proposed to
implement per-pixel nonlinearity and to reduce the number of parameters. NIN uses 1 x 1
convolutional network with ReLU activation. Global average pooling is used to produce logit
vectors.Region based Convolutional Neural Network (R-CNN) detects the object by extracting the
features with region proposals. Forward computions is required to choose the thousand number of
region proposals from a single image increase the processing time. To overcome this, Fast RCNN is
proposed to perform the forward computation as a whole. Entire image is given as an image for
processing each region proposals at a time unlike RCNN where region proposals are given as a input.
Faster R-CNN uses region proposal network, thereby reducing the number of generated region
proposal and maintaining the accuracy of object detection. Mask R-CNN uses fully convolutional
layer for locating the objects at pixel level. Comparatively Mask RCNN enhanced object detection
accuracy rather than Faster RCNN. GoogLeNet uses 22 convolutional layer for transfer learning.
Optimization is increased when image classification is done via GoogLeNet CNN model. Visual
Geometry Group (VGG) is widely used in gait recognition when compared to other CNN model since
it provides high accuracy. VGG performs training faster since it has reduced number of parameters.
Table 3 provides the overview of recent work in gait recognition using deep architectures.
TABLE 3Review of recent gait recognition using deep architectures.
YEA
R
REFERENC
E
APPROA
CH
GAIT
FEATUR
ES
DATASET CLASSIFIE
R PURPOSE
ACCURA
CY
2006 [40] Yu at al Model
free
GEI CASIA B Multilayer
autoencoder
Identificatio
n
-
2015 [28] Chao Yan
et al
Model
free
GEI CASIA B CNN Identificatio
n
96%
2015 [29] Perlin et al - Cloth and
gender
H3D
dataset,
ViPer
dataset,
HATdb
CNN Pattern for
clothes and
gender
-
2016 [48] Hannink et
al
Model
Based
acclerator
,
gyroscope
parameter
s
GAITRite CNN with
modifications
Identificatio
n
-
2017 [37] Alotaibi et
al
Model
free
CASIA B CNN, keras
,python
Identificatio
n
Mis-
classificatio
n rate -
<0.15%
2017 [41] Sokolova
and Konushin
Model
free
Optical
Flow
TUM-
GAID
dataset
CNN - VGG Identificatio
n
16 hours ,
97%
2017 [42] Dehzangi
et al
Model
Based
Gait
measurem
ents
Sensor
collected
data
Deep CNN Identificatio
n
97.60%
2017 [30] Alotaibi et Model GEI CASIA B Deep CNN Identificatio 98.30%
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1648
al free n
2018 [31] Gadaleta et
al
Model
Based
acclerator
,
gyroscope
parameter
s
GaitDataset
McGill CS
CNN Identificatio
n
94%
2018 [32] batchulunn
et al
Model
free
GEI DBMHI-
DB1
CNN - VGG Authenticati
on
Identificatio
n
100%
2018 [47] Liu et al Model
free
GEI CASIA B CNN - VGG Identificatio
n
89.62%
2018 [39] Wu et al Model
free
GEI CASIA B,
OU-ISIR
Deep CNN
with location
weight
descriptor
Identificatio
n
86%
2018 [53] Yuan et al Model
Based
acclerator
,
gyroscope
parameter
s
GITHUB CNN Authenticati
on
Identificatio
n
87%
2018 [54] Batchuluun
et al
Model
free
camera
image
Sensor
collected
data
DBMHI-
DB1,DBM
HI-DB2
LSTM CNN Authenticati
on
Identificatio
n
97%
2018 [55] Sun et al Model
Based
spectogra
m
Self
collected
dataset
from radar
and
microwave
anechoic
chamber
CNN, Stacked
Auto encoder
Authenticati
on
Identificatio
n
96.36%
2018 [43] Yu et al Model
free
GEI CASIA B CNN-
BOLTZMAN
N MACHINE
Identificatio
n
96.50%
2018 [57] Liu et al Model
free
GEI OU-ISIR CNN,
SIAMESE
NEURAL
NETWORK
Identificatio
n
88
2018 [49] Tong et al Model
free
GEI CASIA-B,
OU-ISIR,
and CMU
MoBo
SPATIAL
TEMPORAL
DEEP NN
Identificatio
n
95.67%
2018 [58] Takemura
et al
Model
free
GEI OU-ISIR CNN Identificatio
n
95%
2018 [59] Yao et al Model
free
Skeleton
Gait
image
CASIA B CNN Identificatio
n
92.09%
2018 [64] Verlekar et
al
Model
free
GEI DAI , DAI2
INIT
CNN – LSTM Clinical gait
disease
95%
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1649
classificatio
n
2018 [68] Wang et al Model
based
Threadnill
sensor
Recorded
with Body
weight
support
threadmill
CNN-
AlexNet
Gesture
recognition
99.98%
2019 [27] Maryam et
al
Model
free
GEI OULP and
Casia-B
CNN Identificatio
n
-
2019 [60] Wang et al Model
free
GEI CASIA B ,
OU-ISIR
CNN- 2
BRANCH
Authenticati
on
94%
2019 [52] Hawas et al Model
free
GEI CASIA B CNN Identificatio
n
95%
2019 [56] Wang et al Model
free
GEI CASIA A,
B, OU-
ISIR
CNN Identificatio
n
95%
2020 [63] Shao et al Model
free
GEI CASIA B CNN , LeNet5 Identificatio
n with
Covariates
98.3%
(normal)
89.2 % (
Backpack)
95.8% (
Wearing
coat)
Human Gait Recognition:
Hannink et al [48] used convolutional neural network with ReLU activation function to extract the
gait feature from wearable sensor dataset. Tong et al [49] used spatial-temporal convolutional neural
network to recognize the gait in multi-view angle with 92.54% accuracy. Liu et al[47] combined the
convolutional neural network and support vector machine to predict the gender of an individual using
the CASIA B dataset with 87.94%. The convolution model VGGNet-16 is used to train the network
with GEI input while SVM classify the descriptors provided by optimizing the weight parameters.
Shao et al [63] proposed improved LeNet5 network combined with Residual Network to accelerate
the convergence speed and to improve the recognition rate. CASIA B dataset is trained and tested
with the improved LeNet model of CNN. They focused to work on the covariates like Viewing angle,
backpack, jacket and speed. The accuracy of gait recognition with backpack, wearing coat and normal
are 89.2%, 95.8% and 98.3% respectively. Maryam et al [27] used FCN (Fully Convolutional Neural
Network) to reconstruct the incomplete Gait Energy Image(GEI) into a complete GEI. Experiments
were carried out in OU-ISIR Biometric – OULP database and CASIA B database to obtain the
complete GEI from 0.1 part of a gait cycle. This method helps in matching view-variant incomplete
gait image.
Clinical Gait Analysis: Verlekar[64] proposed a fine tuned VGG-19 Deep architecture to classify gait disease type. The
pathological dataset with diplegia, Hemiplegia, neuropathy, parkinson and normal individual gait
dataset is used to experiment the proposed model. The classification of normal and impaired gait type
is performed with high accuracy. Wang et al [64] used Alexnet CNN model to perform the gesture
recognition system with 99.98% accuracy. Five different gestures such as moving forward, moving
backward, moving left, moving right and stop are recognized with Body Weight Support Treadmill
(BWST) Platform. This recognition system helps in lower limb rehabilitation. This system is
controlled by a trainer to monitor the patient walking sequence in BWST platform.
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1650
6. DISCUSSION
A comprehensive survey of gait recognition and analysis methods based on wearable sensors and
vision based methods, various features and deep architecture techniques were presented. A review of
gait databases with different techniques used to capture gait features were given. Figure 7 depicts the
recurrently used gait features in various fields and obtained the best accuracy with the recent
techniques of deep architectures.
(a) (b)
(c)
Fig. 7. Statistical figures based on the referred papers a) gait features and usage b) Gait analysis
applied in the various fields c) Different combinations of Deep architectures used in gait recognition.
Future Research Directions.
The scope of future research provides an insight into remote health monitoring, surveillance and
robotics. Immense opportunities are open to enhance and innovate in the current technological
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1651
environment in the following fields. Application of Multivariate statistical analysis and deep learning
techniques is becoming increasing popular in the area of clinical gait analysis, thereby providing new
insights into human gait control. According to Global Burden of Disease (GDB), musculoskeletal
disorder is estimated to witness the second highest growth rate across the globe. In recent years,
number of people with abnormal walking posture is increasing. Research were carried out to
differentiate the healthy walking posture and abnormal walking posture. Sports professionals utilized
gait analysis to monitor the movement of leg in terms of angle, position etc. The measurement and
assessment of the players improved their performance as well as decreased the risk of injury. Model
free method is preferred quickly to interpret about the movement of the players. Trajectory
movements of model based human gait helps to design the walking nature of robots.Human gait
recognition system is becoming widely spread due to its nature of identifying individual without their
intervention with the system. Still there exist some limitations such as impact of environmental
changes, user location and orientation, sensing multi-user activity and security considerations. To
perform the longitudinal study over the wide range of covariates, there is a need for a large multi-
modality, multi-covariate standard dataset. Our future work is to carry out the experiments with the
recent network architectures and novel approaches with the huge dataset of different view angles,
clothing and walking style promoting the better automatic gait recognition and gait analysis system.
7. CONCLUSION
The most interesting research in biometrics is automatic gait recognition when compared to other
human unique features. Though there are many approaches to overcome the variations in gait
recognition, still there are challenges to recognize a person. Challenges are distinctive gait datasets,
degree of stability in identifying, sensing modality, covariates and spoofing effects, and exploring
new algorithms. From the developmental perspective, human gait recognition has gained maturity in
adopting recent methodologies to provide the high accuracy and the same was analyzed and studied in
the view of vision based system. In this paper, the basic knowledge about the human gait is identified
and explored. The comparative analysis with the existing techniques of gait recognition were
discussed. In this work, the survey of recent deep architecture model on human gait identification,
authentication and clinical applications were discussed. Future research directions and challenges
were identified.
8. REFERENCES
[1] http://classics.mit.edu/Aristotle/gait_anim.html
[2] George A. Bekey, Fellow, Ieee, Chi-Wu Chang, Jacqueline Perry, And M. Mark Hoffer
,“Pattern Recognition of Multiple EMG Signals Applied to the Description of Human Gait”,
1977.
[3] In-Sheng Cheng, S. H. Koozekanani, Member, IEEE, and M. T. Fatehi, Student Member, IEEE,
“A Simple Computer-Television Interface System for Gait Analysis”,1974
[4] Roger Clarke. 1994, “Human identification in information systems”, Management challenges
and public policy issues. Info. Technol. People 7, 4 (1994), 6–37.
[5] Christopher L. Vaughan, Brian L. Davis, C. O. Jeremy et al. 1999. “Dynamics of human gait”,
Human Kinetics Publishers.
[6] Armin Bruderlin, Thomas W. Calvert, School of Computing Science Simon Fraser University
Burnaby, British Columbia, Canada V5A 1S6, “Goal-Directed, Dynamic Animation of Human
Walking”, Computer Graphics, Volume 23, Number 3, July 1989.
[7] Michael McKenna and David Zeltzer, “Dynamic Simulation of Autonomous Legged
Locomotion” , Computer Graphics and Animation Group The Media Laboratory Massachusetts
Institute of Technology Cambridge, MA , Computer Graphics, Volume 24, Number 4, August
1990
[8] Monika Köhle, Dieter Merkl, “Identification of gait patterns with self-organizing maps based
on ground reaction force”, Institute of Software Technology, Vienna University of
Technology, Proc. of the European Symposium on Artificial Neural Networks (ESANN’96).
Bruges. Belgium. April 24-26. 1996.
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1652
[9] Monika Kohle, Dieter Merkl, “Clinical Gait Analysis by Neural Networks: Issues and
Experiences” , Institut fur Softwaretechnik, Technische Universitat Wien, A-1040 Wien,
Austria, Tenth IEEE Symposium on Computer-Based Medical Systems,1997
[10] Monika KOhle, Dieter Merkl ,“Analyzing Human Gait Patterns for Malfunction Detection”,
Institut for Softwaretechnik Technische UniversitSt Wien Favoritenstral~e 9-11/188 A-1040
Wien, Austria, SAC '00: Proceedings of the 2000 ACM symposium on Applied computing -
Volume 1, 2000
[11] Harold C. Sun, Dimitris N. Metaxas, “Automating gait generation”, University of
Pennsylvania, ACM SIGGRAPH 2001, 12-17 August 2001, Los Angeles, CA, USA, 2001
[12] Aaron E Bobick, Amos Y Johnson , “Gait Recognition Using Static, Activity-Specific
Parameters” , GW Center/College of Computing Georgia Tech Atlanta, Electrical and
Computer Engineering Georgia Tech Atlanta, GA 30332, Proceedings of the 2001 IEEE
Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001.
[13] Arunachalam Somasundaram, Rick Parent,“3D Reconstruction Of Walking Behaviour Using A
Single Camera”, SIGGRAPH Abstracts and Applications 2002
[14] Chiraz BenAbdelkadery, Ross Cutlerz, and Larry Davisy, “Motion-based Recognition of
People in EigenGait Space”, University of Maryland, College Park, Proceedings of the Fifth
IEEE International Conference on Automatic Face and Gesture Recognition (FGR02) ,2002
[15] James B. Hayfron-Acquah, Mark S. Nixon, John N. Carter, “ Automatic gait recognition by
symmetry analysis”,Image, speech and Intelligent Systems Group, Department of Electronics
and computer science, University of Southampton, Pattern Recognition Letters 24(2003) 2175-
2183, 2003
[16] Andrew Crossan, Roderick Murray-Smith, Stephen Brewster, James Kelly, Bojan Musizza,
“Gait Phase Effects in Mobile Interaction”, CHI 2005 , Late Breaking Results: Posters, April
2-7 , Portland, Oregon, USA, 2005
[17] Anuradha Annadhorai, Eric Guenterberg, Jaime Barnes, Kruthika Haraga and Roozbeh Jafari
Department of Computer Engineering The University of Texas at Dallas, Richardson, TX –
75083, USA , “Human Identification by Gait Analysis”, HealthNet'08, June 17, 2008,
Breckenridge, CO Copyright 2008 ACM ISBN 978-1-60558-199-6/08/06,2008.
[18] Yang Hu ; Konstantinos Sirlantzis ; Gareth Howells, “Optimal Generation of Iris Codes for Iris
Recognition”, IEEE Transactions on Information Forensics and Security , Volume: 12 , Issue: 1
, Jan. 2017
[19] Imed Bouchrika and Mark S. Nixon, “Model-Based Feature Extraction for Gait Analysis and
Recognition”, in:proceedings of Computer Vision/Computer Graphics Collaboration
Techniques, Springer Berlin Heidelberg,2007, pp.150-160.
[20] C. Yam, M. Nixon, J. Carter, Automated person recognition by walking and running via model-
based approaches, Pattern Recognition.37(5)(2004) 1057 -1072.
[21] A.F. Bobick, A.Johnson, Gait recognition using static , activity-specific parameters,in:
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern
Recognition, 1,2001,pp.423-430.
[22] C.Yam, M.Nixon, Gait recognition, model-based,Encylopedia of Biometrics,2015,pp,799-805.
[23] Dash, Manoranjan, and Huan Liu. 1997. Feature selection for classification. Intelligent Data
Analysis 1(3): 131–156.
[24] Guyon, Isabelle, Amir Saffari, Dror Gideon, and Gavin Cawley. 2010. Model selection:
Beyond the bayesian/frequentist divide. Journal of Machine Learning Research 11: 61–87.
[25] Model Selection. 2015. http://bactra.org/notebooks/model-selection.html.
[26] Michael Otero. 2005. Application of a continuous wave radar for human gait recognition.
Proceedings of SPIE—The International Society for Optical Engineering 5809, 538–548
[27] Maryam Babaee , Linwei Li , Gerhard Rigoll, "Person identification from partial gait cycle
using fully convolutional neural networks", Neurocomputing 338 (2019) 116–125
[28] C. Yan , B. Zhang , F. Coenen , Multi-attributes gait identification by convolutional neural
networks, in: Proceedings of the Eighth International Congress on Im- age and Signal
Processing (CISP), IEEE, 2015, pp. 642–647
[29] Perlin, H. A., & Lopes, H. S. (2015). Extracting human attributes using a convolutional neural
network approach. Pattern Recognition Letters, 68, 250–259
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1653
[30] Alotaibi, M., & Mahmood, A. (2017). Improved gait recognition based on specialized deep
convolutional neural network. Computer Vision and Image Understanding, 164, 103–110.
[31] Gadaleta, M., & Rossi, M. (2018). IDNet: Smartphone-based gait recognition with
convolutional neural networks. Pattern Recognition, 74, 25–37.
doi:10.1016/j.patcog.2017.09.005
[32] Batchuluun, G., Naqvi, R. A., Kim, W., & Park, K. R. (2018). Body-movement-based human
identification using convolutional neural network. Expert Systems with Applications, 101, 56–
77. doi:10.1016/j.eswa.2018.02.016
[33] Wang L et al (2003) Silhouette analysis-based gait recognitionfor human identification. IEEE
Trans Pattern Anal Mach Intell 25(12):1505–1518
[34] Jeevan M, et al (2013) Gait recognition based on gait pal and pal entropy image. in Image
Processing (ICIP), 2013 20th IEEE International Conference on. IEEE
[35] Degui Xiao and Lei Yang. 2008. Gait recognition using Zernike moments and bp neural
network. In IEEE International Conference on Networking, Sensing and Control. 418–423.
[36] D. Liu , M. Ye , X. Li , F. Zhang , L. Lin , Memory-based gait recognition, in: Pro- ceedings of
the British Machine Vision Conference (BMVC), 2016 .
[37] X. Ben , P. Zhang , W. Meng , R. Yan , M. Yang , W. Liu , H. Zhang , On the distance metric
learning between cross-domain gaits, Neurocomputing 208 (2016) 153–164 .
[38] S. Yu , H. Chen , Q. Wang , L. Shen , Y. Huang , Invariant feature extraction for gait
recognition using only one uniform model, Neurocomputing 239 (2017) 81–93 .
[39] Wu, H., Weng, J., Chen, X., & Lu, W. (2018). Feedback weight convolutional neural network
for gait recognition. Journal of Visual Communication and Image Representation, 55, 424–432.
doi:10.1016/j.jvcir.2018.06.019
[40] Shiqi Yu, Daoliang Tan, and Tieniu Tan. 2006. A framework for evaluating the effect of view
angle, clothing and carrying condition on gait recognition. In International Conference on
Pattern Recognition. 441–444.
[41] A. Sokolova and A. Konushin. 2017. Gait recognition based on convolutional neural networks.
In International Archives of the Photogrammetry, Remote Sensing & Spatial Information
Sciences. 42, 42, 207–212.
[42] O. Dehzangi, M. Taherisadr, and R. Changalvala. 2017. IMU-based gait recognition using
convolutional neural networks and multi-sensor fusion. Sensors 17, 12, 2735.
[43] Yu, T., Liu, X., Mei, Y., Dai, C., & Yan, J. (2018). Identification by Gait Using Convolutional
Restricted Boltzmann Machine and Voting Algorithm. 2018 IEEE International Conference on
Internet of Things (iThings) and IEEE Green Computing and Communications (GreenCom)
and IEEE Cyber, Physical and Social Computing (CPSCom) and IEEE Smart Data
(SmartData).
[44] Hinton, Geoffrey; Salakhutdinov, Ruslan (2012). “A better way to pretrain deep Boltzmann
machines” NIPS'12: Proceedings of the 25th International Conference on Neural Information
Processing Systems.
[45] Lukas Vareka and pavel Mautner “Stacked Autoencoders for the P300 Component Detection”,
Neuroinformatics Research Group, Department of Computer Science and Engineering, Faculty
of Applied Sciences, University of West Bohemia, Pilsen, Czechia, (2017).
[46] Sun, L., Yuan, Y.-X., Zhang, Q., & Wu, Y.-C. (2018). Human Gait Classification Using Micro-
Motion and Ensemble Learning. IGARSS 2018 - 2018 IEEE International Geoscience and
Remote Sensing Symposium.
[47] Liu, T., Ye, X., & Sun, B. (2018). Combining Convolutional Neural Network and Support
Vector Machine for Gait-based Gender Recognition. 2018 Chinese Automation Congress
(CAC).
[48] Hannink, J., Kautz, T., Pasluosta, C. F., Gasmann, K.-G., Klucken, J., & Eskofier, B. M.
(2017). Sensor-Based Gait Parameter Extraction With Deep Convolutional Neural Networks.
IEEE Journal of Biomedical and Health Informatics, 21(1), 85–93.
[49] Tong, S., Fu, Y., Yue, X., & Ling, H. (2018). Multi-View Gait Recognition Based on A
Spatial-Temporal Deep Neural Network. IEEE Access, 1–1.
[50] Mason, James Eric, Issa Traore, and Isaac Woungang. "Gait Biometric Recognition", Machine
Learning Techniques for Gait Biometric Recognition, 2016.
European Journal of Molecular & Clinical Medicine ISSN 2515-8260 Volume 07, Issue 06, 2020
1654
[51] Singh, M., Aujla, G. S. S., Singh, A., Kumar, N., & Garg, S. (2020). Deep Learning based
Blockchain Framework for Secure Software Defined Industrial Networks. IEEE Transactions
on Industrial Informatics, 1–1.
[52] Hawas, A. R., El-Khobby, H. A., Abd-Elnaby, M., & Abd El-Samie, F. E. (2019). Gait
identification by convolutional neural networks and optical flow. Multimedia Tools and
Applications.
[53] Wei Yuan and Linxuan Zhang, Gait Classification and Identity Authentication Using CNN.
Methods and Applications for Modeling and Simulation of Complex Systems, 2018, Volume
946,pp 119-128.
[54] Batchuluun, G., Yoon, H. S., Kang, J. K., & Park, K. R. (2018). Gait-based Human
Identification by Combining Shallow Convolutional Neural Network-stacked Long Short-term
Memory and Deep Convolutional Neural Network. IEEE Access, 1–1.
[55] Sun, L., Yuan, Y.-X., Zhang, Q., & Wu, Y.-C. (2018). Human Gait Classification Using Micro-
Motion and Ensemble Learning. IGARSS 2018 - 2018 IEEE International Geoscience and
Remote Sensing Symposium.
[56] Wang, X., Zhang, J., & Yan, W. Q. (2019). Gait recognition using multichannel convolution
neural networks. Neural Computing and Applications.
[57] Liu, W., Zhang, C., Ma, H. et al. Learning Efficient Spatial-Temporal Gait Features with Deep
Learning for Human Identification. Neuroinform 16, 457–471 (2018).
[58] Takemura, N., Makihara, Y., Muramatsu, D. et al. Multi-view large population gait dataset and
its performance evaluation for cross-view gait recognition. IPSJ T Comput Vis Appl 10, 4
(2018).
[59] L. Yao, W. Kusakunniran, Q. Wu, J. Zhang and Z. Tang, "Robust CNN-based Gait Verification
and Identification using Skeleton Gait Energy Image," 2018 Digital Image Computing:
Techniques and Applications (DICTA), Canberra, Australia, 2018, pp. 1-7
[60] Wang, X., Zhang, J. Gait feature extraction and gait classification using two-branch CNN.
Multimed Tools Appl 79, 2917–2930 (2020).
[61] L. Hoyet, C. Spies, P. Plantard, A. Sorel, R. Kulpa and F. Multon, "Influence of Motion Speed
on the Perception of Latency in Avatar Control," 2019 IEEE International Conference on
Artificial Intelligence and Virtual Reality (AIVR), San Diego, CA, USA, 2019, pp. 286-2863,
doi: 10.1109/AIVR46125.2019.00066.
[62] Mazilu, S., Hardegger, M., Zhu, Z., Roggen, D., Troester, G., Plotnik, M., & Hausdorff, J.
(2012). Online Detection of Freezing of Gait with Smartphones and Machine Learning
Techniques. Proceedings of the 6th International Conference on Pervasive Computing
Technologies for Healthcare.
[63] Shao, X., Nie, X., Zhao, X., Zheng, R., & Guo, D. (2020). Gait Recognition based on
Improved LeNet Network. 2020 IEEE 4th Information Technology, Networking, Electronic
and Automation Control Conference (ITNEC). doi:10.1109/itnec48623.2020.9084747
[64] Ver
[65] Prado, A., Cao, X., Robert, M. T., Gordon, A. M., & Agrawal, S. K. (2019). Gait
Segmentation of Data Collected by Instrumented Shoes Using a Recurrent Neural Network
Classifier. Physical Medicine and Rehabilitation Clinics of North America.
doi:10.1016/j.pmr.2018.12.007
[66] Hu, K., Wang, Z., Wang, W., Martens, K. A. E., Wang, L., Tan, T., … Feng, D. D. (2019).
Graph Sequence Recurrent Neural Network for Vision-based Freezing of Gait Detection. IEEE
Transactions on Image Processing, 1–1. doi:10.1109/tip.2019.2946469
[67] Jun, K., Lee, D.-W., Lee, K., Lee, S., & Kim, M. S. (2020). Feature Extraction Using an RNN
Autoencoder for Skeleton-Based Abnormal Gait Recognition. IEEE Access, 8, 19196–19207.
[68] Wang, P., Zhang, Q., Li, L., Ru, F., Li, D., & Jin, Y. (2018). Deep learning-based gesture
recognition for control of mobile body-weight support platform. 2018 13th IEEE Conference
on Industrial Electronics and Applications (ICIEA). doi:10.1109/iciea.2018.8398001