Anomaly Detection in Pedestrian Walkways for Intelligent Transportation System Using Federated Learning and Harris Hawks Optimizer on Remote Sensing Images

Abstract

Anomaly detection in pedestrian walkways is a vital research area that uses remote sensing, which helps to optimize pedestrian traffic and enhance flow to improve pedestrian safety in intelligent transportation systems (ITS). Engineers and researchers can formulate more potential techniques and tools with the power of computer vision (CV) and machine learning (ML) for mitigating potential safety hazards and identifying anomalies (i.e., vehicles) in pedestrian walkways. The real-world challenges of scenes and dynamics of environmental complexity cannot be handled by the conventional offline learning-based vehicle detection method and shallow approach. With recent advances in deep learning (DL) and ML areas, authors have found that the image detection issue ought to be devised as a two-class classification problem. Therefore, this study presents an Anomaly Detection in Pedestrian Walkways for Intelligent Transportation Systems using Federated Learning and Harris Hawks Optimizer (ADPW-FLHHO) algorithm on remote sensing images. The presented ADPW-FLHHO technique focuses on the identification and classification of anomalies, i.e., vehicles in the pedestrian walkways. To accomplish this, the ADPW-FLHHO technique uses the HybridNet model for feature vector generation. In addition, the HHO approach is implemented for the optimal hyperparameter tuning process. For anomaly detection, the ADPW-FLHHO technique uses a multi deep belief network (MDBN) model. The experimental results illustrated the promising performance of the ADPW-FLHHO technique over existing models with a maximum AUC score of 99.36%, 99.19%, and 98.90% on the University of California San Diego (UCSD) Ped1, UCSD Ped2, and avenue datasets, respectively. Therefore, the proposed model can be employed for accurate and automated anomaly detection in the ITS environment.

1. Introduction

Currently, the easy-to-use unmanned remote sensing vehicles (UAVs) and the availability of low-cost image acquisition systems have made remote sensing imaging (RSI) more popular and convenient [1]. It has become possible to obtain a wide range of high-quality RSI without a considerable amount of time and elaborate planning. The RSI has been adopted in several tasks, namely precision agriculture, cartography, urban studies, and landscape archaeology for over a decade [2]. One effective implementation is to classify and detect the vehicle in RSI. It is progressively adopted to smart transportation for traffic flow estimation, parking space allocation, vehicle identification, and so on. Hence, it is the future trend to use RSI for vehicle and transportation-related applications [3]. Over the past years, vehicle identification in RSI has been extensively used in several fields and thus attracted considerable interest. Despite the serious amount of effort dedicated to these tasks, the current method still requires significant expansion to overcome these problems [4]. Firstly, direction and scale variability makes it highly challenging to precisely locate the vehicle object. Next, complex backgrounds rise interclass similarity and intraclass variability. At last, some RSI is captured in a lower resolution, which can lead to lacking an abundant detailed appearance for distinguishing vehicles from similar objects. As compared with each image of the vehicle [5], the RSI captured from the perpendicular perspective lose the ‘face’ of the vehicle, and the vehicle commonly shows a rectilinear structure. Therefore, the presence of non-vehicle rectilinear objects such as electrical units and trash bins can complicate the task of air conditioning units on the top of buildings, which causes several false alarms [6]. Thus, researchers are trying to employ contemporary DL-based object detection methods to push the boundary of triumph in these regards.

With the latest improvements in machine learning (ML) approaches, we are now capable of achieving higher object detection rates in cluttered scenes [7]. ML is a subfield of artificial intelligence (AI) that focuses on the design of algorithms and statistical models, which allows computers to learn and make predictions without being explicitly programmed. The detection and classification of vehicles using aerial images become more realistic with deep neural networks (DNN). The technique for vehicle detection using aerial images is categorized into two classes, the traditional ML approaches and the DL methods [8]. DL for computer vision (CV) grows more conventional every year, particularly due to Convolutional Neural Networks (CNN) that are capable of learning expressive and powerful descriptors from an image for a great number of tasks: detection, classification, segmentation, and so on [9]. This ubiquity of CNN in CV is now beginning to affect remote sensing, as they could address several tasks such as object detection or land use classification in RSI. Furthermore, a new architecture has appeared, derived from Fully Convolutional Networks (FCN) [5], capable of outputting dense pixel-wise annotation and therefore capable of achieving fine-grained classification. Federated learning (FL) is an ML approach that trains the model using multiple independent sessions. In contrast, the classical centralized ML model, where local datasets are combined into one training session, considers that local data samples can be identically distributed. FL allows many actors for building a common, robust ML approach with no data distribution, thus resolving challenging problems such as data privacy, data security, data access rights, and access to heterogeneous data.

Javadi et al. [10] examined the capability of 3D mapping features for improving the efficiency of DNN for vehicle detection. Charouh et al. [11] presented a structure for reducing the complexity of CNN-oriented AVS approaches, whereas a BS-related element was established as a pre-processing stage for optimizing the count of convolutional functions implemented by the CNN component. A CNN-based detector with a suitable count of convolutions was executed for all the image candidates for handling the overlapping issue and enhancing detection efficiency. In [12], emergency vehicle detection (EVD) for police cars, fire brigades, and ambulances is performed dependent upon its horn sounds. A database in Google Audioset ontology has gathered and extracted features by Mel-frequency Cepstral Coefficient (MFCC). The three DNN approaches such as CNN, dense layer, and RNN, with distinct configurations and parameters, are examined. Afterwards, an ensemble approach was planned with better-chosen methods and executing experimental testing on several configurations with hyperparameter tuning. The authors in [13] examined a model for vehicle detection in multi-modal aerial imagery through an improved YOLOv3 and DNN, which behaviours mid-level fusion. The presented mid-level fusion structure is a primary of its type that is utilized for detecting vehicles in multi-modal aerial imagery using a hierarchical object detection network. Joshi et al. [14] presented the EDL-MMLCC method (ensemble of DL-related multimodal land cover classification) utilizing remote sensing images (RSI). In addition, the trained procedure of DL approaches is improved by utilising the hosted cuckoo optimization (HCO) technique. Lastly, the SSA with regularized ELM (RELM) technique was executed for the land cover classifier.

Li et al. [15] presented an anchor-free target detection approach for solving these challenges. Primarily, a multi-attention feature pyramid network (MA-FPN) has been planned for addressing the control of noise and background data on vehicle target recognition by fusing attention data from the feature pyramid network (FPN) infrastructure. Secondly, a more precise foveal area (MPFA) has been presented for providing the best ground truth for the anchor-free process to determine a further precise positive instance selective region. In [16], the authors established a new vehicle detection and classification method for smart traffic observation which utilizes CNN for segmenting UAVs. These segmentation images can be examined to identify the vehicles with incorporated innovative customized pyramid pooling. Lastly, such vehicles can be tracked using Kalman filter (KF), kernelized filter-based approaches for coping with and accomplishing huge traffic flows with lesser human intervention.

Several models exist in the literature to perform the classification process. Though several ML and DL models for anomaly classification are available in the literature, it is still needed to enhance the classification performance. Owing to the continual deepening of the model, the number of parameters of DL models also increases quickly, which results in model overfitting. At the same time, different hyperparameters have a significant impact on the efficiency of the CNN model. Particularly, hyperparameters such as epoch count, batch size, and learning rate selection are essential to attain an effectual outcome. Since the trial and error method for hyperparameter tuning is a tedious and erroneous process, metaheuristic algorithms can be applied. Therefore, in this work, we employ an HHO algorithm for the parameter selection of the HybridNet model.

This study presents an Anomaly Detection in Pedestrian Walkways for Intelligent Transportation Systems using Federated Learning and Harris Hawks Optimizer (ADPW-FLHHO) algorithm on Remote Sensing Images. The ADPW-FLHHO technique uses the HybridNet model for feature vector generation. In addition, the HHO algorithm is implemented for the optimal hyperparameter tuning process. For anomaly detection, the ADPW-FLHHO technique uses a multi deep belief network (MDBN) model. The experimental results of the ADPW-FLHHO technique can be studied on the UCSD anomaly detection dataset.

The rest of the paper is organized as follows. Section 2 provides the related works and Section 3 offers the proposed model. Then, Section 4 gives the result analysis and Section 5 concludes the paper.

2. Materials and Methods

In this study, a novel ADPW-FLHHO approach was developed for anomaly detection in the pedestrian walkways, to improve the safety of the pedestrians. The presented ADPW-FLHHO technique mainly concentrated on the effectual identification and classification of anomalies, i.e., vehicles in the pedestrian walkways. Figure 1 demonstrates the overall flow of the ADPW-FLHHO algorithm.

2.1. Data Used

The performance of the ADPW-FLHHO technique can be examined under three datasets, namely UCSDPed1 (http://www.svcl.ucsd.edu/projects/anomaly/dataset.htm, accessed on 24 February 2023), UCSDPed2, and avenue dataset (http://www.cse.cuhk.edu.hk/leojia/projects/detectabnormal/dataset.html, accessed on 24 February 2023).

Table 1. Details of datasets.

Dataset	No. of Videos	Training Set	Testing Set	Average Frames	Dataset Length
UCSDPed1 (Bikers, small carts, walking across walkways)	70	34	36	201	5 min
UCSDPed2 (Bikers, small carts, walking across walkways	28	16	12	163	5 min
Avenue (Run, throw, new object)	37	16	21	839	5 min

illustrates the details of the datasets.

2.2. FL Model

Federated learning (FL) is classified into federated transfer learning (FTL), sample-based (horizontal) FL, and feature-based (vertical) FL based on the data distribution amongst different parties [17]. In sample-based or horizontal FL, datasets share a similar feature space but dissimilar sample space (data samples). In feature-based or vertical FL, the dataset shares a similar sample space but varies in feature space. In FTL, the dataset differs in both sample and feature space, which have a smaller intersection. In the presented method, consider a horizontal FL case. A bottom one (decentralized) and a top one (centralized) are the two standard designs of a federated architecture. For centralized design: initially, every partner within the data sharing federation collects local datasets from its data acquisition source (for example, surveillance cameras or vehicle sensors). Next, the raw information is exploited for a local anomaly recognition model that is framed and well-trained with the locally accessible dataset. Then, the local model is periodically synchronized with the cloud server. The failure would disrupt the training method of every client. On the other hand, the bottom of the FL structure does not depend on a central cloud instance, however, it allows to synchronize and direct exchange method parameters between partners. However, both methods are potential and the selected technique may rely on the partner included in a data-sharing federation.

2.3. Feature Extraction Using HybridNet

To derive a set of feature vectors, the Hybrid Net technique is used. For the generation of the visual feature of the input image, the HybridNet model is exploited in the presented study [18]. Overall, classification needs intra-class in-variant features while re-establishment should retain each dataset. HybridNet consists of the unsupervised path ($E_u$ and $D_u$) and a discriminating path ($E_c$ and $D_c$) to overcome the limitation. Both $E_u$ and $E_c$ encoding functions take an $x$ input image and generate $h_c$ and $h_u$ representations, however, the $D_c$ and $D_u$ decoding function correspondingly assume $h_c$ and $h_u$ as an input to create $\widehat{x}$ and $\widehat{x}$ partial reconstruction. Finally, $C$ classification produces a class prediction through discriminating features: $\widehat{y}=C\left(h_c\right)$. While both paths might have the same structure, they have to do complementary and various roles. The discriminating path should extract the $h_c$ discriminating attribute that needs to be crafted well to perform the classifier task and make an ${\widehat{x}}_c$ partial reconstruction that could be precise while maintaining the data. So, the unsupervised path role is corresponding to the discriminating path through $p$ in $h_u$ the data lost in $h_c$. Accordingly, it generates $\widehat{x}$ complementary reconstruction, while adding $\widehat{x}$ and $\widehat{x}$, reconstruction $\widehat{x}$ is close to $x$, and it can be formulated as follows:

$$h_c=E_c\left(x\right){\widehat{x}}_c=D_c\left(h_c\right)\widehat{y}=C\left(h_c\right)$$

$$h_u=E_u\left(x\right){\widehat{x}}_u=D_u\left(h_u\right)\widehat{x}={\widehat{x}}_c+{\widehat{x}}_u$$

The ultimate purpose is to perform the process as regularised for discriminating encoding. The key problem and contribution are to develop an approach for ensuring that both paths are performed.

The two most important challenges that are addressed are in the discriminating path to emphasise discriminating features, and that required both paths to contribute and co-operate toward the reconstruction. Two paths that work individually are a reconstruction path $\widehat{x}={\widehat{x}}_u=D\left(E\right(x\left)\right)$ and a classification path $\widehat{y}=C\left(E\left(x\right)\right)$ and ${\widehat{x}}_c=0$. Next, we resolved the problems by using the encoding and decoding function altogether with an accurate training and loss function. The HybridNet method comprises dual data paths in that one creates a class prediction and both create a partial reconstruction that needs to be integrated. In the presented algorithm, we overcome issues of training those structures efficiently. It involves an intermediate reconstruction with ${\mathcal{L}}_{rec-interb,l}$ (for layer $l$ and branch $b$)$;$ classification with ${\mathcal{L}}_{cls}$; stability with ${\Omega }_{stability}$; and reconstruction with ${\mathcal{L}}_{rec}$.

All of these terms are weighted through the parameter $\lambda$ and it follows the branch complementarity training model.

$$\mathcal{L}={\lambda }_c{\mathcal{L}}_{cls}+{\lambda }_r{\mathcal{L}}_{rec}+\sum _{bϵ\left\{c,u\right\},l}{\lambda }_{rb,l}{\mathcal{L}}_{rec-interb,l}+{\lambda }_s{\Omega }_{stability}$$

(1)

Every batch was incorporated of $n$ samples, categorized into labelled images $n_s$ from ${\mathcal{D}}_{\mathrm{s}\mathrm{u}\mathrm{p}}$ and unlabelled images $n_u$ from ${\mathcal{D}}_{un\mathrm{s}\mathrm{u}\mathrm{p}}$. HybridNet was trained on the partially labelled dataset, that is labelled pairs ${\mathcal{D}}_{\mathrm{s}\mathrm{u}\mathrm{p}}=\{\left(x^{\left(k\right)},y^{\left(k\right)}\right){\}}_{k=1.N_s}$ and unlabelled images ${\mathcal{D}}_{un\mathrm{s}\mathrm{u}\mathrm{p}}=\{x^{\left(k\right)}{\}}_{k=1..N_u}$. The classification term was a regular cross-entropy term that is exploited on $n_s$ labelled instance as follows:

$${\mathcal{l}}_{cls}={\mathcal{l}}_{CE}(\widehat{y},y)=-\sum _i{y}_i\mathrm{l}\mathrm{o}\mathrm{g}\widehat{y},{\mathcal{L}}_{cls}=\frac{1}{n_s}\sum _k{\mathcal{l}}_{cls}({\widehat{y}}^{\left(k\right)},y^{\left(k\right)}).$$

(2)

2.4. Hyperparameter Tuning Using HHO Algorithm

In this work, the HHO technique is applied for the optimal hyperparameter tuning process. HHO was stimulated by the behaviour of Harris Hawks (HHs), which are intellectual birds [19]. Using different states of energy, the fundamental phases of HHO can be obtained. The exploration stage mimics the mechanism if HHs could not exactly locate prey. In that case, the hawk takes a break to locate and track the newest prey. The candidate solution was the hawks in the HHO technique, and the better solution in each step is prey. The hawks arbitrarily perch at the dissimilar location and wait for the prey applying 2 major operators, which are designated according to the $q$ probability, where $q<0.5$ specifies that hawks perch at the place of others and the prey (rabbit). The hawks are at an arbitrary position near the population range if $q\ge 0.5$. The following are symbols utilized in the algorithm to facilitate the understanding of HHO, is given below:

• The initial state of energy, escape energy E₀, E;
• Vector of hawks location (search agent) X_i;
• Location of Rabbit (better agent) X_rabbit;
• Location of a random Hawk X_rand;
• Hawks average location X_m;
• The maximal amount of iterations, swarm size, iteration counter T, N, t 6. Random integers lie within (0,1)r₁, r₂, r₃, r₄, r₅, q;
• Dimension, upper and lower boundaries of parameters, D, LB, UB.

The exploration stage is represented below:

$$X\left(t+1\right)=\left\{\begin{array}{ll}{X}_{rand}\left(t\right)-r_1\left|X_{rand}\left(t\right)-2r_{2}X\left(t\right)\right|& q\ge 0.5\\ \left(X_{rabbit}\left(t\right)-X_m\left(t\right)\right)-r_3\left(LB+r_4\left(UB-LB\right)\right)& q<0.5\end{array}\right.$$

(3)

The average location of the $X_m$ hawks can be shown as:

$$X_m\left(t\right)=\frac{1}{N}\sum _{i=1}^N{X}_i\left(t\right)$$

(4)

where $X\left(t\right)$ denotes the location in the iteration for $t$ Hawk and $N$ symbolizes the overall amount of hawks. The average location was effectuated by applying dissimilar techniques. An effective changeover from exploration to exploitation was needed; herein a shift was predictable among the exploitative behaviour by using escaping energy factor $E$ of prey, which drastically minimizes escaping behaviours. The prey energy is calculated using the following equation:

$$E=2E_0\left(1-\frac{t}{T}\right)$$

(5)

where $E_0$ denotes the initial escape energy and $T$ characterizes the greatest amount of iterations, correspondingly. The soft besiege was a crucial stage in HHO, it can be displayed if $r\ge 0.5$ and $\left|E\right|\ge 0.5$. In such scenarios, the rabbit has adequate energy. The rabbit performs a random misleading shift to escape, but still, in metaphor, it could not. These steps are determined by the following rules:

$$X\left(t+1\right)=\Delta X\left(t\right)-E\left|J{X}_{rabbit}\left(t\right)-X\left(t\right)\right|$$

(6)

$$\Delta X\left(t\right)=X_{rabbit}\left(t\right)-X\left(t\right)$$

(7)

where $\Delta X\left(t\right)$ denotes the different location vector for each rabbit and for a current position at $t$ iteration, and $J=2(1-r_5)$ is the spontaneous jumping ability of the rabbit all over the escaping stage. The J value varies arbitrarily in every iteration for representing rabbit behaviours. In the extreme siege phase, if $r\ge 0.5$ and $\left|E\right|<0.5$, then the prey was tired and had no escape energy. The HHs were barely encircling the trained prey and attacked them. For these cases, the existing location is changed by Equation (8):

$$X\left(t+1\right)=X_{rabbit}\left(t\right)-E\left|\Delta X\left(t\right)\right|$$

(8)

Assume the hawks’ behaviour, they gradually select the better dive for prey if they capture the prey in a competitive situation:

$$Y=X_{rabbit}\left(t\right)-E\left|J{X}_{rabbit}\left(t\right)-X\left(t\right)\right|$$

(9)

The soft besiege given in the prior Equation (9) was implemented in progressive speedy dives just if $\left|E\right|\ge 0.5$ but $r<0.5$. In such cases, the rabbit has enough energy for escaping and can be exploited for a soft siege beforehand so an attack comes as a surprise. The HHO model has diverse paradigms of escape for prey and leap frog movements. Here, the Lévy flights (LF) are developed for emulating the different movements of the rabbit dives and Hawk.

$$Z=Y+S\times LF\left(D\right)$$

(10)

where $S$ signifies the random vector of $1$ × $D$ size and LF indicates the Lévy flight function, utilizing Equation (11):

$$LF\left(x\right)=0.01\times \frac{u\times \sigma }{|\nu {|}^{\frac{1}{\beta }}},\sigma ={\left(\frac{\Gamma \left(1+\beta \right)\times \sin \left(\frac{\pi \beta }{2}\right)}{\Gamma \left(\frac{1+\beta }{2}\right)\times \beta \times 2\left(\frac{\beta -1}{2}\right)}\right)}^{\frac{1}{\beta }}$$

(11)

where, $u,$ and $v$ indicate the random value within $\left(\mathrm{0,1}\right),$ and $\beta$ symbolizes the default constant fixed to 1.5. The last stage was to upgrade the location of hawks using:

$$X\left(t+1\right)=\left\{\begin{array}{ll}y& ifF\left(Y\right)<F\left(X\left(t\right)\right)\\ Z& ifF\left(Z\right)<F\left(X\left(t\right)\right)\end{array}\right.$$

(12)

In Equation (12), $y$ and $Z$ are calculated using Equations (7) and (8). In progressive quick dives, HHO was hard-pressed, if $\left|E\right|<0.5$ and $r<0.5$. Now, the rabbit does not have enough strength to escape and the hard siege is recommended before the surprise attack is made to kill and catch prey. Here, hawks seek to drop the distance between the average position and prey:

$$X\left(t+1\right)=\left\{\begin{array}{ll}Y& ifF\left(Y\right)<F\left(X\left(t\right)\right)\\ Z& ifF\left(Z\right)<F\left(X\left(t\right)\right)\end{array}\right.$$

(13)

where $y$ and $Z$ values are proposed based on the new rules in Equations (14) and (15), where $X\left(t\right)$ can be attained.

$$Y=X_{rabbit}\left(t\right)-E\left|J{x}_{rabbit}\left(t\right)-X_m\left(t\right)\right|$$

(14)

$$Z=Y+S\times LF\left(D\right)$$

(15)

Fitness selection is a crucial factor in the HHO approach. Solution encoding is exploited for assessing the aptitude (goodness) of the candidate solution. Now, the accuracy value is the main condition utilized for designing a fitness function.

$$Fitness=\mathrm{m}\mathrm{a}\mathrm{x}\left(P\right)$$

(16)

$$P=\frac{TP}{TP+FP}$$

(17)

where TP represents the true positive and FP denotes the false positive value.

2.5. MDBN-Based Classification

For anomaly detection, the ADPW-FLHHO technique used the MDBN model. RBM comprises a visible layer (VL) and a hidden layer (HL). The VL has been responsible for input, and HL learns higher-level semantic features in the input dataset [20]. The visible and hidden units can be binary variables with a state of 0 or 1. An entire network refers to a bipartite graph, and there is no connecting edge in the VL/HL that occurs among the visible unit $v=[v_1,\cdots ,v_m,\cdots ,v_k]$ and hidden unit $h=[h_1,\cdots ,h_m,\cdots ,h_k].$ The parameter of RBM is altered to the suitable values for avoiding the local optimum solution. A strong-feature learning capability called RBM is utilized to extract the data. However, the efficiency of FE or RBM has been restricted once it can be executed on difficult non-linear data. For improving the representation capability of a single RBM, the DBN approach has been introduced by stacking several RBMs altogether. Therefore, the DBN search for a deep hierarchical representation of trained instances. The 2 neighbouring layers of DBN are assumed as a single RBM. The greedy layer-wise unsupervised learning trains all the RBMs, and RBM could not assume other RBM in their learning procedure. Figure 2 showcases the framework of the DBN technique.

For extracting deep features to all the classes in remote sensing images, a multi-DBN infrastructure is planned for the complete extraction of the data from all the classes. The first to $(L-1)$-th layer in MMDBN relates to multi-DBN infrastructure. Based on the HSI properties, Gaussian distribution has been established to model inputs that recognize real-valued RBM rather than binary RBM. The conditional probability distribution and energy function can be determined as:

$$E\left(v,h;\theta \right)=-\sum _{m=1}\frac{(v_m^i-b_m^i{)}^2}{2({\sigma }_m^i{)}^2}-\sum _{n=1}{a}_n^i{h}_n^i-\sum _{m=1}\sum _{n=1}{w}_{mn}^i\frac{v_m^i}{{\sigma }_m^i}{h}_n^i$$

(18)

$$P\left(h_n^i|v_i;\theta \right)$$

$$w_{m{n}^{\mathrm{\prime }}}^i\left({\sigma }_m^i{)}^2\right)$$

(19)

where ${\sigma }_i$ denotes standard deviation, and $N(\mu ,{\sigma }^2)$ implies Gaussian distribution with variance $\sigma$ and mean $\mu .$

The MDBN framework feature extraction in the perspective of DL and the feature attained in the $(L-1)$-th layer comprise deep abstract data for the hyperspectral dataset. For improving the discriminative ability of extraction features, MMDBN searches manifold infrastructure in HSI utilizing label data.

3. Results

Figure 3 exhibits the accuracy of the ADPW-FLHHO technique during the training and validation process on the UCSDPed1 dataset. The figure specifies that the ADPW-FLHHO technique reaches higher accuracy values over increasing epochs. As well, the increasing validation accuracy over training accuracy displays that the ADPW-FLHHO approach learns efficiently on the UCSDPed1 dataset.

The loss analysis of the ADPW-FLHHO technique at the time of training and validation is seen on the UCSDPed1 dataset in Figure 4. The results indicate that the ADPW-FLHHO technique reaches closer values of training and validation loss. The ADPW-FLHHO technique learns efficiently on the UCSDPed1 dataset.

Figure 5 examines the accuracy of the ADPW-FLHHO technique during the training and validation process on the UCSDPed2 dataset. The figure highlights that the ADPW-FLHHO technique reaches increasing accuracy values over increasing epochs. In addition, the increasing validation accuracy over training accuracy exhibits that the ADPW-FLHHO technique learns efficiently on the UCSDPed2 dataset.

The loss analysis of the ADPW-FLHHO technique at the time of training and validation on the UCSDPed2 dataset is shown in Figure 6. The figure points out that the ADPW-FLHHO technique reaches closer values of training and validation loss. The ADPW-FLHHO technique learns efficiently on the UCSDPed2 dataset.

Figure 7 inspects the accuracy of the ADPW-FLHHO method during the training and validation process on the Avenue dataset. The figure highlights that the ADPW-FLHHO technique reaches increasing accuracy values over increasing epochs. In addition, the increasing validation accuracy over training accuracy exhibits that the ADPW-FLHHO approach learns efficiently on the Avenue dataset.

The loss analysis of the ADPW-FLHHO technique at the time of training and validation is given on the Avenue dataset in Figure 8. The figure indicates that the ADPW-FLHHO method reaches closer values of training and validation loss. It is observed that the ADPW-FLHHO technique learns efficiently on the Avenue dataset.

4. Discussion

Table 2. TPR outcome of ADPW-FLHHO approach on UCSDPed1 dataset.

True Positive Rate
False Positive Rate	MPPCA	Social Force	MDT	AMDN	EADN	ADPW-FLHHO
0	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
5	0.0904	0.1302	0.2257	0.3239	0.3451	0.5944
10	0.2257	0.2443	0.3981	0.4459	0.5493	0.7430
15	0.3504	0.3663	0.5016	0.5653	0.7536	0.8226
20	0.4300	0.4565	0.5865	0.6979	0.7536	0.8730
25	0.9366	0.9207	0.7509	0.6820	0.5361	0.5175
30	0.5732	0.6316	0.7563	0.8464	0.9419	0.9605
35	0.6342	0.7165	0.7960	0.9048	0.9552	0.9711
40	0.6793	0.8093	0.8544	0.9260	0.9632	0.9764
45	0.7536	0.8809	0.8809	0.9393	0.9711	0.9817
50	0.7960	0.9075	0.9234	0.9605	0.9764	0.9844
55	0.8305	0.9313	0.9579	0.9711	0.9764	0.9844
60	0.8783	0.9393	0.9632	0.9764	0.9764	0.9870
65	0.9101	0.9472	0.9764	0.9791	0.9817	0.9870
70	0.9525	0.9552	0.9791	0.9791	0.9817	1.0000
75	0.9552	0.9738	0.9791	0.9817	0.9870	1.0000
80	0.9605	0.9817	0.9817	1.0000	1.0000	1.0000
85	0.9658	0.9817	1.0000	1.0000	1.0000	1.0000
90	0.9817	0.9817	1.0000	1.0000	1.0000	1.0000
95	0.9791	0.9923	1.0000	1.0000	1.0000	1.0000
100	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000

illustrates the TPR outcomes of the ADPW-FLHHO technique on the UCSDPed1 dataset. The figure indicates that the values of TPR are raised with an increase in FPR. It is noticed that the MPPCA, SF, and MDT models resulted in the least TPR values while the AMDN model accomplishes certainly boosted TPR values. Contrastingly, the EADN model accomplishes near-optimal performance with reasonable TPR values. Nevertheless, the ADPW-FLHHO technique reaches maximum performance with higher TPR values on the UCSDPed-1 dataset.

Table 3. $AU{C}_{score}$ outcome of ADPW-FLHHO approach with other systems on UCSDPed1 dataset.

Methods	AUC Score (%)
ADPW-FLHHO	99.36
EADN	98.36
Binary SVM classifier	96.73
MIL-C3D	94.99
TSN-Optical Flow	92.86
Spatiotemporal	91.57
TSN-RGB	90.49

highlights a comparative $AU{C}_{score}$ inspection of the ADPW-FLHHO method with the recent algorithm on the UCSDPed1 dataset [21]. The figure demonstrates the ineffectual outcomes of the TSN-Optical Flow, spatiotemporal, and TSN-RGB models with closer $AU{C}_{score}$ of 92.86%, 91.57%, and 90.49%, respectively. Contrastingly, the binary SVM and MIL-C3D models have managed to report a certainly increased $AU{C}_{score}$ of 96.73% and 94.99%, respectively. Although the EADN model accomplishes a considerable $AU{C}_{score}$ of 98.36%, the ADPW-FLHHO technique ensures its supremacy with a maximum $AU{C}_{score}$ of 99.36%.

Table 4. TPR outcome of ADPW-FLHHO approach on UCSDPed2 dataset.

True Positive Rate
False Positive Rate	MPPCA	Social Force	MDT	AMDN	EADN	ADPW-FLHHO
0	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
5	0.0704	0.1222	0.2632	0.3769	0.3424	0.5637
10	0.2460	0.2647	0.4221	0.4671	0.5295	0.6407
15	0.3599	0.4126	0.5546	0.5409	0.5969	0.7857
20	0.4845	0.4888	0.6375	0.6719	0.7419	0.8655
25	0.5544	0.6442	0.7570	0.6863	0.7763	0.9025
30	0.6876	0.7279	0.7926	0.8594	0.9215	0.9550
35	0.7134	0.8253	0.7945	0.9171	0.9419	0.9552
40	0.7696	0.8718	0.8289	0.9270	0.9482	0.9639
45	0.7931	0.9217	0.8713	0.9356	0.9544	0.9753
50	0.8266	0.9384	0.9295	0.9541	0.9608	0.9800
55	0.9214	0.9518	0.9631	0.9653	0.9759	0.9856
60	0.9334	0.9549	0.9714	0.9672	0.9795	1.0000
65	0.9460	0.9707	0.9995	0.9854	0.9883	1.0000
70	0.9585	0.9766	1.0000	0.9900	0.9941	1.0000
75	0.9775	0.9835	1.0000	0.9918	0.9987	1.0000
80	0.9832	0.9945	1.0000	0.9974	1.0000	1.0000
85	0.9884	0.9958	1.0000	1.0000	1.0000	1.0000
90	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000
95	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000
100	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000

reveals the TPR results of the ADPW-FLHHO technique on the UCSDPed2 dataset. The figure indicates that the values of TPR are raised with an increase in FPR. It is noticed that the MPPCA, SF, and MDT models resulted in the lowest TPR values while the AMDN model accomplishes certainly boosted TPR values. In contrast, the EADN model accomplishes near-optimal performance with reasonable TPR values. Nevertheless, the ADPW-FLHHO technique reaches maximum performance with maximum TPR values on the UCSDPed-1 dataset.

Table 5. AUC_score outcome of ADPW-FLHHO approach with other systems on UCSDPed2 dataset.

Methods	AUC Score (%)
ADPW-FLHHO	99.19
EADN	98.30
Binary SVM classifier	97.16
MIL-C3D	95.50
TSN-Optical Flow	94.36
Spatiotemporal	92.48
TSN-RGB	90.44

shows a comparative $AU{C}_{score}$ inspection of the ADPW-FLHHO method with recent approaches on the UCSDPed2 dataset. The figure demonstrates the ineffectual outcomes of the TSN-RGB, Spatiotemporal, and TSN-Optical Flow with closer $AU{C}_{score}$ of 90.44%, 92.48%, and 94.36%, correspondingly. Contrastingly, the MIL-C3D and Binary SVM models have managed to report a certainly increased $AU{C}_{score}$ of 95.5% and 97.16%, correspondingly. Although the EADN technique accomplishes a considerable $AU{C}_{score}$ of 98.30%, the ADPW-FLHHO technique ensures its supremacy with a higher $AU{C}_{score}$ of 99.19%.

Table 6. TPR outcome of ADPW-FLHHO approach on Avenue dataset.

True Positive Rate
False Positive Rate	MPPCA	Social Force	MDT	AMDN	EADN	ADPW-FLHHO
0	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
5	0.1084	0.1843	0.2188	0.3292	0.3407	0.5694
10	0.2476	0.2932	0.4303	0.4593	0.5529	0.5850
15	0.3232	0.4174	0.5447	0.5918	0.5960	0.7382
20	0.4629	0.4502	0.5807	0.6878	0.7256	0.8654
25	0.6024	0.6523	0.7218	0.7275	0.7386	0.8868
30	0.6391	0.7687	0.7407	0.8876	0.9217	0.9468
35	0.7099	0.7881	0.7773	0.9022	0.9497	0.9578
40	0.7830	0.8593	0.8935	0.9238	0.9735	0.9597
45	0.8298	0.9280	0.9105	0.9715	0.9773	0.9716
50	0.8308	0.9353	0.9174	0.9730	0.9829	0.9819
55	0.8815	0.9412	0.9633	0.9790	0.9857	0.9858
60	0.8997	0.9533	0.9674	0.9936	0.9917	0.9941
65	0.9449	0.9601	0.9739	0.9936	0.9965	1.0000
70	0.9552	0.9645	0.9789	0.9945	1.0000	1.0000
75	0.9647	0.9882	0.9922	0.9955	1.0000	1.0000
80	0.9726	1.0000	1.0000	1.0000	1.0000	1.0000
85	0.9795	1.0000	1.0000	1.0000	1.0000	1.0000
90	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000
95	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000
100	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000

describes the TPR results of the ADPW-FLHHO technique on the Avenue dataset. The figure indicates that the values of TPR are raised with an increase in FPR. It is noticed that the MPPCA, SF, and MDT models resulted in the lowest TPR values while the AMDN model accomplishes certainly enhanced TPR values. Contrastingly, the EADN model accomplishes near-optimal performance with reasonable TPR values. Nevertheless, the ADPW-FLHHO technique reaches maximum performance with maximal TPR values on the UCSDPed-1 dataset.

Table 7. AUC_score outcome of ADPW-FLHHO approach with other systems on Avenue dataset.

Methods	AUC Score (%)
ADPW-FLHHO	98.90
EADN	97.78
Binary SVM classifier	96.21
MIL-C3D	95.02
TSN-Optical Flow	93.31
Spatiotemporal	91.41
TSN-RGB	89.47

highlights a comparative $AU{C}_{score}$ inspection of the ADPW-FLHHO technique with recent methods on the Avenue dataset. The results demonstrate the ineffectual outcomes of the TSN-RGB, Spatiotemporal, and TSN-Optical Flow with a closer $AU{C}_{score}$ of 89.47%, 91.41%, and 93.31%, respectively. Contrastingly, the binary MIL-C3D and Binary SVM approaches have reported a certainly increased $AU{C}_{score}$ of 95.02% and 96.21%, respectively.

Though the EADN model accomplishes a considerable $AU{C}_{score}$ of 97.78%, the ADPW-FLHHO technique ensures its supremacy with a maximum $AU{C}_{score}$ of 98.90%. From the detailed results and discussion, it is assumed that the ADPW-FLHHO technique highlights proficient results over other models. The proposed model can alert authorities to identify abnormal objects in the pedestrian walkways, which results in a quick response time during an emergency, and potentially prevent accidents or other safety incidents. This can enhance the accessibility for people with disabilities or mobility issues, making pedestrian walkways more inclusive. It has the potential to improve safety, security, efficiency, accessibility, and urban planning, making pedestrian walkways more pleasant and inclusive spaces for everyone.

5. Conclusions

In this study, a novel ADPW-FLHHO algorithm was developed for anomaly detection in pedestrian walkways, to improve the safety of pedestrians. The presented ADPW-FLHHO technique mainly concentrated on the effectual identification and classification of anomalies, i.e., vehicles in the pedestrian walkways. To accomplish this, the ADPW-FLHHO technique uses the HybridNet model for feature vector generation. In addition, the HHO method is implemented for the optimal hyperparameter tuning process. For anomaly detection, the ADPW-FLHHO technique used the MDBN model. The experimental results of the ADPW-FLHHO technique can be studied on the UCSD anomaly detection dataset and the results illustrated the promising performance of the ADPW-FLHHO approach over existing models. In future, the proposed model can be tested on a large-scale real-time remote sensing image dataset. Additionally, the computation complexity of the proposed model can be investigated in future.