YOLOv8-PB-based abnormal behavior detection method and system in bid evaluation process

Abstract

The invention provides a YOLOv8-PB-based abnormal behavior detection method and system in a bid evaluation process, and belongs to the technical field of computer vision. The method comprises the following steps: acquiring monitoring image data of a bid evaluation site, and labeling abnormal behaviors in an image; improvement is carried out on the basis of a YOLOv8 model to obtain a YOLOv8-PB model, and the improvement comprises the steps that a second C2f module is replaced with a MobileViTv3 module in a backbone network; a CBAM attention mechanism is introduced into the neck network; replacing a bounding box regression loss function from CIoULoss to MPDIOULoss, wherein the MPDIOULoss is used as a regression loss function of the bounding box; using the training set to train the YOLOv8-PB model, and using the training set to train the YOLOv8-PB model; inputting bid evaluation field images collected in real time into the trained YOLOv8-PB model, detecting whether abnormal behaviors exist or not, and triggering an alarm when the abnormal behaviors are detected. According to the invention, the abnormal behavior detection of participants in the bid evaluation process is realized, and the supervision of the bid evaluation process is enhanced.

Description

TECHNICAL FIELD

[0001] The application relates to a YOLOv8-PB-based bid evaluation process abnormal behavior detection method and system, and belongs to the technical field of computer vision. BACKGROUND

[0002] In the bid evaluation process, timely discovery and identification of abnormal behavior are of great significance to maintaining a fair, just and open bid evaluation environment, and the behavior of bid evaluation experts needs to be monitored. At present, whether the behavior of bid evaluation experts is abnormal is usually determined manually by bid evaluation supervisors, so that corresponding processing measures can be taken when the behavior of bid evaluation experts is abnormal. This method of manually identifying bid evaluation abnormalities has high labor costs and low identification efficiency. Existing abnormal behavior detection methods face various challenges, such as target multi-scale changes, complex background interference and small target detection difficulties. YOLOv8, as the latest version of the YOLO series model, has shown great potential and advantages in the field of target detection. However, in order to better adapt to the special needs of the bid evaluation process, targeted improvement and optimization are still needed. SUMMARY

[0003] The application aims to provide a YOLOv8-PB-based bid evaluation process abnormal behavior detection method and system, which realizes abnormal behavior detection of participants in the bid evaluation process, strengthens the supervision of the bid evaluation process, and effectively assists the fairness of the bid evaluation activity.

[0004] To achieve the above-mentioned purpose, the following technical solutions are implemented: A YOLOv8-PB-based bid evaluation process abnormal behavior detection method, comprising the following steps: Obtain monitoring image data of the bid evaluation site, and label abnormal behavior in the image to construct a training set, a validation set and a test set; Improve the YOLOv8 model to obtain a YOLOv8-PB model, the improvement including: replacing the second C2f module with a MobileViTv3 module in the backbone network; introducing a CBAM attention mechanism in the neck network; replacing the CIoULoss with the MPDIoULoss as the bounding box regression loss function; Train the YOLOv8-PB model using the training set, verify it using the validation set, calculate the loss using the MPDIoULoss, and update the model parameters; Input the real-time collected bid evaluation site image into the trained YOLOv8-PB model to detect whether there is abnormal behavior, and trigger an alarm when abnormal behavior is detected.

[0005] Preferably, the YOLOv8-PB model feature map processing mode is as follows: the input feature map sequentially passes through the first CBS module, the second CBS module, the first c2f module, the third CBS module, the MobileViTv3 module, the fourth CBS module, the second c2f module, the fifth CBS module, the third c2f module and the SPPF module of the backbone network to obtain a feature map ; The feature map is up-sampled on the neck network and spliced with the feature map output by the second c2f module of the backbone network , which is processed by the first c2f module of the neck network to obtain a feature map , the feature map is up-sampled and spliced with the feature map output by the MobileViTv3 module of the backbone network , which is processed by the second c2f module of the neck network to obtain a feature map , the feature map is up-sampled and spliced with the feature map output by the first c2f module of the backbone network , which is processed by the third c2f module of the neck network to obtain a feature map ; the feature map is output to the detection head through the CBAM attention mechanism, the feature map is processed by the first CBS module of the neck network and spliced with the feature map , and then a feature map is obtained through the fourth c2f module of the neck network , the feature map is output to the detection head through the CBAM attention mechanism, the feature map is processed by the second CBS module of the neck network and spliced with the feature map , and then a feature map is obtained through the fifth c2f module of the neck network , the feature map is output to the detection head through the CBAM attention mechanism, the feature map is processed by the third CBS module of the neck network and spliced with the feature map , and then sequentially passes through the sixth c2f module of the neck network and the CBAM attention mechanism to be output to the detection head.

[0006] Preferably, the MobileViTv3 module extracts local features from the input feature through 1x1 convolution and depth convolution, expands the local features into non-overlapping patches, and encodes the relationship between the patches using a Transformer to obtain a feature , the feature After folding, the data is projected onto the C-dimensional space through a 1×1 convolution, and then combined with the input features through a cascade operation. Finally, local and global features are fused through a depthwise convolution.

[0007] Preferably, the CBAM attention mechanism sequentially weights the feature map in the channel dimension and the spatial dimension through the channel attention module and the spatial attention module.

[0008] Preferably, the channel attention module is specifically as follows: The input feature map is processed by global max pooling and global average pooling in the width and height dimensions, respectively, to obtain two feature maps of size 1×1×C. The two feature maps are input into a multilayer perceptron to perform a dimensionality reduction and then dimensionality increase operation. The two feature maps are added together and then passed through the Sigmoid activation function to obtain the feature vectors on the channels.

[0009] Preferably, the spatial attention module is as follows: The output feature map of the channel attention module is subjected to global max pooling and global average pooling operations in the channel dimension to obtain two feature vectors of size H×W×1. The two feature vectors are concatenated along the channel dimension. The concatenated feature vector is then subjected to a 7×7 convolution to reduce the number of channels in the feature vector. The number of channels after dimensionality reduction is 1. After passing through the Sigmoid activation function, the feature vector of the spatial attention module is obtained.

[0010] Preferably, the formula for MPDIoULoss is: , , in, For MPDIoULoss, For true bounding boxes; For prediction boxes; , Input the width and height of the image from the network, respectively; The Euclidean distance between the top-left corners of the ground truth bounding box and the predicted bounding box; The Euclidean distance between the bottom right corners of the ground truth bounding box and the predicted bounding box; This is the intersection-union ratio of the areas of the ground truth bounding boxes and the predicted bounding boxes. , These are the coordinates of the top left and bottom right corners of the prediction box, respectively. , These are the coordinates of the top left and bottom right corners of the actual bounding box, respectively.

[0011] A YOLOv8-PB-based system for detecting abnormal behavior in the bidding process, comprising: The data acquisition module is used to acquire on-site image or video stream data; The model training module is used to build and train the YOLOv8-PB model. The abnormal behavior analysis module is used to detect and classify abnormal behaviors in input images using a trained YOLOv8-PB model. The data storage module is used to store training data, model parameters, and detection results; The information push module is used to generate and send alarm information when abnormal behavior is detected.

[0012] The advantages of this invention are as follows: The second C2f module is replaced with the MobileViTv3 module in the Backbone section, improving the model's feature extraction capability. The Transformer's attention mechanism captures image feature information globally, enabling the model to better understand the entire image scene in varied evaluation scenarios and provide richer contextual information. It also better captures long-range dependencies and contextual information in the image, helping the model identify and distinguish subtle actions and abnormal behaviors of evaluation experts. Furthermore, the MobileViTv3 module provides stronger feature extraction capabilities, making the model more robust when dealing with complex backgrounds, lighting changes, and occlusion, thereby improving the accuracy and reliability of the monitoring system.

[0013] Introducing the CBAM attention mechanism enhances the model's ability to extract information from small targets. In the Neck stage, CBAM strengthens the fusion of multi-scale features along top-down and bottom-up paths, improving the network's adaptability to multi-scale targets. This is crucial for monitoring abnormal behaviors of different sizes that may occur during the evaluation process. Focusing on global contextual information strengthens the network's ability to identify target boundaries and category attributes. In monitoring abnormal behaviors during evaluation, this means the model can more accurately distinguish between normal and abnormal expert behavior, improving monitoring accuracy.

[0014] The MPDIoU loss function was improved, the bounding box prediction accuracy was optimized, the calculation process was simplified, the model convergence was stabilized, the model's detection accuracy for small targets of evaluation personnel was improved, and the abnormal behavior of evaluation experts was accurately located; the risk of overfitting was reduced, and the generalization ability of the model in practical applications was improved. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0017] Figure 2 This is a schematic diagram of the YOLOv8-PB model structure of the present invention.

[0018] Figure 3 This is a schematic diagram of the MobileViT module structure.

[0019] Figure 4 This is a schematic diagram of the CBAM attention mechanism.

[0020] Figure 5 This is a schematic diagram of the channel attention structure.

[0021] Figure 6 This is a schematic diagram of the spatial attention module.

[0022] Figure 7 This is a schematic diagram of the system structure.

[0023] Figure 8 This is a schematic diagram illustrating the changes in model loss. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1 like Figure 1 As shown, a method for detecting abnormal behavior in the evaluation process based on YOLOv8-PB is proposed. By integrating the MobileViTv3 module into the YOLOv8 backbone network and replacing the original C2f module, the model's ability to extract features is effectively improved, enhancing its ability to identify abnormal behavior in the evaluation process. The CBAM attention mechanism is introduced into the Neck network to improve the model's ability to detect small targets by focusing on features in both channel and spatial dimensions, capturing detailed behaviors in the evaluation process. The loss function is improved by replacing CIoULoss with MPDIoULoss, which more accurately optimizes the prediction of bounding boxes and improves the model's detection accuracy.

[0026] Specifically, it includes the following steps: S1: Obtain monitoring image data of the bidding site, label abnormal behaviors in the images, and construct training set, validation set and test set.

[0027] S2: Based on the YOLOv8 model, an improvement was made to obtain the YOLOv8-PB model. The improvements include: replacing the second C2f module in the backbone network with the MobileViTv3 module; introducing the CBAM attention mechanism in the neck network; and replacing the bounding box regression loss function from CIoULoss to MPDIoULoss.

[0028] S3: Train the YOLOv8-PB model using the training set, validate it using the validation set, calculate the loss using MPDIoULoss, and update the model parameters.

[0029] S4: Input the real-time collected images of the evaluation site into the trained YOLOv8-PB model to detect any abnormal behavior and trigger an alarm when abnormal behavior is detected.

[0030] As a refinement of the above embodiment, in step S1, high-definition cameras are deployed at the bid evaluation site. A total of 242 monitoring points are arranged in the bid evaluation room, tea room, rest area, corridor, dormitory building, and perimeter of the base to achieve blind-spot-free monitoring. Each bid evaluation room is equipped with three monitoring points: a fixed dome camera, a pan-tilt dome camera, and a bullet camera, to achieve blind-spot-free monitoring. The fixed dome camera and bullet camera monitoring points implement a continuous monitoring and recording strategy, while the pan-tilt dome camera implements a motion detection strategy to collect real-time images of the scene. An abnormal behavior risk identification database is established, including abnormal situations such as ① bid evaluation experts using mobile phones; ② bid evaluation experts sleeping; ③ bid evaluation experts leaving their posts; ④ bid evaluation experts bringing in papers, etc.

[0031] Using the labelimg software, labels were generated for the image samples. Images of evaluation experts using their phones were labeled "phone," sleeping, leaving their posts, or carrying papers were labeled "paper," etc. These labels were then saved in TXT format. Based on the unusual behavior of the evaluation personnel, the images were categorized into four types and labeled accordingly. In this process, 60% of the labeled images were used as training samples, 20% as validation samples, and 20% as test samples.

[0032] As a refinement of the above embodiment, in step S2, YOLOv8 is one of the latest object detection algorithms. When applying it to the detection of abnormal personnel behavior in a bidding evaluation scenario, this paper selects the minimum weight model of the YOLOv8n algorithm to meet real-time requirements. However, in actual detection, it was found that the YOLOv8n algorithm has some problems. In special scenarios such as bidding evaluation rooms, there may be challenges such as complex lighting conditions, occlusion, and targets of different scales, requiring more specialized model structures or training strategies; for the detection of abnormal personnel behavior, the model is required to have high accuracy. However, the YOLOv8n algorithm does not perform ideally in handling the detection of abnormal personnel behavior in bidding evaluation scenarios, and is prone to false positives and false negatives.

[0033] To improve the detection performance of abnormal behavior by personnel in bidding evaluation scenarios, this invention proposes the YOLOv8-PB algorithm, with the network structure shown in Figure 2. In the backbone, the C2f module is replaced with the MobileViTv3 module to enhance the feature extraction capability of the abnormal behavior detection model. The second C2f module in the backbone network is replaced with the MobileViTv3 module, which employs the Transformer's attention mechanism to capture image feature information globally. In the neck part, the CBAM attention mechanism is introduced to enhance the ability to extract information from small targets. The CBAM attention mechanism is embedded before the head in the network, making the network focus more on targets of interest. The loss function is improved by optimizing the CIoU loss function to the MPDIoU loss function, simplifying the calculation process, stabilizing model convergence, improving the model's detection accuracy for small targets of bidding evaluation personnel, and optimizing the prediction accuracy of bounding boxes.

[0034] The YOLOv8-PB model feature map processing method is as follows: the input feature map is sequentially processed through the first CBS module, the second CBS module, the first c2f module, the third CBS module, the MobileViTv3 module, the fourth CBS module, the second c2f module, the fifth CBS module, the third c2f module, and the SPPF module of the backbone network to obtain the feature map. .

[0035] The feature map Feature maps sampled from the neck network and output by the second C2F module of the backbone network. The feature map is obtained after concatenation and processing by the first C2F module of the neck network. The feature map After upsampling, the feature maps are compared with those output by the backbone network MobileViTv3 module. The concatenation is processed by the second C2F module of the neck network to obtain the feature map. The feature map After upsampling, the feature map is compared with the output feature map of the first C2F module of the backbone network. The feature maps are obtained after concatenation and processing by the third C2F module of the neck network. The feature map The feature map is output to the detection head via the CBAM attention mechanism. After processing by the first CBS module of the neck network and compared with the feature map The feature maps are then concatenated and passed through the fourth C2F module of the neck network. The feature map The feature map is output to the detection head via the CBAM attention mechanism. After processing by the second CBS module of the neck network and compared with the feature map The feature maps are then concatenated and passed through the fifth C2F module of the neck network. The feature map The feature map is output to the detection head via the CBAM attention mechanism. After processing by the third CBS module of the neck network and compared with the feature map The data is then spliced ​​together and sequentially passed through the sixth C2F module of the neck network and the CBAM attention mechanism before being output to the detection head.

[0036] As a refinement of the above embodiments, such as Figure 3 As shown, the MobileViTv3 module can model local and global information in the input tensor using fewer parameters. First, local features are extracted from the input features using 1×1 convolutions and depthwise convolutions. These local features are then unfolded into non-overlapping patches, and the relationships between the patches are encoded using a Transformer to obtain the features. Unlike ViT, which loses pixel spatial order, MobileViT retains neither the patch order nor the spatial order of pixels within each patch, thus preserving the features. After folding, the data is projected onto the C-dimensional space through a 1×1 convolution, and then combined with the input features through a cascade operation. Finally, local and global features are fused through a depthwise convolution.

[0037] Because convolution is used to encode local information in an n×n region, and The global information at the p-th position in P patches is encoded, therefore It can perceive global information in the input feature map. Therefore, we consider further improving the MobileViT module, utilizing its modeling of local and global information, to further improve the model's detection performance for small targets such as whether or not insulated gloves are worn, while keeping the number of parameters low.

[0038] In this application, the second C2f module in the backbone network is replaced with the MobileViTv3 module. The MobileViTv3 module first extracts local information through convolutional blocks, enabling the model to better recognize the contours, textures, and shapes of objects, which helps distinguish the target from the background or other objects. Simultaneously, the MobileViTv3 module uses the self-attention mechanism in Transformer to perform global correlation calculation on the input feature map. This captures long-distance dependencies and contextual information, helping the model to have a global understanding of the visual scene and providing more accurate feature representations. This allows the model to more accurately understand the position and features of objects in the image, even in low light or occluded conditions, further enhancing the feature expressiveness and improving the model's detection accuracy. The challenge of detecting small targets, such as those wearing insulated gloves, lies in the small area of ​​the hand, making detection difficult. Therefore, the information and features that can be extracted from the image are very limited, while the hand hides a lot of correlation information with the positions of people and objects such as control levers. Traditional object detection methods have limited receptive fields and can only acquire local contextual information. To solve this problem, Transformer was introduced into the field of computer vision. MobileViT combines the features of adaptive input weighting and global processing in CNN and ViT. It uses Transformer as a convolution to learn global information, effectively encoding local and global information, and to some extent solving the problems of high computational cost and complexity of Transformer-based detection models.

[0039] As a refinement of the above embodiments, such as Figure 4 As shown, the CBAM attention mechanism can enhance the expressive power of features for targets. Adding it to YOLOv8 utilizes global information guidance while preserving local details, which can effectively improve the detection performance of the YOLOv8 model.

[0040] Specifically, such as Figure 5 As shown, the CBAM attention mechanism sequentially weights the feature map in both the channel and spatial dimensions through a channel attention module and a spatial attention module. First, the output feature map is passed through the channel attention module to generate a set of channel weights. Then, the input feature map is multiplied by these weights to obtain the input matrix for the spatial attention mechanism. Second, the input matrix is ​​passed through the spatial attention module, and the different weights for the spatial features are assigned in a similar manner to obtain the output feature map of the entire CBAM attention mechanism.

[0041] (1) The channel attention module is as follows: like Figure 6 As shown, the input feature map Two feature maps of size 1×1×C are obtained by using global max pooling and global average pooling in the width and height dimensions, respectively.

[0042] The two feature maps are input into a multilayer perceptron to perform dimensionality reduction and then dimensionality increase.

[0043] The two feature maps are added together, and then the result is obtained by passing the Sigmoid activation function to the channel. .

[0044] The specific formula is as follows: , in, For the Sigmoid activation function, Feature map The output feature map is obtained through global average pooling. For feature map The output feature map after global max pooling, and These represent the dimensionality reduction and dimensionality increase operations in MLP, respectively. After dimensionality reduction, the feature map undergoes a ReLU activation function followed by dimensionality increase. The feature map obtained after dimensionality reduction has a dimension of C / r×C, and the feature map obtained after dimensionality increase has a dimension of C×C / r, where r is a decay factor. After the weights for channel attention are generated, they are multiplied with the input feature map to obtain the output features of the channel attention module.

[0045] (2) The spatial attention module is as follows: like Figure 7 As shown, the output feature map of the channel attention module Global max pooling and global average pooling operations are performed on the channel dimension to obtain two feature vectors of size H×W×1.

[0046] The two feature vectors are concatenated along the channel dimension. The concatenated feature vector is then subjected to a 7×7 convolution to reduce the number of channels in the feature vector, resulting in a channel count of 1.

[0047] After passing through the Sigmoid activation function, the feature vector of the spatial attention module is obtained. .

[0048] The specific formula is as follows: , in, For feature map The output feature map is obtained through global average pooling. For feature map The output feature map after global max pooling.

[0049] (3) Finally, The output features of the CBAM attention mechanism are obtained by multiplying the input feature map of the spatial attention module.

[0050] As a refinement of the above embodiments, the formula for MPDIoULoss is: , , in, For MPDIoULoss, For true bounding boxes; For prediction boxes; , Input the width and height of the image from the network, respectively; The Euclidean distance between the top-left corners of the ground truth bounding box and the predicted bounding box; The Euclidean distance between the bottom right corners of the ground truth bounding box and the predicted bounding box; This is the intersection-union ratio of the areas of the ground truth bounding boxes and the predicted bounding boxes. , These are the coordinates of the top left and bottom right corners of the prediction box, respectively. , These are the coordinates of the top left and bottom right corners of the actual bounding box, respectively.

[0051] The bounding box regression loss function in the YOLOv8 network is CIoU. In small object detection, CIoU for bounding box prediction sometimes fails. This is because if the center of the predicted box coincides with the center of the ground truth box, CIoU, while optimizing for shape, size, and positional deviations between the predicted and ground truth boxes, often results in different aspect ratios for the predicted box compared to the ground truth box. More importantly, CIoU has uncertainties in predicting the aspect ratio of the bounding box, all of which degrade the speed and accuracy of bounding box regression.

[0052] This invention introduces the MPDIoU loss function, replacing the CIoU loss function in the original network. MPDIoU comprises three key factors: overlapping or non-overlapping regions, center point distance, and width and height deviations. It calculates IoU by minimizing the point distance between the predicted bounding box and the ground truth bounding box, thereby simplifying the calculation process and solving the aforementioned problems.

[0053] As a refinement of the above embodiment, in step S3, the constructed dataset of abnormal evaluation behaviors is input into the YOLOv8-PB model for training. Through a feature fusion module, the addition of an attention mechanism, and optimization of the loss function, the detection accuracy and generalization ability of the YOLOv8 model are improved. The specific process is as follows: (1) Initialize parameters: Randomly initialize model parameters or load pre-trained weights.

[0054] (2) Forward propagation: Input the processed data and perform forward propagation through the YOLOv8-PB network to calculate the prediction results.

[0055] (3) Loss calculation: Based on the prediction results and the true labels, calculate the MPDIoU loss and evaluate the model performance.

[0056] (4) Backpropagation: Based on the loss function, backpropagate the model parameters and calculate the gradient.

[0057] (5) Parameter update: The model parameters are updated based on the gradient using the optimizer SGD.

[0058] In this specific embodiment, the parameter settings for training the YOLOv8-PB model are as follows: initial learning rate of 0.01; batch size of 16; number of iterations of 300; and model weights saved every 10 steps. The average loss of the model changes with the number of training batches as follows: Figure 8 As shown, with the increase of training batches, the average loss of the model continuously decreases, from around 32 to around 3, and the entire network tends to converge. The changes of each loss are shown in the figure. With the increase of training batches, the values ​​of the model's cls, regression loss (bbox), and obj continuously decrease, and the entire network tends to stabilize.

[0059] Model Evaluation and Optimization Evaluate the performance of the YOLOv8-PB model on the validation set, monitoring metrics such as accuracy, precision, and recall; adjust hyperparameters such as learning rate and batch size based on the evaluation results to optimize model performance; use early stopping to prevent overfitting, stopping training when performance on the validation set no longer improves.

[0060] Experiments were conducted to compare the detection accuracy of different algorithms, leading to the optimal algorithm selection. The detection results were output in a visual format to monitor for any abnormal behavior by evaluation personnel. If any abnormal behavior was detected, an alarm mechanism was immediately triggered to notify relevant personnel for handling. Simultaneously, the detection results and alarm information were stored in a database for subsequent analysis.

[0061] Experimental results: The YOLOv8-PB model of this invention was compared with the YOLOv8n, YOLOv7, and YOLOv5 models in detecting abnormal behavior in the evaluation process. As shown in Table 1, the results showed that: compared with the YOLOv7 model, the accuracy of the YOLOv8-PB model was improved by 2.2%, and the recall, mAP@0.5, and mAP@0.95 were improved by 12.1%, 9.1%, and 13.9%, respectively; compared with the YOLOv5s and YOLOv8n models, the accuracy of the YOLOv8-PB model was not significantly different, but the recall, mAP@0.5, and mAP@0.95 were improved by 4.4%, 3.0%, and 6.4% compared with YOLOv5s, and by 1.8%, 1.9%, and 1.8% compared with YOLOv8n.

[0062] Table 1 Performance Comparison of YOLOv8-PB and YOLO Series Models To verify the effectiveness of the improved modules of the algorithm in this study, the original model YOL0v8n was used as the baseline model, and ablation experiments were conducted using precision, recall, mAP, and FPS as evaluation indicators. Different combinations of the improved modules were used, and the results are shown in Table 2. The YOLOv8-PB model of this invention achieved a mean precision of 92.1% for evaluating anomalous behavior, which is 2.7 percentage points higher than YOLOv8, demonstrating a significant improvement in detection performance. It also improved recall by 3.3 percentage points and precision by 3.8 percentage points, reducing false negatives and false positives.

[0063] Table 1 Ablation Test Example 2 like Figure 7 As shown, an abnormal behavior detection system for the bidding process based on YOLOv8-PB includes: The data acquisition module is used to acquire on-site image or video stream data.

[0064] The model training module is used to build and train the YOLOv8-PB model.

[0065] The abnormal behavior analysis module is used to detect and classify abnormal behaviors in input images using a trained YOLOv8-PB model; it identifies and classifies abnormal behaviors during the evaluation process, such as playing on a mobile phone, sleeping, leaving one's post, and carrying paper.

[0066] The data storage module is responsible for storing the collected raw data, intermediate results during training, and final detection results; managing data backup, recovery, and access permissions to ensure data security and integrity.

[0067] The information push module is responsible for pushing the analysis results of the abnormal behavior analysis module to relevant personnel or systems in real time or periodically; providing alerts and reports on abnormal behavior to help managers respond to and handle abnormal situations in a timely manner.

[0068] Example 3 This disclosure also provides a YOLOv8-PB-based device for detecting abnormal behavior in the bidding process, including a processor and a memory. Optionally, the device may further include a communication interface and a bus. The processor, communication interface, and memory can communicate with each other via the bus. The communication interface can be used for information transmission. The processor can call logical instructions in the memory to execute the YOLOv8-PB-based method for detecting abnormal behavior in the bidding process described above.

[0069] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0070] Memory, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as the program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor executes the program instructions / modules stored in the memory to perform functional applications and data processing, thereby implementing the YOLOv8-PB-based abnormal behavior detection method for the bidding process described in the above embodiments.

[0071] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory may include high-speed random access memory and may also include non-volatile memory.

[0072] This disclosure provides a computer-readable storage medium storing computer-executable instructions configured to execute the above-described YOLOv8-PB-based method for detecting abnormal behavior in the bidding process.

[0073] The aforementioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0074] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code. It can also be a transient storage medium.

[0075] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting abnormal behavior in a bidding process based on YOLOv8-PB, characterized in that, Includes the following steps: Acquire monitoring image data of the bidding evaluation site, label abnormal behaviors in the images, and construct training, validation and test sets; Based on the YOLOv8 model, the YOLOv8-PB model is improved. The improvements include replacing the second C2f module with the MobileViTv3 module in the backbone network. Introduce the CBAM attention mechanism into the neck network; replace the bounding box regression loss function with MPDIoULoss instead of CIoULoss. The YOLOv8-PB model is trained using the training set, validated using the validation set, and the loss is calculated using MPDIoULoss to update the model parameters. The real-time images of the bidding evaluation site are input into the trained YOLOv8-PB model to detect any abnormal behavior and trigger an alarm when abnormal behavior is detected.

2. The method for detecting abnormal behavior in the bidding process based on YOLOv8-PB according to claim 1, characterized in that, The YOLOv8-PB model feature map processing method is as follows: the input feature map is sequentially processed through the first CBS module, the second CBS module, the first c2f module, the third CBS module, the MobileViTv3 module, the fourth CBS module, the second c2f module, the fifth CBS module, the third c2f module, and the SPPF module of the backbone network to obtain the feature map. ; The feature map Feature maps sampled from the neck network and output by the second C2F module of the backbone network. The concatenation is processed by the first C2F module of the neck network to obtain the feature map. The feature map After upsampling, the feature maps are compared with those output by the backbone network MobileViTv3 module. The concatenation is processed by the second C2F module of the neck network to obtain the feature map. The feature map After upsampling, the feature map is compared with the output feature map of the first C2F module of the backbone network. The concatenation is processed by the third C2F module of the neck network to obtain the feature map. The feature map The feature map is output to the detection head via the CBAM attention mechanism. After processing by the first CBS module of the neck network and compared with the feature map The feature maps are then obtained by concatenating the data and passing it through the fourth C2F module of the neck network. The feature map The feature map is output to the detection head via the CBAM attention mechanism. After processing by the second CBS module of the neck network and compared with the feature map The feature maps are then concatenated and passed through the fifth C2F module of the neck network. The feature map The feature map is output to the detection head via the CBAM attention mechanism. After processing by the third CBS module of the neck network and compared with the feature map The data is then spliced ​​together and sequentially passed through the sixth C2F module of the neck network and the CBAM attention mechanism before being output to the detection head.

3. The method for detecting abnormal behavior in the bidding process based on YOLOv8-PB according to claim 1, characterized in that, The MobileViTv3 module extracts local features from the input features through 1×1 convolution and depthwise convolution, unfolds the local features into non-overlapping patches, and uses a Transformer to encode the relationships between the patches to obtain the features. , will feature After folding, the data is projected onto the C-dimensional space through a 1×1 convolution, and then combined with the input features through a cascade operation. Finally, local and global features are fused through a depthwise convolution.

4. The method for detecting abnormal behavior in the bidding process based on YOLOv8-PB according to claim 1, characterized in that, The CBAM attention mechanism sequentially weights the feature map in both the channel and spatial dimensions through the channel attention module and the spatial attention module.

5. The method for detecting abnormal behavior in the bidding process based on YOLOv8-PB according to claim 4, characterized in that, The channel attention module is specifically as follows: The input feature map is processed by global max pooling and global average pooling in the width and height dimensions, respectively, to obtain two feature maps of size 1×1×C. The two feature maps are input into a multilayer perceptron to perform a dimensionality reduction and then dimensionality increase operation. The two feature maps are added together and then passed through the Sigmoid activation function to obtain the feature vectors on the channels.

6. The method for detecting abnormal behavior in the bidding process based on YOLOv8-PB according to claim 4, characterized in that, The spatial attention module is specifically as follows: The output feature map of the channel attention module is subjected to global max pooling and global average pooling operations in the channel dimension to obtain two feature vectors of size H×W×1. The two feature vectors are concatenated along the channel dimension. The concatenated feature vector is then subjected to a 7×7 convolution to reduce the number of channels in the feature vector. The number of channels after dimensionality reduction is 1. After passing through the Sigmoid activation function, the feature vector of the spatial attention module is obtained.

7. The method for detecting abnormal behavior in the bidding process based on YOLOv8-PB according to any one of claims 1-6, characterized in that, The formula for MPDIoULoss is: ， ， in, For MPDIoULoss, For true bounding boxes; For prediction boxes; , The width and height of the image input from the network are respectively. The Euclidean distance between the top-left corners of the ground truth bounding box and the predicted bounding box; The Euclidean distance between the bottom right corners of the ground truth bounding box and the predicted bounding box; This is the intersection-union ratio of the areas of the ground truth bounding boxes and the predicted bounding boxes. , These are the coordinates of the top left and bottom right corners of the prediction box, respectively. , These are the coordinates of the top left and bottom right corners of the actual bounding box, respectively.

8. A system for detecting abnormal behavior in a bidding process based on YOLOv8-PB, characterized in that, include: The data acquisition module is used to acquire on-site image or video stream data; A model training module for constructing and training the YOLOv8-PB model as described in any one of claims 1-7; The abnormal behavior analysis module is used to detect and classify abnormal behaviors in input images using a trained YOLOv8-PB model. The data storage module is used to store training data, model parameters, and detection results; The information push module is used to generate and send alarm information when abnormal behavior is detected.

9. A device for detecting abnormal behavior in a bidding process based on YOLOv8-PB, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute, when running the program instructions, the YOLOv8-PB-based method for detecting abnormal behavior in the bidding process as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the YOLOv8-PB-based method for detecting abnormal behavior in the bidding process as described in any one of claims 1-7.

Images