Driver abnormal driving behavior detection method based on YOLOv11 improvement

Abstract

The invention discloses a driver abnormal driving behavior detection method based on YOLOv11 improvement, and the method comprises the steps: S1, building a data set: collecting a driver abnormal driving behavior data set, carrying out the preprocessing of the data set, and processing an exposure or low-brightness picture in the data set through employing an image brightness adaptive adjustment mechanism; dividing a data set and marking the data set; s2, a model is constructed and trained, wherein a GSConv module and a Dysample dynamic upsampling technology are introduced into a YOLOv11 model to construct an improved YOLOv11 driver abnormal driving behavior detection model; training is carried out, and a final driver abnormal driving behavior detection model weight is obtained after training is completed; and S3, testing a result and evaluating: testing an input picture by using the weight of the detection model obtained by training to obtain an abnormal driving behavior attribute recognition result of the driver, and then evaluating the performance of the model. The method has higher accuracy, timeliness and anti-interference capability.

Description

Technical Field

[0001] This invention belongs to the field of computer vision behavior detection technology, specifically relating to a method for detecting abnormal driving behavior of drivers based on YOLOv11. Background Technology

[0002] With the rapid development of Chinese society and the continuous improvement of people's living standards, the demand for automobiles is also constantly increasing. Detecting abnormal driving behavior has become a core research direction for ensuring road traffic safety. Statistics show that traffic accidents caused by abnormal driver behavior (such as yawning, drinking water, smoking, and making phone calls) account for more than 70% of all road accidents, seriously threatening public safety and property.

[0003] Traditional methods for detecting abnormal driving behavior primarily rely on sensors, analyzing the driver's physical and physiological characteristics to determine the presence of abnormal driving behavior. Some researchers have installed sensors on the steering wheel to collect and analyze real-time data on the driver's steering wheel rotation angles. Based on changes in this parameter, they determine if the driver is exhibiting abnormal driving behavior leading to decreased operational stability. Other researchers have used smartwatches to acquire the driver's physiological information and movement status, detecting fatigue driving through heart rate variability and yawning patterns. While these sensor-based methods offer good timeliness and accuracy, they suffer from limitations such as invasiveness and inconvenience of wearing the devices. Furthermore, their design and deployment costs are too high, and the data collection process can cause discomfort to the driver, affecting normal driving and hindering widespread application.

[0004] In recent years, deep learning-based computer vision technology has effectively compensated for the limitations of sensors due to its efficient feature extraction capabilities and end-to-end detection framework, gradually becoming the mainstream solution for detecting abnormal driving behavior. Researchers have designed a deep convolutional neural network (CNN) method to automatically extract key features from image samples and classify detected abnormal driving behaviors. Other researchers have proposed an attention-based deep neural network framework (ADNet) for driver behavior recognition, integrating a lightweight attention module into a deep learning model and combining spatial and channel attention modules for adaptive feature extraction. While these computer vision-based methods for detecting abnormal driving behavior are very convenient to deploy and do not interfere with normal driving, they are less timely and accurate compared to sensors. Furthermore, they are susceptible to complex environmental factors (such as low brightness or sudden exposure of the driver's real-time image during nighttime driving), leading to an inability to correctly respond to abnormal driver behavior.

[0005] Chinese patent document CN120340001A discloses a driver behavior recognition method in complex scenes based on a segmentation region framework. The method includes: constructing a standard driver behavior image dataset SAA13 and a special subset dataset; studying the core driver region in the images and labeling the driver regions in all dataset images; constructing a driver region detection and segmentation model based on YOLOv11 to segment the driver region in the images, thereby constructing a segmentation dataset; studying a driver behavior recognition model based on the segmentation dataset and constructing a recognition network based on ConvNeXt; and using the softmax activation function to decode the deep features extracted by the ConvNeXt stacked modules and output the final recognition result. The segmentation region framework achieves 95.55% accuracy and 62.5 FPS on the SAA13 comprehensive dataset and 95.96% recognition accuracy on the special subset dataset, providing a feasible method to improve the network's ability to recognize driver behavior.

[0006] Most existing methods for detecting abnormal driving behavior rely on sensors or computer vision. When a driver exhibits abnormal behavior, physiological signals such as electroencephalogram (EEG), electrocardiogram (ECG), and electromyogram (EMG) also become abnormal. Sensor-based methods can directly capture and analyze these physiological signals to determine whether the driver is engaging in abnormal driving behavior, offering good timeliness and accuracy. However, these methods require drivers to wear complex sensors, impacting the driving experience, and human sensor devices are expensive, hindering widespread application.

[0007] Computer vision-based methods for detecting abnormal driving behavior use real-time video images captured by in-vehicle cameras as input sources and detect abnormal driving behavior by analyzing the driver's visual characteristics. This method is more convenient in the data collection process and does not interfere with normal driving. However, it is easily affected by complex environments (such as low brightness or sudden exposure of the driver's real-time image when driving at night), which may lead to an inability to respond correctly to abnormal driver behavior, resulting in poor timeliness and accuracy.

[0008] Therefore, there is a need for a driver abnormal driving behavior detection method based on YOLOv11, which aims to improve detection efficiency and robustness. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a driver abnormal driving behavior detection method based on YOLOv11, which effectively solves the problem of detail loss caused by traditional fixed interpolation methods and enhances the efficiency and robustness of the model in detecting driver abnormal behavior.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a driver abnormal driving behavior detection method based on YOLOv11, specifically including the following steps:

[0011] S1 Dataset Construction: Collect a dataset of abnormal driving behaviors of drivers, preprocess the dataset, and use an image brightness adaptive adjustment mechanism to process overexposed or low-brightness images in the dataset; finally, divide the dataset and label it.

[0012] S2 Model Construction and Training: An improved YOLOv11 driver abnormal driving behavior detection model is constructed by introducing the GSConv module and Dysample dynamic upsampling technology into the YOLOv11 model; and the model is trained to extract features of different abnormal driving behavior attributes. After training, the final driver abnormal driving behavior detection model weights are obtained.

[0013] S3 test results and evaluation: The input image is tested using the weights of the trained detection model to obtain the driver's abnormal driving behavior attribute recognition results, and then the model performance is evaluated.

[0014] The above technical solution first introduces an adaptive image brightness adjustment mechanism in the preprocessing stage of abnormal driver behavior data. This adjusts low-brightness and overexposed images of abnormal driver behavior to normal brightness, reducing the environmental impact of nighttime driving. Then, a GSConv module is introduced into the backbone layer of the YOLOv11 model. This module optimizes the model by combining grouped convolution and spatial convolution, reducing computational cost while enhancing feature extraction capabilities. This significantly improves the model's ability to recognize abnormal driver behaviors (yawning, drinking water, smoking, and making phone calls). Finally, the DySample dynamic upsampling technique is employed in the YOLOv11 feature pyramid structure. By generating content-aware sampling points, upsampling is achieved, bypassing dynamic convolution. This effectively solves the detail loss problem caused by traditional fixed interpolation methods, enhancing the model's efficiency and robustness in detecting abnormal driver behavior.

[0015] Preferably, the specific steps of step S1 are as follows:

[0016] S11: Collect datasets of abnormal driving behaviors from various websites and public databases; search for suitable datasets in Baidu PaddlePaddle AI Studio and publicly available databases abroad, and collect download addresses for relevant image data from Baidu Images. Then, write scripts to download these images in batches and save them in the dataset. Simultaneously, film videos of multiple people performing different abnormal driving behaviors, and write scripts to extract images frame by frame from the videos, saving the extracted image frames in the same dataset;

[0017] S12: Preprocess the dataset and remove poor-quality data samples; first write a script to clean images smaller than 50×50, then manually clean and remove low-resolution images to ensure that the images contain clear information on abnormal driving behavior of the driver, and remove images that do not contain abnormal driving behavior of non-drivers.

[0018] S13: Input the preprocessed dataset into the image brightness adaptive adjustment mechanism to process images in the dataset that are overexposed or underexposed;

[0019] S14: Randomly divide the processed dataset into training, validation, and test sets, and label the attributes of the abnormal driving behavior dataset.

[0020] Preferably, the image brightness adaptive adjustment mechanism in step S13 is as follows: First, the image is converted to a grayscale image and the average value of the image pixels is calculated. Then, the calculated average brightness is compared with the set brightness thresholds for exposed and low-brightness images, wherein a brightness value greater than 200 is considered exposed and less than 50 is considered low-brightness. Then, it is determined whether the image needs to be enhanced or reduced in brightness and the corresponding brightness coefficient is calculated. Finally, the image brightness is changed according to the brightness coefficient.

[0021] The formula for the image brightness adaptive adjustment mechanism is:

[0022] (1);

[0023] (2);

[0024] Where (i,j) are the coordinates of each pixel in image I, and G' is the grayscale image of image I; This indicates that the pixel values ​​of the three color channels of image I are in the range of 0 to 255;

[0025] (3);

[0026] (4);

[0027] (5);

[0028] Where M*N is the image size, and avg is the average pixel value of the converted grayscale image G'. Based on the brightness thresholds of avg, low-brightness image, and exposed image, Target is set to 128 for the adjusted image brightness value, min is set to 50 for the brightness threshold of the low-brightness image, and max is set to 200 for the brightness threshold of the exposed image.

[0029] Preferably, in step S14, the ratio of the training set, validation set, and test set is 8:1:1, and the image order is shuffled; this improves the model's generalization ability, balances the influence of samples, and improves training efficiency. Then, the labelImg annotation tool is used to annotate the abnormal driving behavior dataset, ensuring that abnormal driving behavior in each image is accurately labeled for subsequent model training and evaluation. The labeling rules for abnormal driving behavior attributes are: 0: yawning, 1: drinking water, 2: smoking, 3: making a phone call.

[0030] Preferably, the specific steps of step S2 are as follows:

[0031] S21 Model Construction and Training: An improved YOLOv11 driver abnormal driving behavior detection model was built by introducing the GSConv module and DySample dynamic upsampling technology into the backbone layer and feature pyramid structure of the YOLOv11 model.

[0032] The S22 training model, namely the abnormal driving behavior detection model, involves setting training parameters, using the improved YOLOv11 model constructed with the training set to extract features from images of abnormal driving behavior, and iterating the training multiple times to obtain the weight file of the abnormal driving behavior detection model.

[0033] Preferably, in step S22, the model is trained based on the PyTorch framework, with 300 training iterations, a learning rate of 0.01, a batch size of 16, a worker size of 8, and an RTX4090 GPU to accelerate the training process.

[0034] The training process uses CIoU Loss as the model loss function, and the calculation process of this loss is presented as follows:

[0035] (13);

[0036] (14);

[0037] (15);

[0038] in, For weighting coefficients, and This is used as a measure of aspect ratio consistency. The Euclidean distance between the ground truth bounding box and the predicted bounding box is calculated, where c represents the diagonal length of the smallest bounding rectangle that simultaneously contains both the predicted and ground truth bounding boxes; w and h represent the width and height of the bounding box, respectively.

[0039] This invention uses several commonly used evaluation metrics, including precision, recall, mAP (mean accuracy), GFLOP, and parameters. Among these metrics, TP (True Positive) represents the case where the classifier correctly predicts a positive sample, and the actual data is also a positive sample. TN (True Negative) represents the case where the prediction is a negative sample, and it is indeed a negative sample, indicating accurate identification of the negative sample. FP (False Positive) and FN (False Negative) represent the number of false positives and false negatives, respectively.

[0040] Precision represents the accuracy of correctly predicting positive samples. Precision is calculated by dividing the number of correctly predicted positive samples by the total number of predicted positive samples. Higher precision means fewer false positives; conversely, lower precision means more false positives. Formula (6) represents the performance evaluation index of precision;

[0041] (6);

[0042] Recall represents the coverage of correctly predicted positive samples. It is calculated by dividing the number of correctly predicted positive samples by the total number of actual positive samples, plus the sum of TP (true positives) and FP (false negatives). The higher the recall, the fewer false negatives; conversely, the lower the recall, the more false negatives. Formula (7) represents the performance evaluation index of recall.

[0043] (7);

[0044] AP (Mean Precision) is a metric used to evaluate the performance of a single class or object detection model. In object detection tasks, the model predicts the location and class of objects in an image; AP quantifies the model's performance by calculating the area under the precision (P) and recall (R) curves. Precision refers to the proportion of correctly predicted positive objects out of all detected positive objects, while recall refers to the proportion of correctly detected objects out of all true positive objects. A higher AP value indicates better model performance; conversely, a lower AP value indicates worse model performance.

[0045] mAP (mean accuracy) is a metric used to comprehensively evaluate the performance of multi-class or multi-object detection models. In multi-class object detection tasks, each class has its own AP. mAP is the average of the APs of all classes, used to measure the average level of detection accuracy across multiple classes. Among these metrics, mAP@0.5 represents the average mAP when the IoU threshold is 0.5, which is commonly used for object detection evaluation, as shown in Equation (8).

[0046] (8);

[0047] Where N represents the total number of classes or object categories, and is the average precision of the classes.

[0048] Preferably, in step S21, the code for the GSConv module and DySample dynamic upsampling is designed first, and then integrated into the core files of the YOLOv11 model (such as tasks.py, conv.py, etc.); finally, the yolo11.yaml configuration file of the YOLOv11 model is modified to introduce the GSConv module into the backbone layer of the YOLOv11 model and the DySample dynamic upsampling technology into the feature pyramid structure, respectively.

[0049] Preferably, in step S21, the convolution of the GSConv module is decomposed into group convolution and spatial convolution. Group convolution is: dividing the input feature map into several groups and performing convolution operation on each group to reduce the computational cost of each group convolution. Spatial convolution is: adding spatial convolution operation to the output after each group convolution operation to capture cross-group features.

[0050] Preferably, the GSConv module first performs a normal convolution downsampling on the input data, then uses DWConv depthwise convolution, concatenates the results of the two convolution operations, and finally performs a shuffle operation to connect the corresponding channel numbers of the two convolutions.

[0051] Preferably, the DySample dynamic upsampling in step S21 is mainly divided into three parts: a sampling point generator, a static scope factor, and a dynamic scope factor;

[0052] The sampling point generator generates a set of sampling points to determine which points to upsample at. Then, the grid sampling method is used to sample the original feature map to generate a new high-resolution feature map. This sampling point-based upsampling method is dynamic, meaning that it can automatically adjust the sampling points according to different input feature maps, avoiding the limitations of traditional interpolation methods and making upsampling more flexible and accurate.

[0053] The static scoping factor is obtained by first generating low-dimensional features from the input feature map through a linear transformation, then multiplying by a fixed factor, performing a pixel shuffle operation, and finally adding the generated high-resolution features to an offset feature map calculated by a fixed scoping factor. The static scoping factor makes upsampling stable and controllable, and is suitable for handling scenarios with consistent scale changes.

[0054] The dynamic scoping factor is similar to the static scoping factor, but it dynamically adjusts the scaling factor based on the input feature map and then amplifies the feature map through a pixel shuffle operation. It is also obtained by adding the generated feature map and the dynamically shifted feature map. The dynamic scoping factor can adaptively adjust the scaling scale of different inputs, making it particularly suitable for processing multi-scale features or images with large resolution variations. It can better capture detailed information at different scales, improving the model's flexibility and adaptability.

[0055] Preferably, in step S3, the data after model training is organized and a scatter plot is drawn to compare with other models, thereby evaluating the model and obtaining the model's performance.

[0056] Each column of the confusion matrix represents the predicted class, and each row represents the actual class; each element in the matrix represents the combination of the actual class and the predicted class; specifically, the confusion matrix consists of the following four parts:

[0057] True Positive (TP): The number of samples that are actually positive and are also predicted to be positive; False Negative (FN): The number of samples that are actually positive but are predicted to be negative; False Positive (FP): The number of samples that are actually negative but are predicted to be positive; True Negative (TN): The number of samples that are actually negative and are also predicted to be negative.

[0058] The following evaluation results for the model can be obtained from the four parts of the confusion matrix described above:

[0059] Accuracy: Represents the proportion of samples correctly predicted by the model out of the total number of samples. The formula is:

[0060] Accuracy = (TP + TN) / (TP + FN + FP + TN) (9);

[0061] Precision: Represents the proportion of samples predicted as positive that are actually positive. The formula is:

[0062] Precision = TP / (TP + FP) (10);

[0063] Recall: Represents the proportion of samples that were actually positive but were correctly predicted as positive. The formula is:

[0064] Recall = TP / (TP + FN) (11);

[0065] PR curve: Precision and recall curves, with recall on the X-axis and precision on the Y-axis;

[0066] F1 Score: The harmonic mean of precision and recall, used to comprehensively evaluate the model's performance. The formula is:

[0067] F1_Score = 2 * (Precision * Recall) / (Precision + Recall) (12).

[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0069] (1) High accuracy: The improved YOLOv11 network model effectively solves the problem of information isolation between groups caused by simple group convolution through the channel shuffling operation of GSConv, and cross-fusions the feature information learned by different groups, which can better capture the subtle features of abnormal driving behavior of drivers, thereby improving the accuracy of detection and reducing false detections or missed detections; Dysample dynamic upsampling is introduced into the feature pyramid structure of YOLOv11 to replace the traditional deconvolution or bilinear interpolation method. DySample dynamic upsampling technology optimizes the fusion process of multi-scale features by dynamically adjusting the upsampling weights, thereby improving the model detection efficiency;

[0070] (2) High timeliness: The improved YOLOv11 network model is a lightweight target detection model with fast inference speed, which can quickly detect and respond to abnormal driving behavior targets in images or videos.

[0071] (3) Multi-scale detection: The improved YOLOv11 network model supports multi-scale feature fusion and can detect abnormal driving behavior of drivers at different scales;

[0072] (4) Anti-interference ability: The image brightness adaptive adjustment mechanism can adaptively adjust the brightness of the image or video, which enhances the robustness of the model to exposure and nighttime environmental interference, that is, enhances the model's adaptability to the environment and reduces the interference of exposure and nighttime environment on the detection of abnormal driving behavior of the driver; enabling the model to effectively identify fall behavior in complex environments.

[0073] (5) Scalability: This invention can be combined with other sensors to enhance the accuracy and reliability of detecting abnormal driving behavior of drivers, and is suitable for a variety of application scenarios. Attached Figure Description

[0074] Figure 1 This is a flowchart illustrating the driver abnormal driving behavior detection method based on YOLOv11 improved according to the present invention.

[0075] Figure 2 This is a structural diagram of the GSConv module in the driver abnormal driving behavior detection method based on YOLOv11 of the present invention;

[0076] Figure 3 The diagram shows the structure of the DySample dynamic upsampling module in the driver abnormal driving behavior detection method based on YOLOv11 of the present invention; wherein (a) is the structure of the sampling-based dynamic upsampling module; and (b) is the structure of the dynamic sampling module using a sampling point generator.

[0077] Figure 4 This is the network model of the abnormal driving behavior detection algorithm in the driver abnormal driving behavior detection method based on YOLOv11 of the present invention;

[0078] Figure 5 for Figure 4 The C3k2 feature extraction module uses a customizable convolution kernel C3k with both correct and incorrect structure diagrams; where (a) represents the case where C3k is correct and (b) represents the case where C3k is incorrect.

[0079] Figure 6 This is a scatter plot showing the model accuracy and GFLOPs in the driver abnormal driving behavior detection method based on YOLOv11 of this invention. Detailed Implementation

[0080] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solutions of the present invention more clearly, and should not be used to limit the scope of protection of the present invention.

[0081] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0082] In this invention, the term "connection" can mean either a direct connection between components or an indirect connection between components through other components.

[0083] The basic known technologies of this invention are: YOLOv11 object detection algorithm, GSConv convolution module, and Dysample upsampling technology.

[0084] (1) YOLOv11 object detection algorithm:

[0085] YOLOv11 is an improved single-stage object detection algorithm based on YOLOv8, integrating a more efficient feature extraction architecture, optimized computational efficiency, and stronger multi-task support capabilities. By introducing innovative modules (such as C3K2 and C2PSA) and a lightweight design, YOLOv11 significantly improves detection accuracy and inference speed, especially in small object detection and adaptability to complex scenes. This invention uses the YOLOv11 model for detecting abnormal driving behavior. The YOLOv11 network structure can be divided into four parts: Input layer, Backbone network, Neck, and Head.

[0086] YOLOv11's input layer inherits the Mosaic image augmentation technique from YOLOv8 and further extends multi-scale data augmentation strategies. Mosaic image augmentation generates training samples by randomly cropping, scaling, and stitching four images, enhancing the model's robustness to targets of different scales and shapes. Simultaneously, adaptive anchor box calculation dynamically adjusts the anchor box size and scale based on dataset characteristics, improving the model's adaptability to target distribution. Furthermore, adaptive image scaling unifies the input image to a standard size (e.g., 640×640) through letterbox padding, addressing the issue of image size variations.

[0087] The backbone network of YOLOv11 consists of the C3K2 module, the SPPF module, and the C2PSA module, which improves the efficiency and flexibility of feature extraction. The C3K2 module is an improvement on the C2f module of YOLOv8, using 3×3 small convolutional kernels instead of traditional large convolutional kernels, reducing computation while preserving feature expressive power. The SPPF module captures contextual information of multi-scale features through parallel max-pooling branches (such as 5×5, 9×9, and 13×13), enhancing the model's ability to perceive complex textures and shapes. The C2PSA module introduces a point-based spatial attention (PSA) mechanism after SPPF, dynamically weighting the spatial dimension of the feature map through multi-head attention, improving the model's ability to focus on key target regions. This combined design allows YOLOv11 to significantly improve the accuracy and efficiency of feature extraction while maintaining a lightweight architecture.

[0088] The YOLOv11 neck module combines FPN (Feature Pyramid Network) and PANet (Path Aggregation Network), and further optimizes the feature fusion process through the C2PSA module. The FPN module passes high-level feature information (such as category information) from top to bottom, fusing it with low-level feature details (such as edge information) to generate multi-scale feature maps for detecting targets of different sizes. The PANet module, on the other hand, passes strong localization features from bottom to top, strengthening the interaction of multi-scale features through lateral connections, thus solving the feature loss problem of traditional FPN in complex scenes. This multi-scale feature fusion strategy significantly improves the detection performance for occluded and dense targets, enabling YOLOv11 to maintain high-precision detection capabilities even in complex backgrounds.

[0089] YOLOv11's detection head uses Depthwise Separable Convolution (DWConv) instead of traditional convolution, significantly reducing the number of parameters and computational overhead, while supporting multi-task extensions (such as segmentation heads and pose estimation heads). Furthermore, in its loss function design, YOLOv11 enhances the model's adaptability to long-tailed class data through a dynamic label assignment strategy, reducing classification loss. In the post-processing stage, YOLOv11 introduces Soft-NMS to reduce false positives and false negatives in dense target scenes.

[0090] (2) GSConv module:

[0091] GSConv (Group Shuffle Convolution) is an efficient and lightweight convolutional module that combines the advantages of standard convolution (SC) and depthwise separable convolution (DSC) to reduce computational costs while preserving or even enhancing feature representation capabilities. Its core idea is to use a "shuffle" operation to mix features generated by SC and features generated by DSC, thereby breaking the limitation of isolated channel information in DSC. GSConv first uses SC to extract global features, then uses DSC to capture local spatial information, and finally achieves inter-channel information interaction through feature concatenation and shuffling operations, making the output of DSC closer to the feature representation capabilities of SC.

[0092] Key features of the GSConv module:

[0093] 1. High real-time detection efficiency: GSConv adopts grouped convolution as the main structure, which significantly reduces the number of model parameters and floating-point operations (FLOPs) compared to standard convolution, making it perform well in real-time object detection. It has a fast detection rate and low computing resource requirements, making it very suitable for platforms with limited computing resources (such as mobile devices, embedded devices, etc.) and application scenarios with high real-time requirements.

[0094] 2. Strong feature extraction capability: GSConv effectively solves the problem of information isolation between groups caused by simple grouped convolution through channel shuffling. Channel shuffling cross-integrates feature information learned from different groups, thereby significantly improving feature representation capability and model accuracy while maintaining high computational efficiency;

[0095] 3. Good robustness: GSConv's channel shuffling mechanism can interweave feature information from different sources, ensuring that subsequent layers of the network can acquire diverse features learned from different convolutional layers, optimizing the information flow within the network, and enhancing the model's ability to learn robust features.

[0096] 4. Wide range of applications: GSConv is suitable for resource-constrained real-time detection tasks, including mobile image recognition, real-time object detection, video analysis, augmented reality (AR), and smart vision tasks on IoT devices.

[0097] (3) DySample dynamic upsampling technique:

[0098] DySample is an innovative dynamic upsampler that primarily employs point resampling instead of traditional kernel-based methods, thus improving resource efficiency. DySample redesigns the upsampling process from a point-sampling perspective, bypassing dynamic convolutions to directly generate content-aware sampling points, thereby improving accuracy while maintaining high efficiency. Its core idea is to generate a continuous feature map from input features using bilinear interpolation, then use linear projection to generate an offset, and dynamically adjust the sampling position and range to achieve more accurate upsampling. By avoiding time-consuming dynamic convolutions and additional self-generated dynamic kernels, DySample significantly reduces computational workload and latency.

[0099] Key features of DySample dynamic upsampling technology:

[0100] 1. High real-time detection efficiency: DySample's lightweight design and highly optimized implementation make it superior to traditional dynamic upsamplers in terms of inference speed and latency, making it suitable for real-time application requirements;

[0101] 2. Good robustness: DySample effectively reduces artifacts and boundary blurring problems through offset range constraints and grouping strategies, thereby improving the quality and stability of output features.

[0102] Example: Figure 1 As shown, this driver abnormal driving behavior detection method based on YOLOv11 improvement specifically includes the following steps:

[0103] S1 Dataset Construction: Collect a dataset of abnormal driving behaviors of drivers, preprocess the dataset, and use an image brightness adaptive adjustment mechanism to process overexposed or low-brightness images in the dataset; finally, divide the dataset and label it.

[0104] In some embodiments, suitable datasets are searched in Baidu PaddlePaddle AI Studio and publicly available databases abroad, and download addresses for relevant image data are collected from Baidu Images. Then, scripts are written to download these images in batches and save them in the dataset.

[0105] The specific steps of step S1 are as follows:

[0106] S11: Collect datasets of abnormal driving behaviors from various websites and public databases; search for suitable datasets in Baidu PaddlePaddle AI Studio and publicly available databases abroad, and collect download addresses for relevant image data from Baidu Images. Then, write scripts to download these images in batches and save them in the dataset. Simultaneously, film videos of multiple people performing different abnormal driving behaviors, and write scripts to extract images frame by frame from the videos, saving the extracted image frames in the same dataset;

[0107] S12: Preprocess the dataset and remove poor-quality data samples; first write a script to clean images smaller than 50×50, then manually clean and remove blurry images to ensure that there is clear information about abnormal driving behavior of the driver in the images, and remove images that do not contain abnormal driving behavior of non-drivers.

[0108] S13: Input the preprocessed dataset into the image brightness adaptive adjustment mechanism to process images in the dataset that are overexposed or underexposed;

[0109] The image brightness adaptive adjustment mechanism in step S13 is as follows: First, the image is converted to a grayscale image and the average value of the image pixels is calculated. Then, the calculated average brightness is compared with the brightness thresholds of the set exposure image and low brightness image, where the brightness value of the image is set to be greater than 200 for exposure and less than 50 for low brightness. Then, it is determined whether the image needs to be enhanced or reduced in brightness and the corresponding brightness coefficient is calculated. Finally, the image brightness is changed according to the brightness coefficient.

[0110] The formula for the image brightness adaptive adjustment mechanism is:

[0111] (1);

[0112] (2);

[0113] Where (i,j) are the coordinates of each pixel in image I, and G' is the grayscale image of image I; This indicates that the pixel values ​​of the three color channels of image I are in the range of 0 to 255;

[0114] (3);

[0115] (4);

[0116] (5);

[0117] Where M*N is the image size, and avg is the average pixel value of the converted grayscale image G'. Based on the brightness thresholds of avg, low-brightness image, and exposed image, Target is set to 128 for the adjusted image brightness value, min is set to 50 for the brightness threshold of low-brightness image, and max is set to 200 for the brightness threshold of exposed image.

[0118] In some embodiments, the specific steps for processing image brightness using an image brightness adaptive adjustment mechanism are as follows:

[0119] (1) Design code for an image brightness adaptive adjustment mechanism;

[0120] (2) Input the image data into the dataset. The image brightness adaptive adjustment mechanism will convert the image into a grayscale image and calculate the average brightness of the image pixels.

[0121] (3) Compare the calculated average brightness with the set brightness thresholds for the exposed image and the low brightness image to determine whether the image needs to be enhanced or reduced and calculate the corresponding brightness coefficient.

[0122] (4) Adjust the image brightness according to the brightness coefficient;

[0123] S14: Randomly divide the processed dataset into training set, validation set and test set and label the attributes of the abnormal driving behavior dataset;

[0124] In step S14, the ratio of the training set, validation set, and test set is 8:1:1, and the image order is shuffled to improve the model's generalization ability, balance the influence of samples, and improve training efficiency. Then, the labelImg annotation tool is used to annotate the abnormal driving behavior dataset, ensuring that abnormal driving behavior in each image is accurately labeled for subsequent model training and evaluation. The labeling rules for abnormal driving behavior attributes are: 0: yawning, 1: drinking water, 2: smoking, 3: making a phone call.

[0125] S2 Model Construction and Training: An improved YOLOv11 driver abnormal driving behavior detection model is constructed by introducing the GSConv module and Dysample dynamic upsampling technology into the YOLOv11 model; and the model is trained to extract features of different abnormal driving behavior attributes. After training, the final driver abnormal driving behavior detection model weights are obtained.

[0126] The specific steps of step S2 are as follows:

[0127] S21 Model Construction and Training: An improved YOLOv11 driver abnormal driving behavior detection model was built by introducing the GSConv module and DySample dynamic upsampling technology into the backbone layer and feature pyramid structure of the YOLOv11 model.

[0128] In step S21, the code for the GSConv module and DySample dynamic upsampling is first designed, and then integrated into the core files of the YOLOv11 model (such as tasks.py, conv.py, etc.); finally, the yolo11.yaml configuration file of the YOLOv11 model is modified to introduce the GSConv module into the backbone layer of the YOLOv11 model and the DySample dynamic upsampling technology into the feature pyramid structure, respectively.

[0129] In some embodiments, the specific steps for improving the YOLOv11 backbone layer and integrating the GSConv module are as follows:

[0130] (1) Design the GSConv module code and integrate it into the core files of YOLOv11 (such as tasks.py, conv.py, etc.);

[0131] (2) Modify the yolo11.yaml configuration file of YOLOv11 and introduce the GSConv module in the backbone layer to replace the traditional convolutional layer; the GSConv module enhances the feature extraction capability and improves the model detection efficiency by combining group convolution and spatial convolution.

[0132] In some embodiments, the specific steps for improving the feature pyramid structure of YOLOv11 and integrating Dysample dynamic upsampling are as follows:

[0133] (1) Design the DySample dynamic upsampling code and integrate it into the core files of YOLOv11 (such as tasks.py, conv.py, etc.);

[0134] (2) Modify the yolo11.yaml configuration file of YOLOv11 to introduce DySample upsampling technology into the feature pyramid structure, replacing the traditional deconvolution or bilinear interpolation method. DySample dynamic upsampling technology optimizes the multi-scale feature fusion process by dynamically adjusting the upsampling weights, thereby improving the model detection efficiency;

[0135] In step S21, the convolution of the GSConv module is decomposed into group convolution and spatial convolution. Group convolution is: dividing the input feature map into several groups and performing convolution operation on each group to reduce the computation of each group convolution. Spatial convolution is: adding spatial convolution operation to the output after each group convolution operation to capture cross-group features. Figure 2 Here is a structural diagram of the GSConv module; Figure 2 In Chinese: input; Conv: regular convolution; channels: number of channels; DWConv: depthwise convolution; Concat: concatenation; shuffle: shuffling operation; output: output.

[0136] The GSConv module first downsamples the input data using a regular Conv convolution with c1 channels, then uses a DWConv depthwise convolution, concatenates the results of the two convolution operations, and finally performs a shuffle operation to connect the corresponding channels of the two convolutions and output the data.

[0137] DySample is an ultra-lightweight and efficient dynamic upsampling method designed to learn upsampling through learning sampling. Unlike traditional kernel-based dynamic upsampling methods, DySample is designed from a point sampling perspective, segmenting a single point into multiple points to achieve sharper edges. Its core technology utilizes dynamic sampling to achieve the upsampling process. DySample performs the sampling process by finding the correct semantic cluster for each upsampling point, thus achieving efficient upsampling. Furthermore, DySample features a customized design for upsampling, sampling a single point at each upsampling location and segmenting a point into s... 2 DySample offers several upsampling points. Compared to other dynamic upsamplers, it has lower costs and faster backpropagation speed in terms of inference time, training memory, training time, GFLOPs, and number of parameters. Figure 3 This is a structural diagram of the DySample dynamic upsampling module; where... Figure 3 (a) in the diagram is a structural diagram of the sampling-based dynamic upsampling module; Figure 3 (b) shows the structure diagram of dynamic sampling using a sampling point generator;

[0138] Figure 3 In Chinese: Sampling based dynamic upsampling;

[0139] sampling point generator; sampling set; gridsample;

[0140] Sampling point generator in DySample: A sampling point generator for dynamic upsampling;

[0141] Static Scope Factor; Dynamic Scope Factor.

[0142] pixel shuffle: pixel shuffling operation; linear: linear layer;

[0143] In step S21, the DySample dynamic upsampling mainly consists of three parts: a sampling point generator, a static scope factor, and a dynamic scope factor.

[0144] The sampling point generator generates a set of sampling points to determine which points to upsample at. Then, the grid sampling method is used to sample the original feature map to generate a new high-resolution feature map. This sampling point-based upsampling method is dynamic, meaning that it can automatically adjust the sampling points according to different input feature maps, avoiding the limitations of traditional interpolation methods and making upsampling more flexible and accurate.

[0145] The static scope factor is obtained by first generating low-dimensional features from the input feature map through a linear transformation, then multiplying by a fixed factor, performing a pixel shuffle operation, and finally adding the generated high-resolution features to an offset feature map calculated by the fixed scope factor. The static scope factor makes upsampling stable and controllable, and is suitable for handling scenarios with consistent scale changes.

[0146] The dynamic scoping factor is similar to the static scoping factor, but it dynamically adjusts the scaling factor based on the input feature map and then amplifies the feature map through a pixel shuffle operation. It is also obtained by adding the generated feature map to the dynamically offset feature map. The dynamic scoping factor can adaptively adjust the scaling scale of different inputs, making it particularly suitable for processing multi-scale features or images with large resolution variations. It can better capture detailed information at different scales, improving the model's flexibility and adaptability.

[0147] S22 training model, i.e. driver abnormal driving behavior detection: set training parameters, use the training set to extract features from driver abnormal driving behavior images using the constructed improved YOLOv11 model, and iterate the training multiple times to obtain the driver abnormal driving behavior detection model weight file.

[0148] In some embodiments, the model training in step S22 is based on the PyTorch framework, with 300 training iterations, a learning rate of 0.01, a batch size of 16, a worker size of 8, and an RTX4090 GPU to accelerate the training process.

[0149] The training process uses CIoU Loss as the model loss function, and the calculation process of this loss is presented as follows:

[0150] (13);

[0151] (14);

[0152] (15);

[0153] in, For weighting coefficients, and This is used as a measure of aspect ratio consistency. The Euclidean distance between the ground truth bounding box and the predicted bounding box is calculated, where b represents the center coordinate of the predicted bounding box. The coordinates of the center of the ground truth bounding box are given; c represents the diagonal length of the smallest bounding rectangle that simultaneously contains both the predicted and ground truth bounding boxes; w and h represent the width and height of the predicted bounding box, respectively. and represents the width and height of the ground truth bounding box, respectively; arctan is the arctangent function; IoU is the intersection-union ratio of the predicted bounding box and the ground truth bounding box;

[0154] In some embodiments, several commonly used evaluation metrics are used to evaluate the model after training, including accuracy, recall, mAP (mean accuracy), GFLOPs, and parameters. Among these metrics, TP (True Positive) indicates that the classifier correctly predicted a positive sample, and the actual data was also a positive sample. TN (True Negative) indicates that the predicted negative sample was indeed a negative sample, representing accurate identification of the negative sample. FP (False Positive) and FN (False Negative) represent the number of false positives and false negatives, respectively.

[0155] Precision represents the accuracy of correctly predicting positive samples. Precision is calculated by dividing the number of correctly predicted positive samples by the total number of predicted positive samples; the higher the precision, the fewer false positives; conversely, the lower the precision, the more false positives; Formula (6) represents the performance evaluation index of precision;

[0156] (6);

[0157] Recall represents the coverage of correctly predicted positive samples. It is calculated by dividing the number of correctly predicted positive samples by the total number of actual positive samples, plus the sum of TP (true positives) and FP (false negatives). The higher the recall, the fewer false negatives; conversely, the lower the recall, the more false negatives. Formula (7) represents the performance evaluation index of recall.

[0158] (7);

[0159] AP (Mean Precision) is a metric used to evaluate the performance of a single class or object detection model. In object detection tasks, the model predicts the location and class of objects in an image. AP quantifies model performance by calculating the area under the precision (P) and recall (R) curves. Precision refers to the proportion of correctly predicted positive objects out of all detected positive objects, while recall refers to the proportion of correctly detected objects out of all true positive objects. A higher AP value indicates better model performance; conversely, a lower AP value indicates worse model performance.

[0160] mAP (mean accuracy) is a metric used to comprehensively evaluate the performance of multi-class or multi-object detection models. In multi-class object detection tasks, each class has its own AP. mAP is the average of the APs of all classes, used to measure the average level of detection accuracy across multiple classes. Among these metrics, mAP@0.5 represents the average mAP when the IoU threshold is 0.5, which is commonly used for object detection evaluation, as shown in Equation (8).

[0161] (8);

[0162] Where N represents the total number of classes or object categories, and is the average precision of the classes;

[0163] S3 Test Results and Evaluation: The input image is tested using the weights of the trained detection model to obtain the driver's abnormal driving behavior attribute recognition results, and then the model's performance is evaluated; In step S3, the data after model training is organized and a scatter plot is drawn for comparison with other models to evaluate the model and obtain the model's effect.

[0164] Each column of the confusion matrix represents the predicted class, and each row represents the actual class; each element in the matrix represents the combination of the actual class and the predicted class; specifically, the confusion matrix consists of the following four parts:

[0165] True Positive (TP): The number of samples that are actually positive and are also predicted to be positive; False Negative (FN): The number of samples that are actually positive but are predicted to be negative; False Positive (FP): The number of samples that are actually negative but are predicted to be positive; True Negative (TN): The number of samples that are actually negative and are also predicted to be negative.

[0166] The following evaluation results for the model can be obtained from the four parts of the confusion matrix described above:

[0167] Accuracy: Represents the proportion of samples correctly predicted by the model out of the total number of samples. The formula is:

[0168] Accuracy = (TP + TN) / (TP + FN + FP + TN) (9);

[0169] Precision: Represents the proportion of samples predicted as positive that are actually positive. The formula is:

[0170] Precision = TP / (TP + FP) (10);

[0171] Recall: Represents the proportion of samples that were actually positive but were correctly predicted as positive. The formula is:

[0172] Recall = TP / (TP + FN) (11);

[0173] PR curve: Precision and recall curves, with recall on the X-axis and precision on the Y-axis;

[0174] F1 Score: The harmonic mean of precision and recall, used to comprehensively evaluate the model's performance. The formula is:

[0175] F1_Score = 2 * (Precision * Recall) / (Precision + Recall) (12).

[0176] In some embodiments, a YOLOv11 model is built using PyCharm, and then the GSConv module and DySample dynamic upsampling are introduced to improve the YOLOv11 model. The improved YOLOv11 model is used to extract features from images of abnormal driving behavior, and the model is trained iteratively multiple times to obtain the driver abnormal driving behavior detection model weight file. The feature weights are then continuously updated iteratively, and the model is evaluated using a validation set. The network model of the driver abnormal driving behavior detection algorithm is as follows: Figure 4 and Figure 5 As shown; where Figure 5 In (a), when the customizable convolution kernel C3k is correct, Figure 5 In (b), the customizable convolutional kernel C3k is set to error; True or False are parameter settings, where c3K=True means the C3 structure is enabled, and False means it is not enabled (in which case the fixed Bottleneck is used).

[0177] Figure 4 and Figure 5 In this context: Input; Output; Backbone; Neck; Head; Conv; C3k2; C3k2 feature extraction module; C3k; Customizable convolution kernel; GSConv; Grouped spatial convolution; SPPF; Pooling technique; C2PSA; Pyramid compression attention module; Concat; DySample; Detect; Split; Bottleneck residual block.

[0178] Most existing studies use YOLO-based object detection algorithms to train drivers on abnormal driving behaviors. This approach often suffers from shortcomings in timeliness and accuracy, and cannot effectively handle image data in complex environments, resulting in insufficient ability to identify abnormal driving behaviors and a high risk of false positives or false negatives. Some studies have improved the YOLO model by enhancing its feature extraction capabilities to increase the accuracy of abnormal driving behavior detection. However, this significantly increases model complexity and drastically reduces detection efficiency, leading to an inability to respond promptly to abnormal driving behaviors and failing to meet the needs of practical applications. Therefore, this invention introduces the GSConv module and DySample dynamic upsampling on top of the YOLOv11 model. This reduces model complexity while maintaining accuracy, achieving a better balance between accuracy and speed. Table 1 shows a comparison of ablation experiments using this invention.

[0179] Table 1 Comparison of ablation experimental results of different models

[0180] Improved network structure Percision / % Recall / % mAP50 / % mAP50-95 / % GFLOPs YOLOv11 95.7 88.5 96.6 61.2 6.3 YOLOv11+DySample 95.8 87.2 96.9 62.1 5.8 YOLOv11+GSConv 96.5 89.7 97.1 62.6 5.4 YOLOv11+DySample+GSConv 96.9 90.8 97.4 63.3 5.1

[0181] As shown in Table 1, compared with YOLOv11 with DySample and GSConv, YOLOv11 with DySample and YOLOv11 with GSConv, the improved model of this invention improves the accuracy of detecting abnormal driving behavior by 1.2%, 1.1%, and 0.4%, respectively; the recall rate by 2.3%, 3.6%, and 1.1%, respectively; the mAP50 value by 0.8%, 0.5%, and 0.3%, respectively; the mAP50-95 value by 2.1%, 1.2%, and 0.7%, respectively; and the GFLOPs (billions of floating-point operations per second) by 1.2, 0.7, and 0.3%, respectively. This indicates that the improved model of this invention is highly suitable for the abnormal driving behavior dataset of this invention. Figure 6 This is a scatter plot of model accuracy and GFLOPs, with accuracy on the ordinate and GFLOPs on the slant side.

[0182] For those skilled in the art, the specific embodiments are merely illustrative descriptions of the present invention. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution to other situations without modification, are all within the protection scope of the present invention.

Claims

1. A method for detecting abnormal driving behavior of drivers based on an improved version of YOLOv11, characterized in that, Specifically, the following steps are included: S1 Dataset Construction: Collect a dataset of abnormal driving behaviors of drivers, preprocess the dataset, and use an image brightness adaptive adjustment mechanism to process overexposed or low-brightness images in the dataset; finally, divide the dataset and label it. S2 Model Construction and Training: An improved YOLOv11 driver abnormal driving behavior detection model is constructed by introducing the GSConv module and Dysample dynamic upsampling technology into the YOLOv11 model; and the model is trained to extract features of different abnormal driving behavior attributes. After training, the final driver abnormal driving behavior detection model weights are obtained. S3 test results and evaluation: The input image is tested using the weights of the trained detection model to obtain the driver's abnormal driving behavior attribute recognition results, and then the model performance is evaluated.

2. The driver abnormal driving behavior detection method based on YOLOv11 as described in claim 1, characterized in that, The specific steps of step S1 are as follows: S11: Collect datasets of abnormal driving behaviors of drivers from major websites and public databases; at the same time, film videos of multiple people performing different abnormal driving behaviors, and write scripts to extract the video frame by frame, and save the extracted image frames in the same dataset. S12: Preprocess the dataset by first writing a script to clean images smaller than 50×50, then removing blurry images to ensure that the images contain clear information about abnormal driving behavior of the driver, and removing images that do not contain abnormal driving behavior of non-drivers. S13: Input the preprocessed dataset into the image brightness adaptive adjustment mechanism to process images in the dataset that are overexposed or underexposed; S14: Randomly divide the processed dataset into training, validation, and test sets, and label the attributes of the abnormal driving behavior dataset.

3. The driver abnormal driving behavior detection method based on YOLOv11 as described in claim 2, characterized in that, The image brightness adaptive adjustment mechanism in step S13 is as follows: First, the image is converted to a grayscale image and the average value of the image pixels is calculated. Then, the calculated average brightness is compared with the brightness thresholds of the set exposure image and low brightness image, where the brightness value of the image is set to be greater than 200 for exposure and less than 50 for low brightness. Then, it is determined whether the image needs to be enhanced or reduced in brightness and the corresponding brightness coefficient is calculated. Finally, the image brightness is changed according to the brightness coefficient. The formula for the image brightness adaptive adjustment mechanism is: （1）； （2）； Where (i,j) are the coordinates of each pixel in image I, and G' is the grayscale image of image I; This indicates that the pixel values ​​of the three color channels of image I are in the range of 0 to 255; （3）； （4）； （5）； Where M*N is the image size, and avg is the average pixel value of the converted grayscale image G'. Based on the brightness thresholds of avg, low-brightness images, and exposed images, Target is set to 128 for the adjusted image brightness value, min is set to 50 for the brightness threshold of low-brightness images, and max is set to the brightness threshold of exposed images.

4. The driver abnormal driving behavior detection method based on YOLOv11 as described in claim 2, characterized in that, In step S14, the ratio of the training set, validation set, and test set is 8:1:1, and the order of the images is shuffled. Then, the abnormal driving behavior dataset is labeled, and the abnormal driving behavior attribute labeling rules are 0: yawning, 1: drinking water, 2: smoking, and 3: making a phone call.

5. The driver abnormal driving behavior detection method based on YOLOv11 as described in claim 2, characterized in that, The specific steps of step S2 are as follows: S21 Model Construction and Training: An improved YOLOv11 driver abnormal driving behavior detection model was built by introducing the GSConv module and DySample dynamic upsampling technology into the backbone layer and feature pyramid structure of the YOLOv11 model. S22 training model, i.e. driver abnormal driving behavior detection: set training parameters, use the improved YOLOv11 model constructed with the training set to extract features from driver abnormal driving behavior images, and iterate the training multiple times to obtain the driver abnormal driving behavior detection model weight file.

6. The driver abnormal driving behavior detection method based on YOLOv11 as described in claim 5, characterized in that, In step S21, the GSConv module and DySample dynamic upsampling are designed first, and then integrated into the core file of the YOLOv11 model. Finally, the yolo11.yaml configuration file of the YOLOv11 model is modified to introduce the GSConv module into the backbone layer of the YOLOv11 model and the DySample dynamic upsampling technology into the feature pyramid structure.

7. The driver abnormal driving behavior detection method based on YOLOv11 as described in claim 5, characterized in that, In step S21, the convolution of the GSConv module is decomposed into group convolution and spatial convolution. Group convolution is: dividing the input feature map into several groups and performing convolution operation on each group to reduce the computation of each group convolution. Spatial convolution is: adding spatial convolution operation to the output after each group convolution operation to capture cross-group features.

8. The driver abnormal driving behavior detection method based on YOLOv11 as described in claim 7, characterized in that, The GSConv module first performs a downsampling operation on the input data using a regular convolution, then uses a DWConv depthwise convolution, concatenates the results of the two convolution operations, and finally performs a shuffle operation to connect the corresponding channel numbers of the two convolutions.

9. The driver abnormal driving behavior detection method based on YOLOv11 as described in claim 5, characterized in that, In step S21, the DySample dynamic upsampling is divided into three parts: a sampling point generator, a static scoping factor, and a dynamic scoping factor; The sampling point generator generates a set of sampling points to determine which points to upsample at; then, the original feature map is sampled using a grid sampling method to generate a new high-resolution feature map. The static scope factor is obtained by first generating low-dimensional features from the input feature map through a linear transformation, then multiplying it by a fixed factor, performing pixel shuffling, and finally adding the generated high-resolution features to an offset feature map calculated by a fixed scope factor. The dynamic scope factor dynamically adjusts the scaling factor based on the input feature map, and then the feature map is magnified through pixel shuffling; similarly, it is obtained by adding the generated feature map and the dynamically offset feature map.

10. The driver abnormal driving behavior detection method based on YOLOv11 as described in claim 5, characterized in that, In step S3, confusion matrix diagram, PR curve and F1 score line are used to evaluate the model and obtain the model's performance. The confusion matrix consists of four parts: TP: True Negative Class, representing the number of samples that are actually positive and are also predicted to be positive; FN: False Negative Class, representing the number of samples that are actually positive but are predicted to be negative; FP: False Positive Class, representing the number of samples that are actually negative but are predicted to be positive; TN: True Negative Class, representing the number of samples that are actually negative and are also predicted to be negative. The following evaluation results for the model can be obtained from the four parts of the confusion matrix described above: Accuracy: Represents the proportion of samples correctly predicted by the model out of the total number of samples. The formula is: Accuracy = (TP + TN) / (TP + FN + FP + TN) (9); Precision: Represents the proportion of samples predicted as positive that are actually positive, expressed by the formula; Precision = TP / (TP + FP) (10); Recall: Represents the proportion of samples that were actually positive but were correctly predicted as positive. The formula is: Recall = TP / (TP + FN) (11); PR curve: Precision and recall curves, with recall on the X-axis and precision on the Y-axis; F1 score: The harmonic mean of precision and recall, used to comprehensively evaluate the performance of the model, and the formula is: F1_Score = 2 * (Precision * Recall) / (Precision + Recall) (12).

Images