Abnormal behavior detection method and device based on improved YOLOv8 and storage medium

Abstract

The invention discloses an abnormal behavior detection method and device based on improved YOLOv8 and a storage medium. According to the method, a new FDPN structure is provided to replace a neck network architecture of an original YOLOv8 model; the FDPN has the advantages of a feature focusing module and a feature diffusion mechanism, the feature focusing module effectively captures and integrates cross-scale feature information through parallel deep convolution operation, and the diffusion mechanism promotes contextual information to flow among scales, so that each feature map contains detailed semantic information, the small target detection precision is improved, and the detection efficiency is improved. And the adaptability of the network to shielding and complex backgrounds is enhanced. In addition, an RCSOSA module and a DyHead dynamic prediction head are also introduced; the RCSOSA module extracts and enhances diversified features through a RepVGG module, optimizes feature selection and channel shuffling through a plurality of RCS modules, and enhances feature fusion between channels; and by using the DyHead dynamic prediction head, the recognition capability of the shielded crowd can be improved. According to the method, the robustness and the precision of abnormal behavior detection in a complex scene are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of behavior detection, and more specifically, to an abnormal behavior detection method, device and storage medium based on improved YOLOv8. Background Art

[0002] With the rapid development of intelligent monitoring and behavior recognition technology, monitoring individual behavior in various complex scenarios is crucial to improving system efficiency and quality. In modern society, with the popularity of remote work and virtual interaction, it is particularly important to ensure individual participation and standardized behavior. By using AI technology, the system can monitor individual behavior in real time and intelligently identify and monitor abnormal behavior. This technology can not only help managers better understand the dynamics, but also optimize the environment and improve the overall user experience.

[0003] As an efficient target detection method, the YOLO (You Only Look Once) series of models are widely used due to their real-time performance and high accuracy. As a newer version of the YOLO series, YOLOv8 significantly improves the accuracy and speed of target detection by introducing a deeper feature extraction network and an optimized loss function. However, in practical applications, especially in complex environments, YOLOv8 still faces some challenges. In these complex scenes, the targets are usually small and densely distributed, which makes it difficult for the model to fully distinguish individuals during feature extraction and target positioning. In addition, individuals are prone to occlude each other, further increasing the complexity of the detection task. These factors lead to a decrease in the performance of the model when dealing with dense crowds and occlusion problems, limiting its application in actual behavior recognition and monitoring scenarios. Therefore, for this type of scenario, it is particularly important to further improve YOLOv8's feature extraction capabilities and robustness in dealing with occlusion problems. Summary of the invention

[0004] In order to overcome the technical problems existing in the above-mentioned prior art, the present invention provides an abnormal behavior detection method, device and storage medium based on improved YOLOv8.

[0005] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0006] An abnormal behavior detection method based on improved YOLOv8 includes the following steps:

[0007] S1. Build an abnormal behavior detection model based on improved YOLOv8;

[0008] The RCSOSA module is introduced to replace the C2f module in the Neck network of the original YOLOv8 model to improve the detection accuracy and recognition effect of dense groups;

[0009] S2. Iteratively train the abnormal behavior detection model based on the improved YOLOv8 to obtain a trained abnormal behavior detection model based on the improved YOLOv8;

[0010] S3, inputting the abnormal behavior image to be detected into the trained abnormal behavior detection model based on improved YOLOv8, and outputting the detection result of the abnormal behavior;

[0011] The working process of the RCSOSA module in step S1 includes:

[0012] The abnormal behavior image to be detected is used as the input feature map, and the input feature map is processed by the RepVGG module to extract and enhance diverse features;

[0013] The processed feature map is divided into two parts in a ratio of 1:1 to obtain the first part and the second part; the first part is directly transmitted, and the second part is optimized for feature selection and channel shuffling through multiple RCS modules to enhance feature fusion between channels;

[0014] The two parts of the feature map are spliced ​​in the channel dimension, and the spliced ​​feature map is processed again by the RepVGG module to further optimize the feature representation.

[0015] Furthermore, in step S1, the improvement of the original YOLOv8 model also includes:

[0016] A feature-focused diffusion pyramid network is constructed to replace the Neck network architecture of the original YOLOv8 model to enhance the feature representation and detection accuracy of small targets.

[0017] Further, the feature-focused diffusion pyramid network includes a FocusFeature module, and the FocusFeature module is used to receive an input feature map and generate a processed feature map;

[0018] The feature map processed by the FocusFeature module is assigned to two branches, which are respectively a first branch and a second branch;

[0019] The first branch includes an Unsample module, which performs upsampling processing on the feature map processed by the FocusFeature module to adjust the resolution of the feature map so that it can be spatially aligned with the feature map of a lower level, thereby laying a foundation for subsequent feature fusion;

[0020] The second branch includes a Conv module, and the Conv module further optimizes the quality of the feature map processed by the FocusFeature module by refining the feature representation;

[0021] The feature maps output by the first branch and the second branch are fused with other feature maps through the Concat module.

[0022] Furthermore, the working process of the FocusFeature module includes:

[0023] The FocusFeature module receives three feature maps of different scales, the three feature maps are feature map P3, feature map P4, and feature map P5;

[0024] The feature map P3, the feature map P4, and the feature map P5 are respectively preprocessed differently; the feature map P3 adjusts the spatial size and the number of channels through the ADown downsampling module; the feature map P4 adjusts the number of channels through a 1×1 convolution layer; the feature map P5 adjusts the spatial size through upsampling, and then adjusts the number of channels through a 1×1 convolution layer;

[0025] The preprocessed feature map P3, the feature map P4, and the feature map P5 are spliced ​​in the channel dimension to form a comprehensive feature map containing multi-scale information;

[0026] The comprehensive feature maps are processed separately by a set of parallel depth-wise separable convolutions. The set of depth-wise separable convolutions uses convolution kernels of different sizes to process the comprehensive feature maps. Each depth-wise separable convolution is used to capture features of different scales.

[0027] Superimposing the outputs of all depth-wise separable convolutions with the comprehensive feature map to obtain fused features;

[0028] The fused features are further compressed and refined through 1×1 convolution to output the final feature map.

[0029] Furthermore, in step S1, the improvement of the original YOLOv8 model also includes:

[0030] The DyHead dynamic prediction head is used to replace the detection head of the original YOLOv8 model to improve the recognition ability of occluded people.

[0031] Furthermore, the DyHead dynamic prediction head includes a spatial scale perception module, a spatial position perception module and a task perception module;

[0032] The spatial scale perception module generates attention weights through global pooling and 1×1 convolution to enhance the perception of multi-scale objects;

[0033] The spatial position perception module strengthens the feature expression of key parts of the target and suppresses background interference through convolution, offset operation and sigmoid operation;

[0034] The task perception module dynamically selects and strengthens task-related features through global pooling and fully connected layers.

[0035] Furthermore, the multi-layer perception fusion formula of the DyHead dynamic prediction head is:

[0036] W(F)=π c (π s (π L (F)·F)·F)·F

[0037] Among them, F is a three-dimensional feature tensor of L×S×C, L represents the level of the feature map, s represents the product of the width and height of the feature map, and c represents the number of channels of the feature map;

[0038] π L (·),π s (·) and π c (·) represent the spatial scale perception module, the spatial position perception module and the task perception attention module respectively, and the spatial scale perception module, the spatial position perception module and the task perception attention module are executed in the three dimensions of L, S and C respectively;

[0039] In the above formula,

[0040]

[0041] π C (F)·F=max(α 1 (F)·F c +β 1 (F),α 2 (F)·F c +β 2 (F)

[0042] Where f(·) is a linear function composed of approximately 1×1 convolution operations to achieve feature dimensionality reduction; σ(x) is the Hard-sigmoid activation function; K is the number of sparse sampling positions; p k +Δp k is a self-learning spatial displacement Δp k The movable position determined to focus on some discriminative positions; Δm k is the position p k The importance scalar of self-learning at c is the feature slice of the Cth channel; θ(·)=[α 1 ,β 1 ,α 2 ,β 2 ] is a hyperfunction for learning to control the activation threshold.

[0043] Further, step S2 specifically includes:

[0044] Obtaining a dataset of abnormal behavior images to be detected, and dividing the dataset of abnormal behavior images to be detected into a training set and a test set;

[0045] The training set is input into the abnormal behavior detection model based on improved YOLOv8, and the abnormal behavior detection model based on improved YOLOv8 is trained for multiple rounds of iterations through the SGD optimizer until the abnormal behavior detection model based on improved YOLOv8 converges, thereby obtaining a trained abnormal behavior detection model based on improved YOLOv8.

[0046] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the aforementioned abnormal behavior detection method based on improved YOLOv8 when executing the computer program.

[0047] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the abnormal behavior detection method based on improved YOLOv8 are implemented.

[0048] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0049] The present invention provides an abnormal behavior detection method, device and storage medium based on improved YOLOv8. By improving the YOLOv8 model, the present invention enhances the feature representation and detection accuracy of small targets, improves the detection accuracy and recognition effect of groups in complex scenes, and improves the recognition ability of obscured participants, effectively reducing false detections and missed detections. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 It is a flow chart of the abnormal behavior detection method based on improved YOLOv8 of the present invention;

[0052] Figure 2 It is the working principle diagram of the RCSOSA module in the present invention;

[0053] Figure 3 It is a working principle diagram of the RCS module in the RCSOSA module of the present invention;

[0054] Figure 4 It is a working principle diagram of the RepVGG module in the RCSOSA module of the present invention;

[0055] Figure 5 It is a structural diagram of the DyHead dynamic prediction head in the present invention;

[0056] Figure 6 This is a structural diagram of the FDPN feature focused diffusion pyramid network in the present invention;

[0057] Figure 7 It is a structural diagram of the FocusFeature module in the FDPN feature focusing diffusion pyramid network of the present invention;

[0058] Figure 8 This is the YOLOv8 network structure diagram before improvement;

[0059] Fig. 9 This is the improved YOLOv8 network structure diagram of the present invention;

[0060] Fig.10 This is a training result diagram of the abnormal behavior detection model based on improved YOLOv8 of the present invention;

[0061] Fig.11 It is the confusion matrix of the abnormal behavior detection model based on improved YOLOv8 of the present invention. DETAILED DESCRIPTION

[0062] In order to better understand the purpose, structure and function of the present invention, the technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific preferred embodiments.

[0063] Embodiment 1:

[0064] like Figure 1 As shown, the present invention provides a technical solution:

[0065] An abnormal behavior detection method based on improved YOLOv8 includes the following steps:

[0066] S1. Build an abnormal behavior detection model based on improved YOLOv8;

[0067] The RCSOSA module is introduced to replace the C2f module in the Neck network of the original YOLOv8 model to improve the detection accuracy and recognition effect of dense groups;

[0068] S2. Iteratively train the abnormal behavior detection model based on the improved YOLOv8 to obtain a trained abnormal behavior detection model based on the improved YOLOv8;

[0069] S3, inputting the abnormal behavior image to be detected into the trained abnormal behavior detection model based on improved YOLOv8, and outputting the detection result of the abnormal behavior;

[0070] Among them, Figure 2-4 As shown, the working process of the RCSOSA module in step S1 includes:

[0071] The abnormal behavior image to be detected is used as the input feature map. The input feature map is processed by the RepVGG module to extract and enhance useful information, and its multi-branch structure is used to capture diverse feature representations.

[0072] The processed feature map is divided into two parts in a ratio of 1:1 to obtain the first part and the second part. The first part is directly transmitted, and the second part is optimized for feature selection through multiple RCS modules to enhance the flow of information between layers. The feature fusion between channels is improved through channel shuffling.

[0073] The two parts of the feature map are spliced ​​in the channel dimension, and the spliced ​​feature map is processed again by the RepVGG module to further optimize the feature representation and provide richer and more accurate feature information for the target detection task.

[0074] Specifically, Figure 2 As shown in the figure, the RCSOSA module first performs preliminary processing on the input feature map, and then passes it through a RepVGG module (such as Figure 4 This step takes advantage of the multi-branch structure of RepVGG and allows the network to capture diverse feature representations in different paths, thus laying the foundation for subsequent feature extraction and optimization.

[0075] Next, the feature map is equally divided into two parts through channel splitting, one of which is directly passed to the next stage, while the other is further processed through a series of stacked RCS modules to further optimize feature selection and enhance the information flow between different layers. This design ensures feature reuse and at the same time, through channel shuffling operations (such as Figure 3 The feature fusion between different channels is enhanced.

[0076] After the last RCS module is processed, all sub-feature maps are concatenated in the channel dimension, and channel shuffling is performed to further promote feature fusion. Finally, the concatenated feature maps are processed again through a RepVGG module to refine and optimize feature representation, providing richer and more accurate feature information for subsequent object detection tasks.

[0077] In complex scenes, due to the diverse and dense distribution of people's postures, it is difficult for the original feature fusion module to effectively capture multi-posture and multi-scale student features. To solve this problem, the present invention introduces the RCSOSA module to replace the traditional C2f structure in the neck network, enhance the multi-scale feature fusion capability, and improve the recognition accuracy of complex posture targets. The RCSOSA module first extracts and enhances the input features through RepVGG, and uses its multi-branch structure to capture diversified features. Then, the feature map is divided into two parts, one part is directly transmitted, and the other part is optimized through multiple RCS modules to select features and enhance the flow of information between layers. The channel shuffling operation ensures the reuse and fusion of features. Finally, all sub-feature maps are spliced ​​and processed by RepVGG again to further optimize the feature representation and provide richer and more accurate information for subsequent target detection.

[0078] Embodiment 2:

[0079] Based on Example 1, Figure 6-7 In step S1, the improvement of the original YOLOv8 model also includes:

[0080] Construct a Focusing Diffusion Pyramid Network (FDPN) to replace the Neck network architecture of the original YOLOv8 model to enhance the feature representation and detection accuracy of small targets;

[0081] Its customized feature focus (FocusFeature) module and feature diffusion mechanism work together to ensure that rich contextual information can be incorporated into feature maps of different scales, thus providing strong support for subsequent target detection and classification tasks.

[0082] Further, the feature-focused diffusion pyramid network includes a FocusFeature module, and the FocusFeature module is used to receive an input feature map and generate a processed feature map;

[0083] The feature map processed by the FocusFeature module is assigned to two branches, which are respectively a first branch and a second branch;

[0084] The first branch includes an Unsample module, which performs upsampling processing on the feature map processed by the FocusFeature module to adjust the resolution of the feature map so that it can be spatially aligned with the feature map of a lower level, thereby laying a foundation for subsequent feature fusion;

[0085] The second branch includes a Conv module, and the Conv module further optimizes the quality of the feature map processed by the FocusFeature module by refining the feature representation;

[0086] The feature maps output by the first branch and the second branch are fused with other feature maps through the Concat module.

[0087] Furthermore, the working process of the FocusFeature module includes:

[0088] The FocusFeature module receives three feature maps of different scales, the three feature maps are feature map P3, feature map P4, and feature map P5;

[0089] The feature map P3, the feature map P4, and the feature map P5 are respectively preprocessed differently; the feature map P3 adjusts the spatial size and the number of channels through the ADown downsampling module; the feature map P4 adjusts the number of channels through a 1×1 convolution layer; the feature map P5 adjusts the spatial size through upsampling, and then adjusts the number of channels through a 1×1 convolution layer;

[0090] The preprocessed feature map P3, the feature map P4, and the feature map P5 are spliced ​​in the channel dimension to form a comprehensive feature map containing multi-scale information;

[0091] The comprehensive feature maps are processed separately by a set of parallel depth-wise separable convolutions. The set of depth-wise separable convolutions uses convolution kernels of different sizes to process the comprehensive feature maps. Each depth-wise separable convolution is used to capture features of different scales.

[0092] Superimposing the outputs of all depth-wise separable convolutions with the comprehensive feature map to obtain fused features;

[0093] The fused features are further compressed and refined through 1×1 convolution to output the final feature map.

[0094] like Figure 6 As shown, Figure 6This is the structure of the FDPN module. By observing the network structure diagram of YOLOv8, it can be found that although the feature fusion method retains shallow and deep information, in the feature extraction process, the network will automatically delete insignificant information, resulting in the loss of many important features related to small-sized objects. This is also the main reason for the poor detection effect of small objects. The improved feature fusion method of the present invention focuses on solving this problem. The core advantage of FDPN lies in its customized feature focus (FocusFeature) module and feature diffusion mechanism. The two work together to ensure that rich contextual information can be incorporated into feature maps of different scales, thereby providing strong support for subsequent target detection and classification tasks. The FocusFeature module of the network can receive and process feature maps from three different scales, and through a carefully designed multi-scale parallel convolution structure (such as Figure 7As shown in Figure 2, the effective fusion and extraction of cross-scale features are achieved. First, the three input feature maps of different scales (P3, P4, and P5) are respectively pre-processed in a customized way: P5 adjusts the spatial size by upsampling, and then adjusts the number of channels by 1×1 convolution; P4 directly adjusts the channels by 1×1 convolution; P3 adjusts the spatial size and number of channels by the downsampling module ADown, which can better retain information compared to the traditional convolution with a step size of 2. These pre-processed feature maps are spliced ​​in the channel dimension to form a comprehensive feature map containing rich multi-scale information. Then, the module processes the feature maps separately through a set of parallel depthwise separable convolutions (DWConv), which use convolution kernels of different sizes (5×5, 7×7, 9×9, 11×11), and each convolution focuses on capturing features of different scales. Then, the outputs of all convolutions are superimposed with the original spliced ​​feature maps to achieve comprehensive fusion of cross-scale features. Finally, the features are further compressed and refined through 1×1 convolution to output the final feature map, which not only enhances the network's detection ability for small objects, but also improves its robustness to complex backgrounds and occlusions. In addition, the diffusion mechanism of FDPN further promotes the flow of contextual information between various detection scales, so that the feature map of each scale is rich in detailed semantic information. Specifically, the feature map processed by the FocusFeature module is assigned to two key branches: the upsampling layer and the convolution layer. First, the feature map is upsampled through the Unsample layer, which adjusts the resolution of the feature map so that it can be spatially aligned with the feature map of the lower level, thus laying the foundation for subsequent feature fusion. In addition, the same feature map is also sent to the Conv convolution layer, which further optimizes the quality of the feature by refining the feature representation. Subsequently, the outputs of these two branches - the upsampled and convolutional feature maps - are fused with other feature maps through the Concat layer at multiple levels. This diffusion mechanism ensures the comprehensiveness and continuity of information and improves the detection ability of the model in complex scenes. This design not only improves the detection accuracy of small targets, but also enhances the network's adaptability to occlusion and complex backgrounds.

[0095] Specifically, Fig. 9 As shown, the feature-focused diffusion pyramid network includes a FocusFeature module at the 10th layer, a Conv module at the 11th layer, a Concat module at the 12th layer, a RCSOSA module at the 13th layer, an Unsample module at the 14th layer, a Concat module at the 15th layer, a RCSOSA module at the 16th layer, a FocusFeature module at the 17th layer, a Conv module at the 18th layer, a Concat module at the 19th layer, a RCSOSA module at the 20th layer, an Unsample module at the 21st layer, a Concat module at the 22nd layer, and a RCSOSA module at the 23rd layer;

[0096] The Backbone network of the original YOLOv8 model includes the Conv module of the 0th layer, the Conv module of the 1st layer, the C2f module of the 2nd layer, the Conv module of the 3rd layer, the C2f module of the 4th layer, the Conv module of the 5th layer, the C2f module of the 6th layer, the Conv module of the 7th layer, the C2f module of the 8th layer, and the SPPF module of the 9th layer, which are connected sequentially;

[0097] The Concat module of the 12th layer is connected to the C2f module of the 4th layer, and receives the feature map P5 input by the C2f module of the 4th layer; the Concat module of the 15th layer is connected to the SPPF module of the 9th layer, and receives the feature map P3 input by the SPPF module of the 9th layer;

[0098] The FocusFeature module of the 10th layer is connected to the C2f module of the 4th layer, the C2f module of the 6th layer, and the SPPF module of the 9th layer respectively, the C2f module of the 4th layer is used to input the feature map P5 to the FocusFeature module of the 10th layer, the C2f module of the 6th layer is used to input the feature map P4 to the FocusFeature module of the 10th layer, and the SPPF module of the 9th layer is used to input the feature map P3 to the FocusFeature module of the 10th layer;

[0099] The feature map processed by the FocusFeature module of the 10th layer is assigned to the Conv module of the 11th layer for convolution processing and the Unsample module of the 14th layer for upsampling processing; the Concat module of the 12th layer receives the feature map processed by the Conv module of the 11th layer after convolution, and performs a splicing operation with the feature map P5 input by the C2f module of the 4th layer; the Concat module of the 15th layer receives the feature map processed by the Unsample module of the 14th layer after upsampling, and performs a splicing operation with the feature map P3 input by the SPPF module of the 9th layer;

[0100] The RCSOSA module of the 13th layer receives the feature map spliced ​​by the Concat module of the 12th layer;

[0101] The RCSOSA module of the 16th layer receives the feature map spliced ​​by the Concat module of the 15th layer, and the feature map output by the RCSOSA module of the 16th layer and the feature map output by the Unsample module of the 14th layer are spliced ​​in the Concat module of the 22nd layer;

[0102] The FocusFeature module of the 17th layer receives the feature maps output by the RCSOSA module of the 13th layer, the FocusFeature module of the 10th layer, and the RCSOSA module of the 16th layer respectively;

[0103] The feature map processed by the FocusFeature module of the 17th layer is assigned to the Conv module of the 18th layer for convolution processing and the Unsample module of the 21st layer for upsampling processing; the Concat module of the 19th layer receives the feature map after convolution processing by the Conv module of the 18th layer, and performs a splicing operation with the received feature map output by the Conv module of the 11th layer and the feature map output by the RCSOSA module of the 13th layer; the Concat module of the 22nd layer receives the feature map after upsampling processing by the Unsample module of the 21st layer, and performs a splicing operation with the received feature map output by the Unsample module of the 14th layer and the feature map output by the RCSOSA module of the 16th layer;

[0104] The RCSOSA module of the 20th layer receives the feature map output by the Concat module of the 19th layer, the RCSOSA module of the 23rd layer receives the feature map output by the Concat module of the 22nd layer, and the FocusFeature module of the 17th layer, the RCSOSA module of the 20th layer, and the RCSOSA module of the 23rd layer respectively output feature maps to the detection head of the YOLOv8 model.

[0105] In behavior detection under complex scenes, due to the long distance of the camera, the target is usually small, and the details are easily overlooked, which affects the detection accuracy. To solve this problem, the present invention constructs a feature-focused diffusion pyramid network (FDPN) to improve the feature fusion capability of the model. The core advantage of FDPN lies in its feature focusing (FocusFeature) module and feature diffusion mechanism, which ensures that rich contextual information is integrated into feature maps of different scales to enhance target detection and classification performance. The FocusFeature module effectively captures and integrates cross-scale feature information through parallel deep convolution operations, and the diffusion mechanism promotes the flow of contextual information between scales, so that each feature map contains detailed semantic information, thereby improving the detection accuracy of small targets and enhancing the network's adaptability to occlusion and complex backgrounds.

[0106] Embodiment 3:

[0107] Based on Example 1, Figure 5 In step S1, the improvement of the original YOLOv8 model also includes:

[0108] Use DyHead dynamic prediction head to replace the detection head of the original YOLOv8 model to improve the recognition ability of occluded people;

[0109] The DyHead dynamic prediction head fuses the spatial scale perception module, the spatial position perception module and the task perception module in the output channel of the target.

[0110] Furthermore, the spatial scale perception module generates attention weights through global pooling and 1×1 convolution to enhance the perception of multi-scale objects;

[0111] The spatial position perception module strengthens the feature expression of key parts of the target and suppresses background interference through convolution, offset operation and sigmoid operation;

[0112] The task perception module dynamically selects and strengthens task-related features through global pooling and fully connected layers.

[0113] Furthermore, the multi-layer perception fusion formula of the DyHead dynamic prediction head is:

[0114] W(F)=π c (π s (π L (F)·F)·F)·F

[0115] Among them, F is a three-dimensional feature tensor of L×S×C, L represents the level of the feature map, s represents the product of the width and height of the feature map, and c represents the number of channels of the feature map;

[0116] v L (·),π s (·) and π c (·) represent the spatial scale perception module, the spatial position perception module and the task perception attention module respectively, and the spatial scale perception module, the spatial position perception module and the task perception attention module are executed in the three dimensions of L, S and C respectively;

[0117] In the above formula,

[0118]

[0119] π C (F)·F=max(α 1 (F)·F c +β 1 (F),α 2 (F)·F c +β 2 (F)

[0120] Where f(·) is a linear function consisting of a 1×1 convolution operation to achieve feature dimensionality reduction; σ(x) is the Hard-sigmoid activation function;

[0121] K is the number of sparse sampling positions; p k +Δp k is a self-learning spatial displacement Δp k The movable position determined to focus on some discriminant positions, Δm k is the position p k A self-learning importance scalar at , both of which can be learned through the input features of the intermediate level of F;

[0122] F c is the feature slice of the Cth channel; θ(·)=[α 1 ,β 1 ,α 2 ,β 2 ] is a hyperfunction that learns to control the activation threshold. It first performs global pooling to reduce the dimension in L×S dimension, then uses two fully connected layers, one is a normalization layer, and finally uses a shifted sigmoid function to normalize the output to [-1,1];

[0123] The DyHead dynamic prediction head dynamically adjusts and optimizes feature representation in multiple dimensions such as scale, space, and channel, thereby significantly improving the model's expressive power and its ability to handle complex scenes and multiple targets.

[0124] Specifically, DyHead enriches the perception range of the object detection model by combining the spatial scale module, spatial position perception module and task perception module. Each module plays a vital role in improving the overall performance of the model at different tasks, scales and spatial positions:

[0125] (1) Spatial scale perception module: It aims to enable the model to recognize and adapt to objects of different scales in the image. The module first performs global average pooling (Avg pool) on the input feature map to extract global spatial information. Then, the pooling result is mapped to the channel dimension through 1×1 convolution to further fuse the information. Next, the ReLU activation function is used to introduce nonlinear relationships, so that the model can learn more complex feature relationships. Finally, the Hard sigmoid function is used to generate attention weights for weighting features of different scales. These weights are used to adjust the input feature map, thereby improving the model's perception of objects of various scales.

[0126] (2) Spatial position perception module: It aims to enable the model to focus on the specific spatial position of the target in the image, especially the key parts of the target. The module first performs an Index operation on the feature map to select and focus on the area of ​​interest. Next, the input features are processed through a 3×3 convolution to generate a feature map. These feature maps are then respectively processed through the Offset and Sigmoid operations, where the Offset is used to dynamically adjust the position and shape of the convolution kernel to more accurately capture the spatial features of the target, and the Sigmoid operation generates attention weights for features related to the target. Finally, the processed features are fused to recalibrate the input feature map, thereby strengthening the feature expression of the target area, while suppressing the interference of irrelevant background, enhancing the model's ability to identify and locate the key parts of the target, and ensuring that it accurately focuses on the spatial area of ​​the target.

[0127] (3) Task-aware module: Aims to enable the model to dynamically adjust feature representation according to different task requirements. The module first performs global average pooling (Avg pool) on the feature map to extract global information. Next, the pooled features are processed by a fully connected layer (FC), and nonlinearity is introduced using the ReLU activation function to improve the expressiveness of the features. Subsequently, the features are further transformed by another fully connected layer (FC), and channel attention weights are generated through a normalization operation. These weights are used to identify and strengthen the feature channels that are most important for the current task, and to achieve dynamic selection and weighting of feature channels. Finally, the module adjusts the activation mode of the feature map according to task requirements to optimize the performance of specific tasks and ensure that the model can accurately focus on task-related feature expressions in different task scenarios.

[0128] The feature maps of three different scales obtained from the neck network are processed by the three perception modules mentioned above in turn, and are enhanced and optimized in scale, space and task dimensions. The optimized feature maps are used for target classification and location prediction. Through this design, DyHead can dynamically adjust feature representations in different dimensions to better handle challenges such as occlusion, small target detection and multi-scale target recognition.

[0129] In complex scenes, due to the complex background of the personnel image, small and dense targets, frequent occlusion, and blurred details, higher requirements are placed on the perception ability of the detection algorithm. To this end, the present invention introduces the dynamic head framework DyHead, which dynamically adjusts and optimizes the feature representation by integrating the spatial scale, spatial position and task perception modules, significantly improving the processing ability of the model. The spatial scale perception module generates attention weights through global pooling and 1×1 convolution to enhance the perception of multi-scale targets; the spatial position perception module strengthens the feature expression of key parts of the target through convolution, Offset and Sigmoid operations, and suppresses background interference; the task perception module dynamically selects and strengthens task-related features through global pooling and fully connected layers. Through the synergy of these three modules, DyHead optimizes the performance of feature maps in scale, space and task dimensions, and improves the model's ability to handle small targets, multiple targets and occlusion problems in complex scenes.

[0130] Embodiment 4:

[0131] Based on Example 1, step S2 specifically includes:

[0132] S21, obtaining an abnormal behavior image dataset to be detected, and dividing the abnormal behavior image dataset to be detected into a training set and a test set;

[0133] S22. Input the training set into the abnormal behavior detection model based on improved YOLOv8, and perform multiple rounds of iterative training on the abnormal behavior detection model based on improved YOLOv8 through the SGD optimizer until the abnormal behavior detection model based on improved YOLOv8 converges, thereby obtaining a trained abnormal behavior detection model based on improved YOLOv8.

[0134] Among them, step S21 also includes preprocessing the abnormal behavior image data set to be detected, and expanding the data set by data enhancement, including seven methods: rotation, translation, scaling, cropping, horizontal flipping, contrast enhancement, and Gaussian noise superposition. The processed data set is divided into a training set and a test set in a ratio of 9:1 and put into model training.

[0135] Furthermore, the training set was input into the improved network model for iterative training, and the number of training iterations was set to 300 to ensure that the model fully learned the training data, while combining the patience parameter to prevent overfitting. The batch size was set to 16 to optimize computational efficiency and memory usage, and the input image size was adjusted to 640 to provide sufficient details to enhance detection accuracy. To improve data loading efficiency, the workers parameter was set to 8 to speed up the data preprocessing process. The optimizer selected stochastic gradient descent (SGD) and adjusted the momentum parameter to 0.937 to accelerate convergence and enhance training stability. In addition, the threshold for turning off mosaic enhancement was set to 10 to make the model more adaptable under different training conditions. Finally, the trained behavior detection model was obtained.

[0136] Furthermore, the test set is input into the trained model for testing to obtain the final detection result.

[0137] Furthermore, this abnormal behavior detection method based on improved YOLOv8 can be used for abnormal behavior detection in meetings. The abnormal behavior image dataset includes abnormal behaviors that do not comply with meeting order, such as leaving seats, using mobile phones or other electronic devices, sleeping, whispering, talking or fighting, answering calls, etc.

[0138] Furthermore, the improvements to the original YOLOv8 model include: constructing an improved behavior detection model based on YOLOv8, introducing the RCSOSA module to replace the original C2f module in the neck network of the YOLOv8 model to improve the detection accuracy and recognition effect of dense groups; constructing a new neck network architecture-Focusing Diffusion Pyramid Network, enhancing the feature representation and detection accuracy of small targets; improving the detection head to the dynamic prediction head DyHead (Dynamic Head) to improve the recognition ability of occluded people, and finally obtaining an improved network model to achieve accurate detection of abnormal behaviors of people in complex scenes. By introducing the FDPN structure, this method significantly enhances the feature expression ability of small targets, ensures the accurate capture of subtle movements and features of participants, and improves the recognition effect of dense groups under complex backgrounds. The RCSOSA module further optimizes multi-scale feature fusion, improves feature reuse efficiency and cross-layer information flow, thereby improving the recognition accuracy of the model for student targets in multi-pose and complex scenes. In addition, the dynamic detection head DyHead effectively enhances the model's feature representation capabilities in different dimensions through the synergy of spatial scale perception, position perception and task perception modules, especially for target detection of occluded groups, significantly reducing false detections and missed detections.

[0139] In summary, the present invention has significantly improved the overall performance of abnormal behavior detection through a number of technological innovations, and has achieved breakthroughs in accuracy, robustness, and scene adaptability. It has broad application value and good promotion prospects.

[0140] Embodiment 5:

[0141] The present invention provides a technical solution:

[0142] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the aforementioned abnormal behavior detection method based on improved YOLOv8 when executing the computer program.

[0143] Embodiment 6:

[0144] The present invention provides a technical solution:

[0145] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the abnormal behavior detection method based on improved YOLOv8 are implemented.

[0146] Experiment and result analysis:

[0147] 1 Experimental details

[0148] 1.1 Dataset

[0149] This experiment uses a publicly available behavioral dataset (POCO) to evaluate the effectiveness of the behavior detection method proposed in this paper. The POCO dataset after data enhancement includes 5709 images, which divides the status of people into three types: positive, negative and neutral, covering a variety of behaviors such as attending meetings, reading, recording, chatting, sleeping, playing with mobile phones, etc. This experiment randomly divides the data into training sets and validation sets in a ratio of 9:1, and invests them in model training and testing to evaluate the performance of the entire framework.

[0150] 1.2 Evaluation indicators

[0151] In order to objectively evaluate the performance of the behavior detection model, the evaluation indicators used include precision (P), recall (R), average precision AP, and mean average precision mAP. P, R, and mAP reflect the detection effect of the model. The larger these values ​​are, the better the model detection effect is. The specific calculation formulas are shown in (1) to (4).

[0152]

[0153] Among them, TP is the number of correctly predicted positive samples, TN is the number of correctly predicted negative samples, FP is the number of negative samples classified as positive samples, FN is the number of positive samples classified as negative samples, and N is the number of behavior categories.

[0154] 1.3 Experimental setup

[0155] The hardware platform and environment parameter configurations used in the experimental training phase are shown in Table 1. The number of training iterations is set to 300 to ensure that the model fully learns the training data, and the patience parameter is combined to prevent overfitting. The batch size is set to 16 to optimize computational efficiency and memory usage, and the input image size is adjusted to 640 to provide sufficient details to enhance detection accuracy. To improve data loading efficiency, the workers parameter is set to 8 to speed up the data preprocessing process. The optimizer selects stochastic gradient descent (SGD) and adjusts the momentum parameter to 0.937 to accelerate convergence and enhance training stability. In addition, the threshold for turning off mosaic enhancement is set to 10 to make the model more adaptable under different training conditions.

[0156] Table 1 Training environment and parameter configuration table

[0157]

[0158] 2 FRD-YOLO model evaluation based on POCO dataset

[0159] When training the network model for behavior detection, the Stochastic Gradient Descent (SGD) optimizer is used to optimize the model, and the mosaic data enhancement technology is used in the last 10 training stages. This adjustment is designed to improve the robustness of the model in detecting the behavior of participants. The training results are shown in Figure 2. Fig.10 shown.

[0160] These graphs show the changes in loss and evaluation indicators of the improved model FRD-YOLO during training and validation. Each graph shows how different losses or evaluation indicators change with training Epoch. The following is an analysis of each graph: The first three columns show the localization loss, target loss, and classification loss of the improved model. These curves depict the loss trend, where the X-axis represents the number of training rounds (Epoch) and the Y-axis represents the overall loss value. It can be seen from the curves that as the training progresses, the overall loss value gradually decreases and tends to stabilize. These results show that the proposed FRD-YOLO exhibits good fitting performance, high stability, and accuracy. The last two columns show important evaluation indicators of the model, where the X-axis represents the training time and the Y-axis represents the changes in accuracy, recall, and average precision under different IoU thresholds. These curves show the changes in the performance of evaluating target detection as the number of training rounds increases: the curve value gradually rises and tends to stabilize, and the performance of the model gradually improves and reaches convergence. From Fig.10 It can be seen that the proposed FRD-YOLO is effective.

[0161] Fig.11 The confusion matrix of the improved YOLOv8 model we proposed is shown, showing the model's prediction results for different participant behaviors in the dataset, and explaining the confusion between the behaviors. In the figure, the horizontal axis represents the true category label, the vertical axis represents the predicted category label, and the diagonal elements represent the correct detection rate. The darker the color, the more accurate the model's prediction on the category. It can be observed that the model proposed in this article has achieved good accuracy in each category.

[0162] 3 Ablation experiments

[0163] In order to verify the performance advantages of the new module and improved model proposed in this paper in behavior detection, an ablation study was conducted on FRD-YOLO. As shown in Table 2 below, this paper makes improvements based on the YOLOv8n basic model, improving the model detection accuracy while achieving model lightweight.

[0164] Table 2 Comparison of ablation experiment results

[0165]

[0166] In YOLOv8n, the C2f in the original feature fusion stage is replaced with the RCSOSA module to enhance the multi-scale feature fusion capability and improve the recognition accuracy of complex posture targets. Compared with the original YOLOv8n model, the precision, recall rate, mAP50 and mAP50-95 are increased by 1.4%, 1.0%, 1.5% and 2.7% respectively, which shows that the improved model with the introduction of the RCSOSA module has significantly improved the detection accuracy.

[0167] After replacing the original detection head module of YOLOv8 with the DyHead module, the model's ability to handle small targets, multiple targets and occlusion problems in complex scenes is improved. Compared with the original YOLOv8n model, the precision, recall rate, mAP50 and mAP50-95 are increased by 1.0%, 2.2%, 1.9% and 2.1% respectively. This shows that the addition of the DyHead module effectively improves the performance of the detection head in positioning and classification tasks.

[0168] After introducing the new FDPN network structure in the Neck layer, the feature fusion capability of the model is improved, and the accuracy of small target detection is improved. Compared with the original YOLOv8n model, the recall rate, mAP50 and mAP50-95 are increased by 1.6%, 0.9% and 1.8% respectively. This shows that the introduction of the FDPN structure effectively improves the detection accuracy of the model.

[0169] From the ablation experiment results, the improved four modules effectively improve the accuracy of the entire model and achieve lightweight, indicating that these improved methods have good coupling. Therefore, FRD-YOLO has significantly improved its performance compared to the original YOLOv8n.

[0170] 4 Comparative test

[0171] In order to further verify the accuracy and lightweightness of the model proposed in this paper, 7 representative network models Faster R-CNN, SSD, YOLOv5n, YOLOv5s, YOLOv6s, YOLOv7-tiny, YOLOv8n, and 3 improved models SBD-Net, VWM-YOLOv8, and SimAM-YOLOv8 in the same scenario were selected for control experiments with FRD-YOLO. All models were trained using the same dataset. Precision, recall, mAP50, and mAP50-95 were selected as evaluation indicators of model accuracy. The experimental results are shown in Table 3.

[0172] Compared with the above ten models, the model FRD-YOLO proposed in this paper has the best comprehensive performance in various indicators. Compared with Faster R-CNN, SSD, YOLOv5s, YOLOv6s, YOLOv7-tiny, YOLOv8n, SBD-Net, VWM-YOLOv8, and SimAM-YOLOv8, the accuracy is improved by 16.2%, 20.3%, 4.0%, 1.5%, 6.2%, 12.0%, 1.9%, 1.8%, 0.9%, and 1.7%, respectively, and the recall rate is improved by 14.4%, 11.6%, and 13.8%, respectively. , 2.1%, 6.6%, 11.3%, 4.3%, 2.1%, 2.1% and 4.0%, mAP50 increased by 14.9%, 22.3%, 10.4%, 2.2%, 3.6%, 5.7%, 3.2%, 1.7%, 1.6% and 2.7%, respectively, and mAP50-95 increased by 19.8%, 25.6%, 6.2%, 5.1%, 8.5%, 10.1%, 6.1%, 3.8%, 3.9% and 5.6%, respectively. In summary, the FRD-YOLO model proposed in this paper has a better effect on the behavior detection of participants in complex scenarios and has high practicality.

[0173] Table 3 Performance comparison of different models

[0174]

[0175] 5Test results

[0176] By comparison, it can be found that the use of YOLOv8n model detection in complex scenes will not effectively capture the detailed features of the complex background, and the problem of long-distance small targets being easily missed will occur. The use of FRD-YOLO significantly improves this situation. Therefore, in the case of complex and changeable crowd behavior, the present invention effectively improves the feature extraction and feature fusion capabilities for complex targets, so that the model shows better performance in behavior recognition work in complex scenes.

[0177] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. An abnormal behavior detection method based on improved YOLOv8, characterized in that: The following steps are involved: S1. Build an abnormal behavior detection model based on improved YOLOv8; The RCSOSA module is introduced to replace the C2f module in the Neck network of the original YOLOv8 model to improve the detection accuracy and recognition effect of dense groups; S2. Iteratively train the abnormal behavior detection model based on the improved YOLOv8 to obtain a trained abnormal behavior detection model based on the improved YOLOv8; S3, inputting the abnormal behavior image to be detected into the trained abnormal behavior detection model based on improved YOLOv8, and outputting the detection result of the abnormal behavior; The working process of the RCSOSA module in step S1 includes: The abnormal behavior image to be detected is used as the input feature map, and the input feature map is processed by the RepVGG module to extract and enhance diverse features; The processed feature map is divided into two parts in a ratio of 1:1 to obtain the first part and the second part; the first part is directly transmitted, and the second part is optimized for feature selection and channel shuffling through multiple RCS modules to enhance feature fusion between channels; The two parts of the feature map are spliced ​​in the channel dimension, and the spliced ​​feature map is processed again by the RepVGG module to further optimize the feature representation.

2. The abnormal behavior detection method based on improved YOLOv8 according to claim 1, characterized in that: In step S1, the improvements to the original YOLOv8 model also include: A feature-focused diffusion pyramid network is constructed to replace the Neck network architecture of the original YOLOv8 model to enhance the feature representation and detection accuracy of small targets.

3. The abnormal behavior detection method based on improved YOLOv8 according to claim 2, characterized in that: The feature-focused diffusion pyramid network includes a FocusFeature module, and the FocusFeature module is used to receive an input feature map and generate a processed feature map; The feature map processed by the FocusFeature module is assigned to two branches, which are respectively a first branch and a second branch; The first branch includes an Unsample module, which performs upsampling processing on the feature map processed by the FocusFeature module to adjust the resolution of the feature map so that it can be spatially aligned with the feature map of a lower level, thereby laying a foundation for subsequent feature fusion; The second branch includes a Conv module, and the Conv module further optimizes the quality of the feature map processed by the FocusFeature module by refining the feature representation; The feature maps output by the first branch and the second branch are fused with other feature maps through the Concat module.

4. The abnormal behavior detection method based on improved YOLOv8 according to claim 3, characterized in that: The working process of the FocusFeature module includes: The FocusFeature module receives three feature maps of different scales, the three feature maps are feature map P3, feature map P4, and feature map P5; The feature map P3, the feature map P4, and the feature map P5 are respectively preprocessed differently; the feature map P3 adjusts the spatial size and the number of channels through the ADown downsampling module; the feature map P4 adjusts the number of channels through a 1×1 convolution layer; the feature map P5 adjusts the spatial size through upsampling, and then adjusts the number of channels through a 1×1 convolution layer; The preprocessed feature map P3, the feature map P4, and the feature map P5 are spliced ​​in the channel dimension to form a comprehensive feature map containing multi-scale information; The comprehensive feature maps are processed separately by a set of parallel depth-wise separable convolutions. The set of depth-wise separable convolutions uses convolution kernels of different sizes to process the comprehensive feature maps. Each depth-wise separable convolution is used to capture features of different scales. Superimposing the outputs of all depth-wise separable convolutions with the comprehensive feature map to obtain fused features; The fused features are further compressed and refined through 1×1 convolution to output the final feature map.

5. The abnormal behavior detection method based on improved YOLOv8 according to claim 1, characterized in that: In step S1, the improvements to the original YOLOv8 model also include: The DyHead dynamic prediction head is used to replace the detection head of the original YOLOv8 model to improve the recognition ability of occluded people.

6. The abnormal behavior detection method based on improved YOLOv8 according to claim 5, characterized in that: The DyHead dynamic prediction head includes a spatial scale perception module, a spatial position perception module and a task perception module; The spatial scale perception module generates attention weights through global pooling and 1×1 convolution to enhance the perception of multi-scale objects; The spatial position perception module strengthens the feature expression of key parts of the target and suppresses background interference through convolution, offset operation and sigmoid operation; The task perception module dynamically selects and strengthens task-related features through global pooling and fully connected layers.

7. The abnormal behavior detection method based on improved YOLOv8 according to claim 6, characterized in that: The multi-layer perception fusion formula of the DyHead dynamic prediction head is: W(F)=π c (p s (p L (F)·F)·F)·F Among them, F is a three-dimensional feature tensor of L×S×C, L represents the level of the feature map, s represents the product of the width and height of the feature map, and c represents the number of channels of the feature map; π L (·),π s (·) and π c (·) represent the spatial scale perception module, the spatial position perception module and the task perception attention module respectively, and the spatial scale perception module, the spatial position perception module and the task perception attention module are executed in the three dimensions of L, S and C respectively; In the above formula, p C (F)·F=max(α 1 (F)·F c +b 1 (F),a 2 (F)·F c +b 2 (F)) Where f(·) is a linear function composed of approximately 1×1 convolution operations to achieve feature dimensionality reduction; σ(x) is the Hard-sigmoid activation function; K is the number of sparse sampling positions; p k +Δp k is a self-learning spatial displacement Δp k The movable position determined to focus on some discriminative positions; Δm k is the position p k The importance scalar of self-learning at c is the feature slice of the Cth channel; θ(·)=[α 1 ,β 1 ,α 2 ,β 2 ] is a hyperfunction for learning to control the activation threshold.

8. The abnormal behavior detection method based on improved YOLOv8 according to claim 1, characterized in that: Step S2 specifically includes: Obtaining a dataset of abnormal behavior images to be detected, and dividing the dataset of abnormal behavior images to be detected into a training set and a test set; The training set is input into the abnormal behavior detection model based on improved YOLOv8, and the abnormal behavior detection model based on improved YOLOv8 is trained for multiple rounds of iterations through the SGD optimizer until the abnormal behavior detection model based on improved YOLOv8 converges, thereby obtaining a trained abnormal behavior detection model based on improved YOLOv8.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the abnormal behavior detection method based on improved YOLOv8 described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the abnormal behavior detection method based on improved YOLOv8 according to any one of claims 1 to 8 are implemented.

Images