Accurate metaheuristic deep convolutional structure for a robust human gait recognition

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 13, No. 6, December 2023, pp. 7005~7015
ISSN: 2088-8708, DOI: 10.11591/ijece.v13i6.pp7005-7015  7005
Journal homepage: http://ijece.iaescore.com
Accurate metaheuristic deep convolutional structure for a
robust human gait recognition
Reem Nehad Yousef1
, Abeer Tawkool Khalil2
, Ahmed Shaaban Samra2
, Mohamed Maher Ata3
1
Delta Higher Institute for Engineering and Technology, Mansoura, Egypt
2
Electronics and Communications Engineering Department, Faculty of Engineering, Mansoura University, Mansoura, Egypt
3
Department of Communications and Electronics Engineering, MISR Higher Institute for Engineering and Technology, Mansoura, Egypt
Article Info ABSTRACT
Article history:
Received May 29, 2023
Revised Jul 5, 2023
Accepted Jul 9, 2023
Gait recognition has become a developing technology in various security,
industrial, medical, and military applications. This paper proposed a deep
convolutional neural network (CNN) model to authenticate humans via their
walking style. The proposed model has been applied to two commonly used
standardized datasets, Chinese Academy of Sciences (CASIA) and Osaka
University-Institute of Scientific and Industrial Research (OU-ISIR). After
the silhouette images have been isolated from the gait image datasets, their
features have been extracted using the proposed deep CNN and the
traditional ones, including AlexNet, Inception (GoogleNet), VGGNet,
ResNet50, and Xception. The best features were selected using genetic, grey
wolf optimizer (GWO), particle swarm optimizer (PSO), and chi-square
algorithms. Finally, recognize the selected features using the proposed deep
neural network (DNN). Several performance evaluation parameters have
been estimated to evaluate the model’s quality, including accuracy,
specificity, sensitivity, false negative rate (FNR), and training time.
Experiments have demonstrated that the suggested framework with a genetic
feature selector outperforms previous selectors and recent research, scoring
accuracy values of 99.46% and 99.09% for evaluating the CASIA and
OU-ISIR datasets, respectively, in low time (19 seconds for training).
Keywords:
CASIA gait dataset
Convolutional neural network
Gait recognition
Genetic algorithm
OU-ISIR
This is an open access article under the CC BY-SA license.
Corresponding Author:
Mohamed Maher Ata
Department of Communications and Electronics Engineering, MISR Higher Institute for Engineering and
Technology
Mansoura, Egypt
Email: mmaher844@yahoo.com
1. INTRODUCTION
Human gait refers to the variations in posture of the entire body, particularly the legs, during
walking. Gait recognition and its applications have attracted much attention in the previous decade.
Compared to other biological features like deoxyribonucleic acid (DNA), face, fingerprint, iris, and so on,
gait has several advantages, which are represented in, it is non-invasive, identifies humans from long
distances, does not require subject cooperation, and is resistant to low-resolution images. Moreover, it is
difficult to conceal because it is based on gait features of entire-body posture changes while walking; it is
hard to hide with gloves and masks such as fingerprints and face [1]. The advantages of the gait
authentication technique have led to its wide variety of applications, including banks, hospitals, airports,
trade centers, and military bases [2].
Gait identification has several benefits, but certain negative aspects are brought on by changes in
clothes or carrying objects, which affect recognition accuracy [3]. Due to this, a number of recent studies

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 13, No. 6, December 2023: 7005-7015
7006
have focused on enhancing gait detection traits by model-based and appearance-based techniques using deep
learning. In this paper, after the pedestrian frames have been segmented from the background, enhanced, and
normalized. Its features have been extracted using the proposed convolutional neural network (CNN) and
other pre-trained models, then selected the best of them by genetic, particle swarm optimizer (PSO), grey
wolf optimizer (GWO), and chi-square models. Finally, recognize these features using the proposed deep
neural network (DNN) model. The main contribution of the proposed methodology can be listed as follows:
i) isolating pedestrians from the data gait video frames, improving and normalizing them; ii) proposing the
well-designed CNN structure for extracting the gait features from Chinese Academy of Sciences (CASIA)
and Osaka University-Institute of Scientific and Industrial Research (OU-ISIR) datasets; iii) selecting the
significant features using genetic, PSO, GWO, and Chi-square models; iii) proposing a DNN for identifying
the gait features accurately; and iv) comparing the gait recognition performance metrics with the recent
studies and the other pre-trained models. The remaining article has the following structure: the methodology
for the proposed model is described in section 2, the proposed CNN model and the feature selector models
are explained in section 3, the main results of the study are presented in section 4, and the study’s
conclusions and recommendations are discussed in section 5.
Wen and Wang [1] have proposed a gait authentication model by extracting a gait energy image
from video frames and dimensionality reduction by sparse linear subspace. Then, classify them by support
vector machine model, which uses Gaussian radial basis function (RBF) kernels to attain cross-view gait
identification. Finally, the performance of the proposed gait recognition system is demonstrated using
CASIA and OU-ISIR datasets. Their experimental results scored 90.5% and 91.3% of accuracy values on
CASIA-B and OU-ISIR datasets, respectively. George et al. [4] presented a color-mapped contour gait image
as a gait characteristic for tackling cofactor variations and the identification of gait across perspectives.
Moreover, they compared and chose the best edge detection methods for gait blueprint generation on the
CASIA-B dataset and the OULP database. Their experimental results significantly improve gait recognition
for different views.
Premalatha and Chandramani [5] proposed a gait authentication technique to overcome the gait
covariate factors. Their model extracted the gait features from silhouette images using a histogram of
oriented gradients (HOG) algorithm and then recognized them by a K-nearest neighbor (K-NN) technique.
The evaluation result scored a high identification rate when applied to the CASIA B gait database, ranging
from 95.5% to 98.7% of accuracy value. Sokolova and Konushin [6] have presented a pose-based CNN
model for gait authentication. Firstly, they estimated the optical flow among gait frames to extract the motion
information. They then proposed a DNN model for computing pose-based gait descriptors. Finally, the
experimental experiments show that their study outperforms state-of-the-art studies.
Parashar et al. [7] used a pose-based gait recognition algorithm to attempt people wearing an
overcoat, carrying things, or other factors. It intends to use CNN to estimate human movement. They used a
multilayer recurrent neural network (RNN) to recognize gait datasets from the pose images. Moreover, they
presented a graphical user interface (GUI) for gait detection to be accurate, easy, and quick application. The
experimental result offered an accuracy ranging from 60.88% to 95.23%. Chen et al. [8] proposed a gait
identification framework called gait aggregation multi-feature representation (GaitAMR) for obtaining the
most selective subject features. The GaitAMR extracts global and local body movement descriptors by a
complete and partial temporal accumulation technique, besides effectively collecting feature representations
from many domains. Experimental results on the most common gait datasets prove that GaitAMR enhances
gait recognition and outperforms the other current studies.
In order to collect information about human locomotion, Dong et al. [9] proposed a recognition
framework based on the feature fusion of human gait frames and phases. Support vector machines (SVM),
backpropagation (Bp) neural networks, AlexNet, and LeNet5 algorithms are just a few examples of the deep
learning models and machines that have been utilized to recognize and assess the fused characteristics. The
proposed framework’s efficacy was demonstrated by the proposed model’s accuracy of 97.7% for gait phase
recognition. In order to recognize gait phases during human movement, Zhang et al. [10] created a model
called SBLSTM that integrates DNN, bidirectional long short-term memory (BiLSTM), and sparse
autoencoder (SAE) models. The four phases of the gait cycle heel strike, foot flat, heel off, and swing phase
can be distinguished using this model. The BiLSTM learns the temporal patterns and periodic changes in the
gait image after the SAE extracts the gait data. The gait stages are identified, and DNN obtains the output
classification results. The presented model demonstrates that the SBLSTM algorithm is successful at
recognizing gaits, and its performance results outperform those of other algorithms, supporting this claim.
Altilio et al. [11] applied various machine learning classification methods on the gait dataset to
classify patient motions from gait frames automatically by probabilistic neural network (PNN) technique.
The proposed model achieved an accuracy rating of 91% for home-based and cost-effective rehabilitation
techniques. In research [12], the human gait Frames of the CASIA (B) database have been classified using a

Int J Elec & Comp Eng ISSN: 2088-8708 
Accurate metaheuristic deep convolutional structure for a robust human gait … (Reem Nehad Yousef)
7007
SVM and a HOG models. First, they extract silhouette images, then create gait energy images (GEIs) to use
as input for the classifier models. The evaluation results achieve an accuracy value of 87.9% in wearing a
coat and a mean value of 83.33% for all classes.
2. METHOD
The main procedures of the proposed gait recognition technique have been depicted in Figure 1,
which contains five main steps: data gathering and preprocessing, feature extraction, feature selection, and
image label recognition. After the dataset had been collected and preprocessed, it is features were extracted
using the proposed CNN architecture and several pre-trained models, including, AlexNet, GoogleNet, Vgg,
ResNet50, and Xception. Then, the best among them were selected using PSO, GWO, Chi-square, and the
genetic model. Finally, recognize them by the fine-tuned DNN. The following subsections will thoroughly
demonstrate the approach of the suggested framework using the CASIA and OU-ISIR gait datasets.
Figure 1. The suggested gait authentication model framework
2.1. Gait dataset
Several datasets have been created for gait authentication research, such as CASIA and OU-ISIR.
The Institute of Automation Chinese Academy has established the CASIA dataset, which includes the four
datasets A, B, C, and D. Over 20K frames in the CASIA gait dataset [13] vary in pace, view angle, and
garment changes. The Institute of Scientific and Industrial Research developed the OU-ISIR gait dataset [14],
which contains 63,846 people ages 2 to 90 who walked along a path captured by the camera at 30 frames per
second. The sample of gait datasets from CASIA and OU-ISIR is shown in Figures 2(a) to 2(d).
2.2. Pre-processing criteria
The preprocessing steps of this study are shown in Figure 3 and are mentioned below. The main aim
of the preprocessing approaches is to increase image quality, minimize distortion, and reduce disturbance in
order to identify the images with clearly discernible silhouettes from the backdrop. The proposed procedures
of gait image preprocessing pass through the following steps:

 ISSN: 2088-8708
7008
Step 1: The gait frames were reduced to 140×140 to reduce the computing time and then converted to gray images.
Step 2: Subtract each frame from its background to obtain silhouette images.
Step 3: Contrast silhouette images using the histogram equalization model [15].
Step 4: Binarize result frames by the Otsu threshold [16] technique, then apply some morphological operations
such as dilating and filling to remove contaminants from the isolated binary successive frames using a
disk structure element (SE) of radius one.
Step 5: Normalize all binary images [17] and merge one cycle from frames to obtain GEIs for each individual.
Step 6: Split the result images into three groups: training (70%), validation (15%), and testing (15%).
(a) (b) (c)
(d)
Figure 2. The sample of dataset (a) CASIA-A with view angle 0º, (b) CASIA-A with view angle 45º,
(c) CASIA-A with view angle 90º, and (d) OU-ISIR with different ages
Figure 3. The preprocessing steps of the proposed model
2.3. Feature extraction and selection criteria
After the gait images have been isolated, enhanced, and normalized, their features have been
extracted using the proposed CNN model called deep gait network (DGNet). Moreover, the other features
have been extracted using the standard pre-trained models, including AlexNet, GoogleNet, VGGNet,
ResNet50, and Xception. These models contain layers comprising linear and nonlinear processes learned in a
joint modality. Convolution, pooling, and fully connected layers are the most prevalent layers in any CNN;
the convolution layer extracts the features, the pooling layer downsamples the feature map, and the fully
connected layer flattens the result downsample feature map. Six convolutional layers and three max pool
layers constitute the DGNet. Each layer is stacked after two convolutional layers with mask sizes 2×2. Each
convolution layer has 32, 64, and 128 filters with a 3×3 filter size, one stride, and a rectified linear unit
(ReLU) activation function, respectively. As illustrated in Figure 4, the output features were flattened using a
fully connected layer, a dense layer with a depth of 512, and a dropout layer to prevent overfitting. The most
salient features among the extracted ones from the proposed and pre-trained models have been selected by
PSO, GWO, Chi-square, and genetic models, which have the following characteristics,
2.3.1. GWO
A popular type of meta-heuristic algorithm that examines the gray wolves’ motions while pursuing
prey, which contains four kinds of wolves called: Alpha (α), Beta (β), Delta (δ), and Omega (ω). The GWO

7009
algorithm’s optimal solutions are Alpha (Gα), Beta (Gβ), and Delta (Gδ), where Alpha is the best, and the
remaining two are the second and third best. Generally, the GWO prediction process is divided into three
stages: surrounding, hunting, and attacking. When the grey wolves sense the presence of prey, they round it
using the (1) to (4).
𝐸
⃗ = |𝑑. 𝑀
⃗⃗ 𝑝(𝑡) − 𝑀
⃗⃗ (𝑡)| (1)
𝑀
⃗⃗ (𝑡 + 1) = 𝑀
⃗⃗ 𝑝(𝑡) − 𝐵
⃗ . 𝐷
⃗
⃗ (2)
where 𝐵
⃗ and 𝑑 are the coefficient vectors, 𝑀
⃗⃗ 𝑝 is the prey’s position vector, and 𝑀
⃗⃗⃗⃗ refers to a wolf’s position
vector. The vectors 𝐵
⃗ and 𝑑 are computed by (3) and (4) [18]:
𝐵
⃗ = 2𝑏
⃗ . 𝑟1 − 𝑏
⃗ (3)
𝑑 = 2. 𝑟2 (4)
where 𝑏
⃗ and its features are linearly value from 2 to 0 throughout the course, and r1 and r2 are chosen random
values ∈ [0, 1].
2.3.2. PSO
In 1995, Kennedy and Eberhart created the PSO technique according to swarm intelligence (SI), which
finds the problem’s optimal solution. It imitates the social behavior of groups of live organisms, such as fish
schools and bird colonies, to find food. It starts by randomly scattering all the group members, then updates
their paths based on the nearest spot to the food. Similarly, PSO [19] begins with several solutions, each with a
unique velocity and position, then is updated later based on interactions with other swarm members until it
reaches the optimal solution. Mathematically, the speed updated by (5) and position from (6) [20],
𝑣𝑖𝑗(𝑡 + 1) = 𝑣𝑖𝑗(𝑡) + 𝑟1 ∗ (𝑃𝑏𝑒𝑠𝑡𝑖𝑗 − 𝑋𝑖𝑗(𝑡)) + 𝑟2 ∗ (𝐺𝑏𝑒𝑠𝑡𝑗 − 𝑋𝑖𝑗(𝑡)) (5)
𝑋𝑖𝑗(𝑡 + 1) = 𝑋𝑖𝑗(𝑡) + 𝑣𝑖𝑗(𝑡 + 1) (6)
where 𝑣𝑖𝑗 and 𝑋𝑖𝑗 are the initial speeds and positions for the ith
particle and the jth
dimension, respectively,
and r1 and r2 are randomly chosen values between 0 and 1. 𝑃𝑏𝑒𝑠𝑡 stands for the individual agent’s best
solution, while 𝐺𝑏𝑒𝑠𝑡 stands for the best solution received from the best agent globally [21].
2.3.3. Chi-square algorithm
A feature selection technique uses a numerical test to evaluate divergence from the expected
distribution without considering any association between the occurrence of the feature and the class [22].
Additionally, it draws out the characteristics that each class values the most from the population.
2.3.4. Genetic algorithm
It is a meta-heuristic algorithm that optimizes issues and provides valuable solutions. Furthermore, it
is sensible and can identify the best solution by changing the evolution of population chromosomes [23]. It
contains four techniques, inheritance, mutation, crossover, and selection using fitness.
Figure 4. The structure of the proposed CNN model DGNet

 ISSN: 2088-8708
7010
2.4. Image recognition
Finally, the gait images have been recognized using the proposed DNN, which contains input, two
hidden, and output layers. The input layer contains the selected features, the two hidden layers have 512 and
124 dense, respectively, and the output layer has the image label. The training process of any DNN passes
through two phases: the feed forward (FF) phase and the back propagation (BP) phase. Each neuron in the FF
phase has been unsupervised training using the training part of data, and the other inputs have been used as
the output for the last neuron.
In the backpropagation phase, the back propagation (BP) neural network parameters have been
utilized to compute the error value [24]. The suggested framework of the proposed work is illustrated in
algorithm 1. Moreover, the proposed strategy complexity analysis is presented in Table 1 based on algorithm.
Using feature population number symbolled by n and iterations number Mt.
Algorithm 1. The proposed methodology
Input: RGB gait dataset and background images.
Output: Identified gait Images
Start procedure
Import Model libraries
Read RGB gait dataset (G) and background images (Bg).
Preprocess image (resizing, gray scale image, contrast)
Binarize I then apply filtering process with some morphological operation (dilating and
filling holes)
Get the Isolated Gait dataset (ROI).
Normalize all isolated images (I).
Split images (I) to 70% train, 15% test, and 15% valid.
− Training Set: train_X;
− Label of Training Data: train_ y, M=Model;
Training and testing Algorithms
for M=[AlexNet, VGG, GoogleNet, ResNet50, Xception]
Extract gait features Xi (population)
for N=[PSO, GWO, Chi-square, Genetic]
Evaluate fitness Fn for each Xi
Find the best features 𝑋 ∗
⃗⃗⃗⃗⃗⃗
Train DNN network layers using Xi features.
Summary the model
Compile Model
where {loss=Li /Opt=yi/Metrics=Acc.}
Fit Model
Where {epochs=30/Batch size=32/Verb.=2}
Evaluate the model
Plot accuracy (Acc.)/Loss curves
Plot Conf. Matrix
Calculate precision/recall/F1-score
Apply statistical analysis ANOVA/Wilcoxon Test
Store the Model
end
end
End procedure
Table 1. The proposed model computational complexity
No. Operation Complexity
1 Reading silhouettes images O (1)
2 Initializing population features O (1)
3 Evaluating fitness function Fn O (n)
4 Getting best features X ∗
⃗⃗⃗⃗⃗ O (n)
5 Updating selector parameters O (Mt×n)
6
Evaluating agents’ fitness function using PSO, GWO,
Chi-square, Genetic O (Mt×n)
7 Evaluating the best selector O (Mt×n)
8 Updating parameters O (Mt)
9 Producing the best fitness O (1)
10 Evaluating the model O (1)
3. PERFORMANCE METRICS
There are several performance metrics evaluated from the proposed model confusion matrix
parameters, including true positive value (TP), true negative value (TN), false positive value (FP), and false-
negative value (FN). These metrics, such as accuracy, sensitivity, specificity, F-score, training time, and

7011
intersection over union (IoU), verify the model’s robustness in both the training and testing phase. The
mentioned metrics have been computed from the (7) to (11) [25],
Accuracy =
TP+TN
TP+TN+FP+FN
× 100 (7)
TPR =
TP
TP+FN
× 100 (8)
TNR =
TN
TN+Fp
× 100 (9)
F − score =
2TP
2TP+FP+FN
× 100 (10)
IoU =
TP
TP+FP+FN
× 100 (11)
4. RESULTS AND DISCUSSION
The proposed DNN has been compiled and fitted based on the number of tuning parameters, such as
batch normalization with size 64, the learning rate value with step size 0.01, and the momentum factor 0.5,
which enhance the accuracy value and the training speed. Furthermore, Ata et al. [26] was utilized as the
optimizer, which used a more powerful stochastic optimization approach to consume memory than others,
with low computational power and the sparse categorical cross-entropy as a loss function for gathering all
used datasets, which have multiple outputs. Table 2 refers to the selected tuning hyper-parameters in our
proposed study, implemented with the Tensor Flow framework, an open-source deep learning library by
Kaggle. All proposed experiments have been evaluated using a personal computer (PC) with the following
characteristics: Microsoft Windows 10 operating system, processor 7-core @4.0 GHz, RAM size 12 GB, and
GPU NVidia Tesla 16 GB.
Table 2. The tuning hyper-parameter
Variable Power (kW)
Batch num 64
Learning Rate 1×10−2
Epochs 30
Optimizer Adam
Momentum Factor 0.5
Loss Function Sparse categorical
4.1. CASIA dataset analysis
From the optimal hyperparameters in Table 2, the proposed structural technique was estimated by
the CASIA gait dataset, which includes various view angle degrees for a class label. Performance metrics
results have been extracted from the confusion matrix for different mentioned experiments that proved the
proposed CNN with genetic feature selectors scored the best values among the other pre-trained models. The
parameters of the feature selectors models are tabulated in Table 3, and the sample of the confusion matrix
for DGNet with different feature selectors is shown in Figures 5(a) to 5(d).
Table 3. The main parameters of feature selectors models
Model Parameter (s) Value (s)
General Number of Iterations 20
Search domain [0,1]
PSO Inertia weight, W 0.9
Acceleration factors C1, C2 [2,2]
K-neighbors 5
GWO Number of agents 3
a 2 to 0
r [0,1]
GENETIC Mutation ratio 0.1
Crossover 0.9
Selection Mechanism Roulette wheel

 ISSN: 2088-8708
7012
(a) (b)
(c) (d)
Figure 5. The confusion matrix of DGNet with genetic feature selector (a) PSO, (b) GWO, (c) chi-square,
and (d) genetic algorithm
The performance metrics of the DGNet with different feature selectors have been tabulated in
Table 4, which scored an accuracy value of 99.46% using genetic feature selectors in 20 seconds only.
Figures 6(a) and 6(b) shows the accuracy and loss curves of DGNet with the genetic selector model based on
12 runs and 20 iterations (maximum iterations) for the n feature from the DGNet model (population size).
Moreover, Table 5 presents the different feature selector test results using a one-way ANOVA test [27]. In
contrast, Table 6 discusses the different applied feature selectors model comparison results using the
Wilcoxon Signed-Rank test. From the table results, the p-value scored less than 0.05, demonstrating the
significant difference between the feature selector results [28].
Table 4. Performance metrics of the DGNet with different feature selectors for CASIA gait dataset
Model Accuracy (%) SENS (%) IOU (%) SPEC (%) FNR (%) FPR (%) TIME
PSO 96.2 94.38 89.36 97.19 5.61 2.8 20.4s
GWO 97.2 95.7 91.84 97.8 4.25 2.12 23.5s
Chi-square 97.6 96.4 93.1 98.21 3.57 1.78 16s
Genetic 99.46 99.02 98.06 99.5 0.97 0.48 19.7s
4.2. OU-ISIR dataset analysis
Moreover, the DGNet and pre-trained models have been tested using the OU-ISIR gait dataset
according to the same optimal tunning parameters with different feature selectors. The performance metrics
of the DGNet with different feature selectors have been tabulated in Table 7. The mentioned results proved
that the proposed model with the Genetic feature selector scored the best in low training time.

7013
(a) (b)
Figure 6. The DGNet with Genetic feature selector (a) accuracy and (b) loss curve
Table 5. ANOVA test for different feature selection algorithms on DGNet
SS DF MS F (DFn, DFd) P value
Treatment (between columns) 0.008824 3 0.002941 F (3, 44) =13.31 P<0.0001
Residual (within columns) 0.009725 44 0.0002210 _ _
Total 0.01855 47 _ _ _
Table 6. Wilcoxon test on DGNet for various feature selection techniques
PSO GWO Chi-Square Genetic
Theoretical median 0 0 0 0
Actual median 0.96 0.9725 0.971 0.99
Number of values 12 12 12 12
Wilcoxon signed rank test
Sum of signed ranks (W) 78 78 78 78
Sum of positive ranks 78 78 78 78
Sum of negative ranks 0 0 0 0
P value (two tailed) 0.0005 0.0005 0.0005 0.0005
Exact or estimate? Exact Exact Exact Exact
P value summary *** *** *** ***
Significant (alpha=0.05)? Yes Yes Yes Yes
How big is the discrepancy?
Discrepancy 0.96 0.9725 0.971 0.99
Table 7. Performance metrics of the DGNet with different feature selectors for OU-ISIR gait dataset
Model Accuracy (%) SENS (%) IOU (%) SPEC (%) FNR (%) FPR (%) TIME
PSO 97.8 90.3 82.3 98.7 9.65 1.206 9 m 5 s
GWO 98.1 94.02 88.7 99.2 5.9 0.74 15 m 22 s
Chi-square 97.9 90.5 82.7 98.8 9.4 1.17 4 m 10 s
Genetic 99.09 95.9 92.2 99.4 4.05 0.506 6 m 24 s
4.3. Comparative study
The DGNet and pre-trained models have been tested using the same optimum tuning hyper
parameters listed in Table 8. Moreover, the proposed technique has been compared with the recent
state-of-art studies and tabulated in Table 9, which approved that the proposed technique outperforms the
current studies despite its deep structure and using many models. Still, it enhanced the accuracy value by
1.46% than [4], 0.67% than [5], 4.24% than [10], and 8.07% than [1].
Table 8. Accuracy values of the pre-trained models with different feature selectors
Model DGNet AlexNet GoogleNet VGGNet ResNet50 Xception
PSO 97.8 94.3 96.2 98.1 98.08 93.2
GWO 98.1 93.5 94.9 97.3 97.4 96.2
Chi-square 97.9 94.3 96.9 97.6 96.8 94.5
Genetic 99.09 97.9 97.8 98.5 98.1 98.6

 ISSN: 2088-8708
7014
Table 9. Comparator study with the current studies
Reference Year Accuracy
George et al. [4] 2020 98%
Premalatha and Chandramani [5] 2020 98.79%
Wen and Wang [1] 2021 91.39%
Zhang et al. [10] 2022 95.22%
Han et al. [3] 2022 87.5%
Dong et al. [9] 2023 97.7%
Ours 2023 99.46%
5. CONCLUSION
This paper proposes a methodology for gait authentication successive frames with a well-designed
approach. Firstly, the gait frames have been extracted, enhanced, and normalized. Secondly, the gait features
have been extracted using the proposed CNN model called DGNet and other pre-trained models. Then the
best was selected using genetic, PSO, chi-square, and GWO models. Finally, the proposed DNN model
identifies the gait images using the selected features. According to performance metrics results, the DGNet
with the Genetic feature selector outperforms the other pre-trained models and the current studies, which
scored 99.46% and 99.09% in recognition accuracy for CASIA and OU-ISIR datasets, respectively. The
proposed structure will likely be tested on various gait datasets in the future to verify its robustness, and
several models, such as RNN, graph CNN, and generative adversarial network (GAN), will likely be used.
REFERENCES
[1] J. Wen and X. Wang, “Gait recognition based on sparse linear subspace,” IET Image Processing, vol. 15, no. 12, pp. 2761–2769,
Oct. 2021, doi: 10.1049/ipr2.12260.
[2] H. Li et al., “GaitSlice: A gait recognition model based on spatio-temporal slice features,” Pattern Recognition, vol. 124, Apr.
2022, doi: 10.1016/j.patcog.2021.108453.
[3] F. Han, X. Li, J. Zhao, and F. Shen, “A unified perspective of classification-based loss and distance-based loss for cross-view gait
recognition,” Pattern Recognition, vol. 125, May 2022, doi: 10.1016/j.patcog.2021.108519.
[4] M. L. George, T. Govindarajan, K. A. Rajasekaran, and S. R. Bandi, “A robust similarity based deep Siamese convolutional
neural network for gait recognition across views,” Computational Intelligence, vol. 36, no. 3, pp. 1290–1319, Aug. 2020, doi:
10.1111/coin.12361.
[5] G. Premalatha and P. V Chandramani, “Improved gait recognition through gait energy image partitioning,” Computational
Intelligence, vol. 36, no. 3, pp. 1261–1274, Aug. 2020, doi: 10.1111/coin.12340.
[6] A. Sokolova and A. Konushin, “Pose‐based deep gait recognition,” IET Biometrics, vol. 8, no. 2, pp. 134–143, Mar. 2019, doi:
10.1049/iet-bmt.2018.5046.
[7] A. Parashar, A. Parashar, and R. S. Shekhawat, “A robust covariate‐invariant gait recognition based on pose features,” IET
Biometrics, vol. 11, no. 6, pp. 601–613, Nov. 2022, doi: 10.1049/bme2.12103.
[8] J. Chen, Z. Wang, C. Zheng, K. Zeng, Q. Zou, and L. Cui, “GaitAMR: Cross-view gait recognition via aggregated multi-feature
representation,” Information Sciences, vol. 636, Jul. 2023, doi: 10.1016/j.ins.2023.03.145.
[9] D. Dong, C. Ma, M. Wang, H. T. Vu, B. Vanderborght, and Y. Sun, “A low-cost framework for the recognition of human motion
gait phases and patterns based on multi-source perception fusion,” Engineering Applications of Artificial Intelligence, vol. 120,
Apr. 2023, doi: 10.1016/j.engappai.2023.105886.
[10] Z. Zhang, Z. Wang, H. Lei, and W. Gu, “Gait phase recognition of lower limb exoskeleton system based on the integrated
network model,” Biomedical Signal Processing and Control, vol. 76, Jul. 2022, doi: 10.1016/j.bspc.2022.103693.
[11] R. Altilio, A. Rossetti, Q. Fang, X. Gu, and M. Panella, “A comparison of machine learning classifiers for smartphone-based gait
analysis,” Medical and Biological Engineering and Computing, vol. 59, no. 3, pp. 535–546, Mar. 2021, doi: 10.1007/s11517-020-
02295-6.
[12] M. Asif, M. I. Tiwana, U. S. Khan, M. W. Ahmad, W. S. Qureshi, and J. Iqbal, “Human gait recognition subject to different
covariate factors in a multi-view environment,” Results in Engineering, vol. 15, Sep. 2022, doi: 10.1016/j.rineng.2022.100556.
[13] A. Chauhan, “GaitSilhouetteDataset,” the kaggle.com, 2019. https://www.kaggle.com/datasets/watermasterz/gaitsilhouettedataset
(accessed Mar. 01, 2023).
[14] “OU-ISIR biometric database,” The Institute of Scientific and Industrial Research. http://www.am.sanken.osaka-
u.ac.jp/BiometricDB/GaitLPAge.html (accessed Mar. 01, 2023).
[15] Y. Xie, L. Ning, M. Wang, and C. Li, “Image enhancement based on histogram equalization,” Journal of Physics: Conference
Series, vol. 1314, no. 1, Oct. 2019, doi: 10.1088/1742-6596/1314/1/012161.
[16] L. Qingge, R. Zheng, X. Zhao, W. Song, and P. Yang, “An improved Otsu threshold segmentation algorithm,” International
Journal of Computational Science and Engineering, vol. 22, no. 1, 2020, doi: 10.1504/IJCSE.2020.10029225.
[17] Y. Donon, R. Paringer, and A. Kupriyanov, “Image normalization for blurred image matching,” CEUR workshop proceedings,
pp. 127–131, 2020.
[18] Ambika, Virupakshappa, and S.-J. Lim, “Hybrid image embedding technique using steganographic signcryption and IWT-GWO
methods,” Microprocessors and Microsystems, vol. 95, Nov. 2022, doi: 10.1016/j.micpro.2022.104688.
[19] N. Kunhare, R. Tiwari, and J. Dhar, “Particle swarm optimization and feature selection for intrusion detection system,” Sādhanā,
vol. 45, no. 1, Dec. 2020, doi: 10.1007/s12046-020-1308-5.
[20] R. Pramanik, S. Sarkar, and R. Sarkar, “An adaptive and altruistic PSO-based deep feature selection method for Pneumonia
detection from Chest X-rays,” Applied Soft Computing, vol. 128, Oct. 2022, doi: 10.1016/j.asoc.2022.109464.
[21] A. Singh, A. Sharma, S. Rajput, A. Bose, and X. Hu, “An investigation on hybrid particle swarm optimization algorithms for
parameter optimization of PV cells,” Electronics, vol. 11, no. 6, Mar. 2022, doi: 10.3390/electronics11060909.

7015
[22] I. S. Thaseen and C. A. Kumar, “Intrusion detection model using fusion of chi-square feature selection and multi class SVM,”
Journal of King Saud University-Computer and Information Sciences, vol. 29, no. 4, pp. 462–472, Oct. 2017, doi:
10.1016/j.jksuci.2015.12.004.
[23] N. Kunhare, R. Tiwari, and J. Dhar, “Intrusion detection system using hybrid classifiers with meta-heuristic algorithms for the
optimization and feature selection by genetic algorithm,” Computers and Electrical Engineering, vol. 103, Oct. 2022, doi:
10.1016/j.compeleceng.2022.108383.
[24] L. Sun, B. Zou, S. Fu, J. Chen, and F. Wang, “Speech emotion recognition based on DNN-decision tree SVM model,” Speech
Communication, vol. 115, pp. 29–37, Dec. 2019, doi: 10.1016/j.specom.2019.10.004.
[25] R. N. Yousef, A. T. Khalil, A. S. Samra, and M. M. Ata, “Model-based and model-free deep features fusion for high performed
human gait recognition,” The Journal of Supercomputing, vol. 79, no. 12, pp. 12815–12852, Aug. 2023, doi: 10.1007/s11227-
023-05156-9.
[26] M. M. Ata, M. L. Francies, and M. A. Mohamed, “A robust optimized convolutional neural network model for human activity
recognition using sensing devices,” Concurrency and Computation: Practice and Experience, vol. 34, no. 17, Aug. 2022, doi:
10.1002/cpe.6964.
[27] T. Si, D. K. Patra, S. Mondal, and P. Mukherjee, “Breast DCE-MRI segmentation for lesion detection using chimp optimization
algorithm,” Expert Systems with Applications, vol. 204, Oct. 2022, doi: 10.1016/j.eswa.2022.117481.
[28] A. Mishrif and A. I. Khan, “Shipping and transportation traffic of medical and non-medical goods before and during COVID-19
in Oman,” Procedia Computer Science, vol. 201, pp. 223–230, 2022, doi: 10.1016/j.procs.2022.03.031.
BIOGRAPHIES OF AUTHORS
Reem Nehad Yousef is an Assistant lecturer at Delta Higher Institute for
Engineering and Technology, Mansoura, Egypt. She received the M.Sc. degree in
communications engineering from Mansoura University in 2019. She is currently a Ph.D.
student in the Faculty of Engineering, Mansoura University. Her research interests are in
image processing, communication systems, and both machine and deep learning
methodologies. She can be contacted at email: Reem.nehad@hotmail.com.
Abeer Tawkool Khalil received the B.Sc. and Ph.D. degrees from the Electronics
and Communications Engineering Department, Faculty of Engineering, Mansoura University,
Mansoura, Egypt, in 2001 and 2013, respectively. She is currently working as an Associate
Professor at the Faculty of Engineering, Mansoura University. She has published more than 27
articles and supervised ten postgraduate students in many universities. She is interested in
wireless networking and hardware realizations of digital systems. She can be contacted at
email: Abeer.twakol@mans.edu.eg.
Ahmed Shaaban Samra was born in Mansoura, Egypt 1954. He received the
B.Sc. and the M.Sc. degree in communications engineering from Menoufia University in 1977,
1982 respectively, and the Ph.D. degree in optical communications and integrated optics from
ENSEG, Greroble, France in 1988. He is now an associate professor at the faculty of
engineering, Mansoura University. His research interests are in the field of optical
communications and optical measurement techniques. He can be contacted at email:
ahsamra@yahoo.co.uk.
Mohamed Maher Ata is an assistant professor, School of Computational
Sciences and Artificial Intelligence (CSAI), Zewail City of Science and Technology. He is also
an assistant professor at MISR Higher Institute for Engineering and Technology, Mansoura,
Egypt. He has received his Ph.D. from the Faculty of Engineering, Electrical Communication
and Electronics Department, Tanta University, Egypt with the cooperation of Regina
University in Canada. His research area of interest was utilized in the field of signal
processing, image processing, multimedia, machine and deep learning, video processing, and
computer vision. He has published many indexed research articles (SJR indexed-ISI indexed)
in the state-of-the-art of computer vision, biomedical engineering, astrophysics, electrical
communication, bioinformatics, encryption, cyphering and intelligent transportation systems
(ITS). He can be contacted at email: mmaher844@yahoo.com.

Accurate metaheuristic deep convolutional structure for a robust human gait recognition

More Related Content

Similar to Accurate metaheuristic deep convolutional structure for a robust human gait recognition (20)

More from IJECEIAES (20)

Recently uploaded (20)

Accurate metaheuristic deep convolutional structure for a robust human gait recognition