YOLO-ABD: A Multi-Scale Detection Model for Pedestrian Anomaly Behavior Detection

Abstract

Public safety and intelligent surveillance systems rely on anomaly detection for effective monitoring. In real-world pedestrian detection scenarios, Pedestrians often exhibit various symmetrical features such as body contours, facial features, posture, and clothing. However, the accuracy of pedestrian anomaly detection is affected by factors such as complex backgrounds, pedestrian obstruction, and small target sizes. To address these issues, this study introduced YOLO-ABD, a lightweight method for anomaly behavior detection that integrated small object detection and channel shuffling. This approach enhanced the YOLOv8n baseline model by integrating a small-object detection mechanism at the head and employing the symmetric GSConv convolutional module in the backbone network to improve perceptual capabilities. Furthermore, it incorporated the SimAM attention mechanism to mitigate complex background interference and thus enhance target detection performance. Evaluation on the IITB-Corridor dataset showed mAP50 and mAP50-95 scores of 89.3% and 60.6%, respectively. Generalization testing on the street-view-gdogo dataset further underscored the superiority of YOLO-ABD over advanced detection algorithms, demonstrating its effectiveness and generalization capabilities. With relatively fewer parameters, YOLO-ABD provided an excellent lightweight solution for pedestrian anomaly detection.

1. Introduction

Human anomaly detection involves real-time analysis and recognition of human behaviors in surveillance videos or images, aiming to identify deviations from typical behavior patterns [1,2]. Such anomalies may include activities like falling, climbing, running, and carrying hazardous items. This technology has widespread applications in public safety and intelligent monitoring systems, facilitating prompt identification of security threats and emergencies to effectively mitigate accidents [3,4,5,6].

With advancements in modeling, mainstream methods for anomaly behavior detection can be categorized into three types: frame reconstruction-based methods, frame prediction-based methods, and end-to-end anomaly score calculation methods [7,8,9,10,11,12,13,14]. Frame reconstruction methods leveraging deep learning often rely on Auto-Encoders [15]. Gong et al. [16] introduced the Memory-Augmented Autoencoder (MemAE), enhancing reconstruction by incorporating a memory module. MemAE selects the most relevant memory items for reconstruction using the encoded query, ensuring that testing results closely resemble normal samples. This approach aims to minimize the reconstruction error for normal behavior while maximizing it for anomalous behavior. Frame prediction-based methods involve feeding a video segment into a predictor to predict a frame; significant deviations between the predicted and current input frames indicate abnormal behavior [17,18]. In [19], a frame prediction network based on multipath ConvGRU was employed, capable of handling semantically informative objects and areas of various scales, capturing spatial-temporal dependencies in normal videos. To mitigate interference from background noise, a noise tolerance loss was introduced during training. However, samples in a static state may be misclassified as background due to the lack of differences between adjacent action frames, potentially leading to missed detections [20,21].

The YOLOv8 model adopts an end-to-end detection approach, enabling real-time object detection with high accuracy across diverse scenes and object types [22]. Its concise and lightweight architecture makes it suitable for deployment in resource-constrained environments. However, anomaly behavior detection presents challenges such as variations in spatial orientation, size, shape, overlapping backgrounds, and limited detection precision. To address these challenges, this paper proposes an enhanced YOLOv8-based method for anomaly behavior detection, termed YOLO-ABD, integrating small object detection and Group Shuffle Convolution (GSConv) [23]. The key contributions of this study are summarized as follows:

• Introducing an end-to-end pedestrian anomaly detection method that utilizes the SimAM attention mechanism [24] to reduce background interference. Additionally, it incorporates a custom-designed small-object detection head to identify pedestrian anomalies at various scales.
• Integrating the GSConv (Group Shuffle Convolution) module with a symmetrical structure enhances the model’s accuracy. Moreover, adopting shuffling strategies decreases the computational complexity, thereby achieving the goal of lightweighting the model.
• The proposed method is trained and validated on public datasets for anomaly behavior detection. Generalization testing in traffic scene detection demonstrates significant performance improvements over existing methods.

2. Related Works

Classical methods for anomaly detection typically involve extracting features from videos or images to identify abnormal behaviors. Motion features are derived from various dimensions such as optical flow maps, heat maps, and depth maps. These motion features can be integrated with trajectory information or motion magnitude to fuse high-level semantic features with low-level details, achieving precise detection of abnormal behaviors [25,26,27,28]. For instance, Xie et al. [29] applied spatiotemporal representation learning to identify behaviors such as sleeping and using mobile phones by analyzing college students’ movement trajectories. Banerjee et al. [30] introduced a deep convolutional network architecture for detecting behavioral patterns of students and teachers in laboratory settings. Guan et al. [31] employed a 3D-CNN to extract features from optical flow and motion history images, using LSTM networks to capture spatiotemporal features for anomaly detection. These studies highlight the effectiveness of classical approaches in recognizing abnormal behaviors. However, these methods may struggle with detecting diverse abnormal behaviors across different scenarios and can be susceptible to interference from complex backgrounds, which poses challenges for multi-person anomaly detection in complex environments.

Modern anomaly detection methods also incorporate multi-object detection techniques to identify abnormal behaviors [32,33,34,35,36]. Researchers apply these approaches to detect abnormal behaviors in individuals or crowds within specific contexts. Object detection frameworks broadly fall into two categories: two-stage and single-stage approaches [37]. Two-stage algorithms, exemplified by Faster R-CNN, have been used for anomaly detection and classification, enhancing accuracy but requiring substantial computational resources and being less suitable for real-time applications, as demonstrated by Mansour et al. [38] and Hongchao et al. [39], who proposed enhanced models for identifying hazardous behaviors in factory workers.

In contrast, single-stage algorithms, exemplified by the YOLO series, are also employed for anomaly detection [40,41]. For example, in [42], the YOLOv3 model was used to detect unsafe behaviors at gas stations like smoking and mobile phone usage. Benjumea et al. [43] modified YOLOv5 to improve its ability to detect small objects in autonomous driving scenarios, while Xiao et al. [44] applied YOLOv5 to monitor safety in substations. The YOLO series simplifies the detection process, offering efficient computation and faster speeds, making it suitable for practical scenarios.

3. Methodology

3.1. The General Structure of YOLO-ABD

In the network architecture of YOLO-ABD, we adopt YOLOv8n as the base model, enhancing it with a small-object detection head and integrating GSConv convolutional blocks along with the SimAm attention mechanism for improvement. The network structure, illustrated in Figure 1, consists of three main components: the backbone, the neck, and the detection head. In the backbone, GSConv modules replace the original Conv modules of YOLOv8n, while the SimAm attention is integrated before the SPPF module. GSConv maintains implicit channel connections to reduce model complexity while retaining the learning capacity of standard convolutions, thereby enhancing feature extraction for abnormal behavior characterization. The SimAm attention mechanism mitigates background interference while maintaining parameter efficiency, thereby improving accuracy. To further enhance the network’s capability to perceive abnormal behaviors across multiple scales, new pathways and detection heads for small objects are designed within the neck and head sections to prevent overlooking or misclassifying distant pedestrians.

3.2. Baseline Model

YOLOv8 models are categorized into five versions based on network depth, width, and maximum number of channels: n, s, m, l, and x. Among these, YOLOv8n stands out with the least parameters and floating-point operations, making it highly efficient and suitable for real-time deployment while maintaining high accuracy. This paper adopts YOLOv8n as the baseline model, depicted in Figure 2. YOLOv8 consists of three main components: backbone, neck, and head. The backbone includes Conv modules, C2f modules, and an SPPF module focused on extracting features from input images. Conv modules process images or feature maps, generating new maps with reduced resolution and increased channels. C2f modules excel in deep and multi-scale feature extraction. The neck integrates multi-scale features through a combination of Feature Pyramid Network (FPN) [45] and Path Aggregation Network (PAN) [46], employing techniques like upsampling, channel concatenation, and deep feature extraction. These features are then passed to the head, which utilizes a decoupled head structure [47] to separate classification and prediction tasks. This design enables precise object detection regarding size and position using feature maps of varying scales.

3.3. Small Object Detection Head

In the IITB-Corridor dataset [48], distant pedestrians occupy a relatively small portion of the image. When resizing the dataset images to 640 × 640 pixels, the input image includes numerous small targets. However, after multiple upsampling and pooling operations in the neck network, many detection features associated with small targets are lost, resulting in missed detections. The baseline model includes detection heads sized at 80 × 80, 40 × 40, and 20 × 20. The 80 × 80 detection head’s receptive field is overly extensive for small targets, causing the baseline model to struggle in accurately detecting abnormal behaviors of distant pedestrians.

To address this issue, we enhanced the baseline model by introducing a 160 × 160 small target detection head. The structural diagram for this new detection head is depicted in Figure 1. Minimal modifications were made to the original model structure, as shown in Figure 1 for clarity. The initially upsampled feature map in the neck layer, originally sized at 40 × 40, undergoes two additional upsampling steps to produce a detailed 160 × 160 feature map rich in small target information. This feature map is then concatenated with the 160 × 160 feature map from the second layer of the backbone network, enhancing the model’s capability to detect small target behaviors at the 160 × 160 scale. Finally, this refined feature map is fed into the detection head layer, creating a new small target detection head that effectively reduces missed and false detections of abnormal behaviors across various scales.

3.4. GSConv Module

Currently, lightweight network design heavily relies on Depth-wise Separable Convolution (DSConv) to reduce model parameters and floating-point operations [49,50,51]. DSConv conducts convolutions independently across three channels, which minimizes redundant feature information but also fully separates the channel information of input data, thereby reducing the model’s feature extraction capability compared to dense channel convolutions, such as standard convolutions. To address this limitation, Zhang et al. [52] introduced “ShuffleNet”, which employs pointwise group convolution and channel shuffle operations to significantly decrease computational cost while maintaining accuracy. GhostNet [53] generates multiple feature maps through a series of low-cost linear transformations, effectively revealing intrinsic feature information.

Li et al. [23] introduced GSConv as an alternative to standard convolutions, offering a computational cost approximately 60–70% lower while retaining comparable learning capability. Illustrated in Figure 3, GSConv initially applies a standard convolution to the input feature map with $C_1$ channels, resulting in an intermediate feature map with $C_2$/2 channels. Subsequently, this intermediate feature map undergoes Depth-wise Separable Convolution (DSConv) to produce another intermediate feature map, also with $C_2$/2 channels. Finally, the two intermediate feature maps are concatenated and shuffled to yield an output feature map with $C_2$ channels. This approach mitigates information loss that can occur with DSConv’s channel separation, while maintaining an output similar to that of standard convolution. The time complexity formulas for standard convolution (SC), Depth-wise Separable Convolution (DSC), and GSConv are shown in Equations (1), (2) and (3), respectively:

$$O\left(W·H·K_1·K_2·C_1·C_2\right)$$

(1)

$$O\left(W·H·K_1·K_2·C_1\right)$$

(2)

$$O\left(\left[W·H·K_1·K_2·C_2\right]/2·\left(C_1+1\right)\right)$$

(3)

In the above formulas, W and H represent the width and height of the output feature map, $K_1$ and $K_2$ are the sizes of the convolution kernels, $C_1$ is the number of input feature channels, and $C_2$ is the number of output feature channels.

3.5. SimAm Attention Module

To enhance the model’s capability in detecting objects and mitigate interference from complex backgrounds, Liu et al. [54] introduced the SimAM attention mechanism before the detection head, effectively reducing background noise interference, which allowed the model to focus more precisely on the small target of corn silk, significantly enhancing the model’s ability to extract distinguishable features. In the study in [55], to address the challenges of small target detection and significant sea surface interference in maritime environments, the authors enhanced the YOLOv7 model by incorporating a small-target detection head to better detect tiny targets. Additionally, they integrated the SimAM module to identify attention regions within the scene, thereby reducing sea surface interference and markedly improving the model’s performance. Based on the case study, the SimAM module demonstrates a significant impact on reducing background interference and enhancing model performance in object detection tasks.

This paper introduces SimAm, a parameter-free attention mechanism. SimAm, rooted in neural network theory, distinguishes itself from existing attention mechanisms [56,57], which typically focus on either the channel or spatial domain. Unlike CBAM [58] and GAM [59], which extract features from both channel and spatial dimensions and merge them into a hybrid attention mechanism, SimAm integrates spatial, channel, and feature dimensions to generate 3D weights. Figure 4 illustrates the generation of these 3D weights.

In visual neuroscience, the activity of an active neuron can induce spatial suppression in the surrounding neurons. SimAm utilizes this concept by prioritizing neurons that induce spatial suppression. The priority of each neuron is determined by an energy function, represented by Equation (4).

$$e_t(w_t,b_t,x_i)={\displaystyle \frac{1}{M-1}}\sum _{i=1}^{M-1}{(-1-(w_t{x}_i+b_t))}^2+{(1-(w_t+b_t))}^2+\lambda w_t^2$$

(4)

In the formula, t and $x_i$ represent the target neuron and other neurons of a single channel in the input features, respectively. Here, i is the index in the spatial dimension, and $M=H\times W$ denotes the number of neurons in that single channel. The weight and bias of the target neuron, denoted as $w_t$ and $b_t$, respectively, are computed using the formulas shown in (5) and (6):

$$w_t=-{\displaystyle \frac{2(t-{\mu }_t)}{{(t-{\mu }_t)}^2+2{\sigma }_t^2+2\lambda }}$$

(5)

$$b_t=-{\displaystyle \frac{1}{2}}(t+{\mu }_t)w_t$$

(6)

In the formula, ${\mu }_t={\displaystyle \frac{1}{M-1}}{\sum }_{i=1}^{M-1}{x}_i$ and ${\sigma }_t^2={\displaystyle \frac{1}{M-1}}{\sum }_{i=1}^{M-1}{(x_i-{\mu }_t)}^2$. When ensuring that all pixels within a single channel follow the same distribution, the minimum energy function can be obtained:

$$e_t^{*}={\displaystyle \frac{4(\widehat{\sigma }+\lambda )}{{(t-\widehat{\mu })}^2+2{\widehat{\sigma }}^2+2\lambda }}$$

(7)

In Formula (7), $\widehat{\mu }$ and $\widehat{\sigma }$ represent the mean and variance of all neurons except t. A smaller $e_t^{*}$ indicates higher importance, and the importance calculation is $1/e_t^{*}$. Scaling the energy of neurons for feature refinement within each neuron yields the final feature map, as shown in Formula (8):

$$\tilde{X}=sigmoid({\displaystyle \frac{1}{E}})\odot X$$

(8)

To assess the effectiveness of the SimAm attention module, two images showcasing abnormal behaviors were selected from the IITB-Corridor dataset. Feature extraction was conducted using both the baseline model and the SimAm-enhanced model, followed by visualization of the resulting feature heatmaps, as shown in Figure 5. It is evident that following the integration of the SimAM attention module, the neural network’s focus on human actions exhibits a higher degree of congruence with the human brain’s attention to action-related features. The distribution of heatmap values (represented by warm-colored regions) is more prominently concentrated on the human body or the actions being performed by it, thereby mitigating the interference caused by extraneous background elements. Thus, the introduction of the SimAM attention module has significantly enhanced the baseline model’s capability to detect abnormal human behavior.

4. Experiments

4.1. Dataset

The IITB-Corridor dataset [48] was curated to investigate abnormal human activities within the corridors of the Indian Institute of Technology Bombay. These videos were captured by fixed cameras overlooking pedestrian corridors on the Mumbai campus. The dataset encompasses a variety of both normal and abnormal behaviors. Normal activities typically include common behaviors such as walking and standing, whereas abnormal behaviors encompass relatively rare and potentially hazardous activities like running, fighting, playing soccer, and riding a bicycle. Notably, some abnormal behaviors are naturally occurring incidents captured by surveillance videos, reflecting unexpected situations in daily life, while others are intentionally arranged by the dataset creators to simulate behavior patterns in specific contexts. These abnormal behaviors vary at both individual and group levels, providing a rich analytical dimension for research. Furthermore, in public settings such as campus hallways, video data are usually obtained through surveillance cameras, exhibiting several distinct characteristics. Firstly, the high density of people in the data results in numerous pedestrians within the frame. Secondly, due to the limitations of camera position and angle, distant human targets are small and difficult to identify, while close human targets may be partially out of the camera’s view. Additionally, overlapping individuals and shadow occlusion significantly increase the difficulty of behavior recognition. These features not only complicate data analysis but also pose higher demands on the accuracy and robustness of algorithms.

To manage the extensive collection of abnormal and non-abnormal video frames in the IITB-Corridor dataset, abnormal frames were sampled at a rate of three frames per second from each video depicting abnormal behavior. Subsequently, all abnormal behaviors were annotated using the labelImg annotation tool, as illustrated in

Table 1. Categories and number of labels in the IITB-Corridor dataset.

Dataset	Abnormal Behavior	Number of Boxes
IITB-Corridor	Bag Exchange	209
	Cycling	577
	Suspicious Object	2255
	Running	2301
	Fighting	2072
	Hiding	396
	Playing With Ball	2058
	Protest	5575

.The adjusted dataset consists of 18,674 video frame images, covering eight categories of abnormal behaviors: Bag Exchange, Cycling, Suspicious Object, Running, Fighting, Hiding, Playing With Ball, and Protest, with representative abnormal frame images shown in Figure 6. In this paper, the first three letters of each category denote a specific class (e.g., “Bag” corresponds to “Bag Exchange”). To facilitate model training and validation, the dataset was divided into training and validation sets in an 8:2 ratio.

The street-view-gdogo dataset [60] contains images captured by fixed traffic cameras positioned along urban roads in Turkey. Through manual annotation, common objects encountered in urban street traffic, such as Person, Car, Bus, Motor, and Bicycle, were categorized into five classes. The dataset comprises a total of 6685 images, which were divided into training, validation, and test sets at a ratio of 8:1:1.

4.2. Training Setting

The experiment on abnormal behavior detection was conducted on the Ubuntu 20.04 operating system, utilizing Python 3.8 as the programming language and CUDA 12.2 and PyTorch 1.9.0 for developing the deep learning framework. The hardware configuration included an NVIDIA GeForce RTX 3080 GPU with 10GB of memory and an Intel Xeon Silver 4210R CPU with 128GB of memory. Experimental parameters were set as follows: a batch size of 16 and training for 200 epochs and employing the Adam optimizer with an initial learning rate of 0.01, weight decay set to 0.0005, and a confidence threshold for object detection of 0.25. Image preprocessing involved resizing images to dimensions of 640 × 640 pixels.

4.2.1. Evaluating Indicators

To evaluate the practical performance of YOLO-ABD, the experiments employ established model evaluation metrics from the object detection domain. These metrics include precision, recall, average precision (AP), mean average precision (mAP), giga floating-point operations per second (GFLOPs), parameters, and frames per second (FPS). Precision measures the ratio of correctly identified positive samples to all samples classified as positive by the model, while recall evaluates the ratio of correctly identified positive samples to all actual positive samples. The definitions of precision and recall are as follows:

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(9)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(10)

TP represents true positives, indicating the number of positive samples correctly predicted as positive by the model; FP represents false positives, indicating the number of negative samples incorrectly predicted as positive by the model; FN denotes false negatives, indicating the number of positive samples incorrectly predicted as negative by the model. The precision–recall (PR) curve illustrates the relationship between recall (X-axis) and precision (Y-axis) across various thresholds. AP represents the area under the PR curve for each class, while mAP denotes the average AP value across all classes. The calculation formulas are as follows:

$$AP={\int }_0^{1}P\left(R\right)dR$$

(11)

$$mAP={\displaystyle \frac{1}{N}}\sum _{i=1}^{N}A{P}_i$$

(12)

mAP@50 denotes the mean average precision at an IoU threshold of 0.5, while mAP@50:95 signifies the mean average precision across IoU thresholds ranging from 0.5 to 0.95. Parameters and GFLOPs are utilized to assess model size and computational efficiency, as well as to evaluate the model’s hardware requirements. Frames per second (FPS) represents the number of images the model can process per second, providing insights into the model’s real-time performance and detection speed. These metrics collectively offer comprehensive insights into the effectiveness and efficiency of the model in practical scenarios.

4.2.2. Result Analysis

To validate the performance improvement of YOLO-ABD, comparative experiments were conducted on the IITB-Corridor dataset to evaluate various object detection models. Key evaluation metrics such as AP, mAP@50, GFLOPs, and parameters were used to assess model performance and compare it with several state-of-the-art methods. Each category of abnormal behavior is represented by the first three letters: “Bag” for “Bag Exchange”, “Cyc” for “Cycling”, “Sus” for “Suspicious Object”, “Run” for “Running”, “Fig” for “Fighting”, “Hid” for “Hiding”, “Pla” for “Playing With Ball”, and “Pro” for “Protest”. The experimental results are summarized in

Table 2. Comparison results on the IITB-Corridor dataset.

Methods	Bag	Cyc	Sus	Run	Fig	Hid	Pla	Pro	mAP50	GFLOPs	Parameter	FPS
Faster-RCNN	76.5%	87.3%	96.0%	70.2%	90.4%	85.2%	66.9%	87.5%	82.5%	-	-	-
YOLOv3	81.0%	78.6%	97.7%	71.3%	94.6%	98.1%	81.7%	90.5%	86.7%	18.9	12.1M	467
YOLOv5s	70.4%	82.0%	94.8%	55.4%	89.0%	95.7%	66.8%	89.2%	80.4%	23.8	8.6M	381
YOLOv5n	48.6%	76.0%	94.1%	55.8%	88.8%	92.4%	66.8%	88.5%	76.4%	7.1	2.3M	503
YOLOv6s	73.7%	85.3%	96.9%	67.2%	93.6%	95.0%	72.8%	91.4%	84.5%	44.0	15.5M	352
YOLOv6n	68.8%	81.6%	96.4%	59.7%	91.7%	94.2%	67.1%	89.6%	81.2%	11.8	4.2M	509
YOLOv8s	76.7%	88.1%	96.6%	62.0%	91.1%	96.5%	72.4%	90.9%	84.3%	28.5	11.1M	380
YOLOv8n	75.2%	84.4%	96.9%	79.6%	94.8%	94.8%	73.7%	91.5%	86.4%	10.7	2.49M	347
YOLOv9t	64.9%	75.8%	94.5%	70.1%	93.3%	93.7%	65.7%	89.6%	81.0%	8.7	2.86M	504
YOLOv10n	69.3%	77.0%	95.1%	77.8%	89.1%	92.9%	69.2%	90.5%	82.6%	8.2	2.57M	689
Ours	76.7%	92.5%	96.7%	84.6%	95.6%	95.7%	80.5%	92.2%	89.3%	11.4	2.56M	436

, offering a comprehensive comparison of model performance.

All models were trained from scratch to ensure fairness, without employing pre-training methods. Analysis of the experimental results indicates that YOLO-ABD achieves superior detection performance on the IITB-Corridor dataset, with an mAP@50 score of 89.3%. YOLO-ABD shows a performance improvement of 6.8 percentage points compared to the early-stage two-stage object detection method Faster R-CNN [61]. Significant advancements are observed compared to the YOLO series, particularly in the “Cyc” and “Run” categories. Compared to YOLOv3, YOLO-ABD demonstrates superior performance across most categories, notably in “Run”, “Pla”, and “Pro”, with a 2.6 percentage point increase in mAP@50. Compared to the YOLOv5s and YOLOv5n models, YOLO-ABD achieves higher accuracy across all categories, with mAP@50 scores improving by 8.1 and 12.9 percentage points, respectively. When compared to the YOLOv6s and YOLOv6n models, YOLO-ABD notably enhances accuracy in the “Pro” category, with mAP@50 scores increasing by 4.8 and 8.1 percentage points, respectively. Against the YOLOv8s and YOLOv8n models, YOLO-ABD surpasses them in nearly all categories, particularly in “Run”, “Pla”, and “Pro”, with significant improvements in accuracy and mAP@50 scores. Compared to the latest YOLOv9t and YOLOv10n models, the YOLO-ABD model achieves higher accuracy across all categories, with mAP@50 scores improving by 8.3 and 6.7 percentage points, respectively. Overall, YOLO-ABD exhibits significant performance advantages by achieving higher mAP@50 values for pedestrian anomaly detection while reducing the number of parameters. It retains the lightweight characteristics of the YOLO series, demonstrating that the proposed model offers considerable advancements.

The PR curves shown in Figure 7 provide a visual comparison between our proposed method and the YOLO variants with similar computational complexity on the IITB-Corridor dataset. Our method exhibits a smoother curve, positioned closer to the top-right corner, indicating superior precision and recall performance. This demonstrates a more balanced detection capability, with notable improvements across various classes, particularly evident in the “Bag Exchange” category. These findings underscore the enhanced effectiveness of our proposed method in detecting abnormal behaviors, validating its superiority over existing YOLO variants.

Comparison of the actual detection results between the baseline model and our proposed model, shown in Figure 8 and Figure 9, respectively, reveals significant differences. The baseline model YOLOv8n demonstrates noticeable missed detections on the IITB-Corridor dataset, particularly in scenarios involving crowded and overlapping individuals. In contrast, our proposed model exhibits enhanced accuracy, effectively detecting abnormal targets even under challenging conditions. This improvement is reflected in the higher accuracy and precision of the detection boxes for individuals. Upon analysis of the results depicted in Figure 8 and Figure 9, our proposed model demonstrates superior performance in detecting abnormal behaviors such as “Cycling”, “Suspicious Object”, “Protest”, and “Fighting” compared to the baseline model. Notably, it accurately identifies the “Protest” abnormal behavior, which the baseline model fails to detect, without any missed detections, highlighting the efficacy of our approach. In summary, these findings validate the advanced and effective performance of our proposed model.

4.2.3. Generalization Study

To evaluate the generalization performance of YOLO-ABD in object detection tasks, this study utilized the street-view-gdogo traffic object detection dataset for testing and compared it with several contemporary advanced object-detection models. The experimental results, summarized in

Table 3. Comparison results on the street-view-gdogo dataset.

Methods	Bicycle	Bus	Car	Motor	Person	mAP50	GFLOPs	Parameter	FPS
Faster-RCNN	86.7%	89.8%	94.7%	80.9%	79.1%	86.3%	-	-	-
YOLOv3	92.1%	94.2%	94.3%	80.9%	68.2%	85.9%	18.9	12.1M	95
YOLOv5s	93.4%	95.7%	96.6%	89.8%	86.6%	92.3%	23.8	8.6M	806
YOLOv5n	91.9%	96.1%	96.6%	89.3%	84.9%	91.8%	7.1	2.3M	111
YOLOv6s	93.4%	95.7%	96.3%	90.3%	85.2%	92.2%	44.0	15.5M	63
YOLOv6n	88.2%	91.3%	95.7%	86.8%	76.6%	87.7%	11.8	4.2M	105
YOLOv8s	94.7%	96.8%	96.7%	91.1%	87.3%	93.3%	28.5	11.1M	82
YOLOv8n	92.5%	95.7%	96.5%	89.1%	85.5%	91.9%	8.1	2.49M	65
YOLOv9t	89.3%	94.2%	95.9%	87.3%	76.1%	88.5%	8.7	2.86M	92
YOLOv10n	91.2%	94.9%	96.3%	87.5%	83.1%	90.6%	8.2	2.57M	139
Ours	93.5%	95.4%	97%	89.9%	88.1%	92.8%	11.4	2.56M	122

, demonstrate the performance of our model across five classes: Bicycle, Bus, Car, Motor, and Person. Our model achieved notable accuracy rates of 93.5%, 95.4%, 97%, 89.9%, and 88.1% for these respective classes. Comparative analysis with the baseline model YOLOv8n reveals that our model consistently outperforms in most categories, with a significant 2.6% increase observed in the Person class. These findings highlight the effectiveness and robust generalization capabilities of our model across diverse scenarios.

4.2.4. Ablation Study

Table 4. Comparisons of SC, DSC, and GSConv modules on the IITB-Corridor dataset.

Basic Convolutional Method	mAP50	mAP50-95	GFLOPs	Parameter	FPS
SC	86.4%	57.4%	8.1	2.86M	504
DSC	85.6%	55.8%	7.3	2.5M	505
GSConv	87.5%	58.1%	7.7	2.68M	552

compares the accuracy of the YOLOv8n model integrated with three different convolution modules (SC, DSC, and GSConv) on the IITB-Corridor dataset. To ensure the reliability of the evaluation, each convolution module was applied to the backbone network under identical conditions. The results indicate that the GSConv module achieved superior performance with a lower time complexity. The advantage of the GSConv module lies in the efficient collaboration between its main and auxiliary branches: the main branch captures channel features using the SC module, while the auxiliary branch focuses on spatial information using the DSC module, with a final fusion of these features through a mixing operation. Consequently, the GSConv module captures more spatial and channel features while maintaining lower computational costs, a feat that other convolution methods struggle to achieve simultaneously in optimizing computational efficiency, spatial information, and channel features. Figure 10 displays the visualized feature maps of the SC, DSConv, and GSConv modules. The texture features of the GSConv module are noticeably more similar to the SC module than those of the DSConv module compared to the SC module.

To validate the effectiveness of YOLO-ABD and assess the contributions of SimAm, GSConv, and the small-object detection heads to the model’s accuracy and computational efficiency, eight sets of comparative experiments were conducted. Using YOLOv8n as the baseline model, ablation experiments were conducted on the IITB-Corridor dataset with an input image size of 640 × 640. In the experiments, the positions and quantities of module replacements correspond precisely to the improvements in the YOLO-ABD model. The experimental results are shown in

Table 5. Experimental study of ablation on the IITB-Corridor dataset.

SimAm	GSConv	Small	Precision	Recall	mAP50	mAP50-95	GFLOPs	Parameter	FPS
-	-	-	83.8%	79.3%	86.4%	57.4%	8.1	2.86M	504
✓	-	-	83.8%	79.5%	86.7%	57.4%	8.1	2.86M	507
-	✓	-	84.4%	78.1%	87.5%	58.1%	7.7	2.68M	552
✓	✓	-	85.0%	79.0%	88.0%	58.2%	7.7	2.68M	504
-	-	✓	85.1%	79.4%	86.8%	57.7%	11.8	2.75M	420
✓	-	✓	85.0%	81.1%	87.4%	58.2%	11.8	2.75M	405
-	✓	✓	84.1%	83.0%	88.9%	59.2%	11.4	2.56M	422
✓	✓	✓	83.9%	81.5%	89.3%	60.6%	11.4	2.56M	436

.

From the experimental findings, it is evident that integrating the SimAm module alone improved the model’s mAP50 metric by 0.3 percentage points. The SimAm module enhances performance without adding parameters, thus avoiding increased model complexity or computational cost, and has negligible impact on floating-point operations.Introducing the GSConv module alone improved the model’s mAP50 and mAP50-90 metrics on the dataset by 1.1 and 0.7 percentage points, respectively, while reducing floating-point operations by 0.4 GFLOPs. Adding the small-object detection head alone enhanced the model’s mAP50 and mAP50-90 metrics by 0.4 and 0.3 percentage points, respectively. However, integrating the small-object detection path increased floating-point operations from 8.1 to 11.8 GFLOPs, slightly reducing detection speed, though still suitable for real-time applications.

Combining pairs of the three modules significantly improved both mAP50 and mAP50-90 metrics. By incorporating a small-object detection head and the GSConv module, the mAP50 metric was increased to 89%, and the mAP50-90 metric was improved to 59.2%. This enhancement was achieved while keeping additional floating-point operations to a minimum and significantly reducing the parameter count of the baseline model. Integrating all three modules seamlessly merged with the baseline model, maximizing efficiency and resulting in improvements of 2.9 and 3.2 percentage points in mAP50 and mAP50-90 metrics, respectively. Although integrating the small-object detection head increased floating-point operations, the GSConv module effectively reduced the baseline model’s overall floating-point operations.

Ablation experiments were conducted by individually integrating different modules based on the baseline model, followed by anomaly behavior testing and comparative diagram generation, as depicted in Figure 11. These diagrams illustrate that while individual integration of SimAm, GSConv, and the small-object detection head each improves performance over the original YOLOv8n model, their simultaneous integration yields substantial enhancements in anomaly behavior detection performance.

5. Conclusions

This study introduces the YOLO-ABD model, a lightweight pedestrian anomaly detection method based on the YOLOv8n baseline network. YOLO-ABD addresses the challenge of small-object detection performance by incorporating a dedicated small target detection head. The model integrates the GSConv module to optimize channel connections while minimizing computational cost. Additionally, the inclusion of the SimAM attention module further reduces background interference and enhances model accuracy. Experimental results on the IITB-Corridor dataset demonstrate that YOLO-ABD achieves mAP50 and mAP50-95 scores of 89.3% and 60.6%, respectively, without requiring additional pre-training. Compared to current state-of-the-art object detection algorithms, YOLO-ABD strikes an optimal balance between speed and accuracy. It effectively reduces false negatives and false positives caused by background interference and occlusion, achieving an excellent combination of accuracy and lightweight design. Experiments on the street-view-gdogo dataset highlight the generalization capability of YOLO-ABD.

This study proposes a supervised learning-based method for pedestrian anomaly detection. Given the scarcity of pedestrian anomaly behavior data, initial experiments require extensive manual annotation of pedestrian anomaly behaviors, resulting in substantial data processing and manual labor. Future research could explore the use of unsupervised or semi-supervised learning methods to reduce the dependency of anomaly detection models on labeled data, achieving high performance and accuracy with a small amount of data.