Driver distraction behavior recognition method based on optimized YOLO fusion network model

Abstract

The invention provides a driver distraction behavior recognition method based on a deep learning optimization YOLO fusion network model. The method comprises the following steps: constructing a vehicle driver distraction behavior standard image data set G20 in a highway scene; a deep learning-based driver distraction behavior recognition method in an expressway scene is researched, three models of YOLOv5, YOLOv8 and YOLOv11 for driver distraction behavior recognition are constructed, and the comprehensive detection performance of the models is analyzed and compared; the method comprises the following steps: constructing driver distraction behavior recognition models YOLOv11n-ECA, YOLOv11n-BiFPN-GLAS and YOLOv11n-SIoU based on an optimization method, and analyzing the improvement advantages and comprehensive detection performance of the driver distraction behavior recognition models; the improved methods are fused to construct an optimized fusion network YOLOv11n-Driver, and the test accuracy, the test recall rate, the F1-Score, the average precision and the consumed time of each image on a constructed G20 data set respectively reach 93.1%, 95.4%, 94.2%, 98.3% and 3.4 ms / image, so that a feasible method is provided for optimizing the recognition performance of a YOLO model on driver distraction behaviors in an expressway scene.

Description

Technical Field

[0001] The present invention belongs to the research fields of intelligent transportation and autonomous driving, and specifically relates to a method for identifying driver distraction behavior based on an optimized YOLO fusion network model. Background Art

[0002] Drivers who are on their phones, chatting with passengers, or otherwise distracted while driving increase the likelihood of a traffic accident. Studies have found that drivers are four times more likely to be involved in an accident while on their phones. Other distractions also significantly increase the likelihood of an accident. Currently, detecting distracted drivers and alerting them promptly is a major challenge in traffic management.

[0003] Traffic management on highways primarily relies on manual patrols and traditional monitoring, but this approach has its drawbacks: it doesn't detect incidents in a timely manner and its monitoring coverage is limited. When traffic volume is high and the environment is complex, existing traffic management methods are unable to effectively handle sudden violations. Traditional machine learning methods, such as support vector machines and the K-nearest neighbor algorithm, are computationally simple and easy to implement. However, when faced with complex data and small sample sizes, these methods are ineffective and cannot meet the requirements of real-time recognition in traffic scenarios. Deep learning differs from traditional methods in that it offers significant advantages in image and pattern recognition, particularly in scenarios like autonomous driving. Deep learning algorithms can automatically optimize the process of feature extraction and data reconstruction, resulting in more accurate recognition, higher learning efficiency, and applicable to more complex traffic scenarios and tasks. Using deep learning to identify driver distraction can accurately detect distracted behavior through surveillance video, enabling intelligent warning systems to provide timely alerts, fundamentally reducing the likelihood of traffic accidents and improving traffic management. Existing methods for identifying driver distraction have shortcomings, leading to the increasing adoption of convolutional neural networks in this area. The YOLO series of target detection algorithms has performed well in the field of computer vision. Their speed and accuracy make them particularly well-suited for real-time identification of driver distraction. The YOLO model can instantly identify and locate targets in videos, offering high speed and accuracy, addressing the inaccurate recognition issues faced by traditional methods in complex traffic environments. Therefore, integrating multiple improved methods to construct a YOLO fusion model for driver distraction identification is a viable approach for improving driver distraction identification performance. This paper proposes a driver distraction identification method based on an optimized YOLO fusion network model. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of the existing technology, a method for identifying driver distraction behavior based on an optimized YOLO fusion network model is provided. The method is based on a deep learning fusion network, integrates multiple optimization methods, and constructs a YOLO fusion model to perform high-precision identification of driver distraction behavior, providing certain technical guarantees for the safety of drivers in highway scenarios.

[0005] Technical Solution: To achieve the above objectives, the present invention provides a method for identifying driver distraction behavior based on an optimized YOLO fusion network model, comprising the following steps:

[0006] S1: Construct a standard image dataset G20 of vehicle driver distraction behavior in highway scenes;

[0007] S2: Research on the ECA attention mechanism to enhance the network's ability to extract key features;

[0008] S3: Research on Bidirectional Feature Pyramid Network (BiFPN) for optimizing feature fusion and weighted processing;

[0009] S4: Study the global local spatial attention (GLSA) module to improve model robustness and accuracy;

[0010] S5: Study the SIoU loss function for enhancing positioning accuracy and robustness;

[0011] S6: Build an optimization-based driver distraction behavior recognition model, YOLOv11n-ECA, and analyze its improved advantages and comprehensive detection performance.

[0012] S7: Build an optimization-based driver distraction recognition model, YOLOv11n-BiFPN-GLAS, and analyze its improved advantages and comprehensive detection performance.

[0013] S8: Build an optimization-based driver distraction recognition model, YOLOv11n-SIoU, and analyze its improved advantages and comprehensive detection performance.

[0014] S9: The above-mentioned improved methods are integrated based on rules to construct the optimized fusion network YOLOv11n-Driver and perform driver distraction behavior recognition.

[0015] Furthermore, the specific method for constructing the driver distraction behavior standard image dataset in step S1 is as follows:

[0016] S1-1: For the G20 Highway Driver Distraction Behavior Dataset, image data was collected from surveillance cameras along the G20 Qingyin Expressway in November 2024, with consistent resolution across all images. This dataset covers six different driver behavior categories: head-turn glance, in-vehicle ingestion behavior, smoking, head-down behavior, phone use, and normal driving. After data cleaning and secondary review, invalid or unclassifiable images were deleted, resulting in the construction of the Driver Distraction Behavior Dataset. This dataset contains approximately 150,000 valid images. However, due to the particularity of highway sections, drivers may actively regulate their driving behavior when they are aware of the presence of monitoring. This results in fewer instances of driver distraction in these images, and the number of distraction images in the original dataset is relatively limited. Subsequently, distraction samples were supplemented through data augmentation, and the dataset was finally randomly divided into training, validation, and test sets in a ratio of 7:2:1. The specific number of categories is shown in Table 1.

[0017] Table 1 Statistics of driver distraction behavior image datasets

[0018]

[0019] Furthermore, the specific contents of the ECA attention mechanism used to enhance the key feature extraction capability in step S2 are as follows:

[0020] S2-1: The ECA attention mechanism is an efficient and simplified attention method that can effectively improve the model's ability to identify target areas by optimizing the cross-channel information interaction process. The ECA mechanism only focuses on the interaction between each channel and its k neighboring channels, avoiding the complex calculation of all channels in traditional methods. This allows ECA to improve the model's feature extraction capabilities and enhance the model's discriminability without adding additional computational costs.

[0021] S2-2: Embedding the ECA attention mechanism into the YOLOv11 model's head improves the model's performance in complex backgrounds. In particular, it enhances the model's ability to focus on key driver action areas (such as the driver's arms and head outline) and important features (such as the contact point between the palm and the steering wheel). These features play a crucial role in identifying driver distraction. Furthermore, the ECA mechanism suppresses interference from complex background areas or irrelevant objects, further improving recognition capabilities.

[0022] Furthermore, the specific contents of the Bidirectional Feature Pyramid Network (BiFPN) used to optimize feature fusion and weighted processing in step S3 are as follows:

[0023] S3-1: The Bidirectional Feature Pyramid Network (BiFPN) is a highly efficient feature fusion architecture that overcomes the limitations of the traditional FPN's unidirectional information flow through a bidirectionally connected, cross-scale feature fusion network. BiFPN forms a closed-loop structure through top-down and bottom-up cross-connection paths, enabling more efficient information flow and feature interaction. This not only improves the fusion capabilities of multi-scale features but also enhances the model's expressiveness.

[0024] S3-2: Embedding the BiFPN module into the YOLOv11n model significantly improves the network's ability to weightedly fuse features at different levels. Each fusion module weights the input features based on their importance, automatically learning and determining the weights for each layer. This allows for more intelligent utilization of multi-scale feature information and improves feature extraction. This weighted fusion mechanism is crucial for multi-dimensional feature optimization, particularly in object detection tasks in complex backgrounds. It effectively improves model performance and reduces interference from irrelevant background information.

[0025] Furthermore, the specific contents of the global local spatial attention (GLSA) module used to improve the robustness and accuracy of the model in step S4 are as follows:

[0026] S4-1: The GLSA attention mechanism is a global-local spatial feature aggregation module that adaptively assigns feature weights at both global and local scales, thereby enhancing the model's focus on key driver behaviors. By integrating global and local features, GLSA can highlight the driver's behavioral characteristics in complex backgrounds while suppressing interference from irrelevant background information, significantly improving the model's detection accuracy.

[0027] S4-2: Combining the GLSA attention mechanism with the BiFPN mechanism overcomes BiFPN's limited receptive field in complex backgrounds or with stacked objects, improving the model's ability to recognize specific target features. The BiFPN-GLSA fusion YOLOv11 model performs well in complex object detection tasks while maintaining lightweight, high real-time performance, and enhanced robustness. The GLSA mechanism effectively enhances global feature extraction, improves focus on regional features of the driver, and reduces interference from background noise, further optimizing the recognition of local details.

[0028] Furthermore, the specific content of the SIoU loss function used to enhance positioning accuracy and robustness in step S5 is:

[0029] S5-1: The calculation formula of SIoU loss is:

[0030] The four components of the loss function are distance cost Δ, shape cost Ω, and overlap area cost (IoU). SIoU incorporates the vector angle information between the predicted and true bounding boxes, allowing for improved positioning accuracy and robustness by taking into account the scale and angle of the bounding box when calculating IoU. When the area of ​​the target box is correlated with the SIoU value, this ensures a more stable training process.

[0031] S5-2: During training, network performance is optimized by minimizing the SIoU loss to achieve more accurate object localization. The SIoU loss function consists of four functions: angle cost, distance cost, shape cost, and overlap area cost. It aims to control the loss in each aspect, thereby improving the accuracy of box-to-box matching. This loss function helps predict boxes quickly adjust during optimization, speeding up training and enhancing robustness.

[0032] Furthermore, in step S6, an optimization-based driver distraction recognition model, YOLOv11n-ECA, was constructed. The specific improvement method analyzed is as follows: Based on the original YOLOv11 model, the Backbone layer embeds an Efficient Channel Attention (ECA) mechanism to improve the network's feature extraction capabilities. The ECA module is added after the convolutional layer. This efficient channel attention mechanism enhances the network's focus on important features while reducing computational complexity. Compared to the traditional SE module, ECA does not rely on global pooling operations, improving computational efficiency. This results in the construction of a YOLOv11n-ECA model for driver distraction recognition.

[0033] Furthermore, in step S7, a driver distraction behavior recognition model YOLOv11n-BiFPN-GLAS based on the optimization method is constructed. The specific method of its improvement is analyzed as follows: Based on the original YOLOv11 model, BiFPN (Bidirectional Feature Pyramid Network) and GLSA (Global Local Attention Spatial) are introduced in the Neck part to significantly enhance the fusion of multi-scale features and the spatial attention mechanism. The GLSA module is immediately after the Backbone output and performs global and local attention adjustments on the feature map before feature fusion. BiFPN is located between the feature fusion layers, immediately after the Backbone feature output and the GLSA module. Its function is to efficiently fuse features from different scales through a bidirectional information flow (top-down and bottom-up), thereby constructing a YOLOv11n-BiFPN-GLAS network for driver distraction behavior recognition.

[0034] Furthermore, in step S8, an optimization-based driver distraction behavior recognition model, YOLOv11n-SIoU, is constructed. The specific improvement method is analyzed as follows: Based on the original YOLOv11 model, SIoULoss (Scalable Intersection over Union Loss) is introduced to further enhance the model's positioning accuracy and robustness. Compared with the traditional IoU loss function, the SIoU loss function introduces the angle information between the predicted box and the true box, enabling the model to consider the scale, angle, and shape of the bounding box in addition to calculating the IoU. This improves the accuracy and stability of the bounding box regression in object detection, thereby constructing a YOLOv11n-SIoU network for driving posture behavior recognition.

[0035] Furthermore, in step S9, all improved methods are integrated to construct an optimized YOLO fusion network model, YOLOv11n-Driver. The specific method is to integrate the improved methods of the ECA attention mechanism, the bidirectional feature pyramid network (BiFPN), the GLSA attention mechanism, and the SIoU loss function to form the YOLOv11n-Driver driver distraction behavior recognition fusion model. The YOLOv11n-Driver fusion model workflow is as follows: the input image first passes through the Backbone. The Backbone improves the attention mechanism, focusing on key image areas such as the driver's face and hands, suppressing irrelevant information, and enhancing feature expression. The processed features enter the Neck. The Neck has an improved BiFPN-GLSA structure with bottom-up and top-down connections, fusing features of different scales, preserving details, and improving the ability to detect objects of different sizes. At the same time, the BiFPN uses a weighted strategy to assign weights according to the importance of feature layers to optimize the fusion effect. The multi-scale features are then output and input into the detection head. The detection head has a feature enhancement module to strengthen features related to driver behavior. At the same time, the SIoU loss function is used to further optimize the model, optimize the center point positioning capability of the anchor frame, accelerate the convergence and training speed of the network, accurately identify the driver's gestures, posture and other behaviors, and output the detection results.

[0036] Beneficial effects: Compared with the existing technology, the present invention realizes the automatic recognition of driver distraction behavior based on deep learning optimized YOLO fusion network model technology, improves the comprehensive detection performance such as recognition accuracy and model real-time performance, and provides certain technical guarantees for the active safety of drivers while driving in highway scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Schematic diagram of the structure of the YOLOv11n-Driver fusion model.

[0038] Figure 2 This is an image example from the G20 dataset.

[0039] Figure 3 Implement a flowchart for the ECA attention mechanism.

[0040] Figure 4 This is an example of the bidirectional feature pyramid network process.

[0041] Figure 5 Implementing a flowchart for the GLSA attention mechanism.

[0042] Figure 6 This is the prediction confusion matrix result diagram of the YOLOv11n-Driver fusion model on the G20 dataset. DETAILED DESCRIPTION

[0043] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0044] like Figure 1 As shown, the present invention provides a method for identifying driver distraction behavior based on an optimized YOLO fusion network model, comprising the following steps:

[0045] S1: Construct a standard image dataset G20 of vehicle driver distraction behavior in highway scenes, such as Figure 2 As shown, the specific ones are:

[0046] S1-1: For the G20 Highway Driver Distraction Behavior Dataset, image data was collected from surveillance cameras along the G20 Qingyin Expressway in November 2024, with consistent resolution across all images. This dataset covers six different driver behavior categories: head-turn glance, in-vehicle ingestion behavior, smoking, head-down behavior, phone use, and normal driving. After data cleaning and secondary review, invalid or unclassifiable images were removed to construct the Driver Distraction Behavior Dataset. This dataset contains approximately 150,000 valid images. However, due to the unique nature of highways, drivers may actively regulate their driving behavior when aware of the surveillance cameras. This results in fewer instances of driver distraction in these images, resulting in a relatively limited number of distraction images in the original dataset. Data augmentation was subsequently used to supplement these distraction samples, and the dataset was randomly divided into training, validation, and test sets in a 7:2:1 ratio.

[0047] S2: Research on the ECA attention mechanism used to enhance the key feature extraction capability of the model. The structure example is as follows Figure 3 As shown, the specific ones are:

[0048] S2-1: The ECA attention mechanism is an efficient and simplified attention method that can effectively improve the model's ability to identify target areas by optimizing the cross-channel information interaction process. The ECA mechanism only focuses on the interaction between each channel and its k neighboring channels, avoiding the complex calculation of all channels in traditional methods. This allows ECA to improve the model's feature extraction capabilities and enhance the model's discriminability without adding additional computational costs.

[0049] S2-2: Embedding the ECA attention mechanism into the YOLOv11 model's head improves the model's performance in complex backgrounds. In particular, it enhances the model's ability to focus on key driver action areas (such as the driver's arms and head outline) and important features (such as the contact point between the palm and the steering wheel). These features play a crucial role in identifying driver distraction. Furthermore, the ECA mechanism suppresses interference from complex background areas or irrelevant objects, further improving recognition capabilities.

[0050] S3: Research on Bidirectional Feature Pyramid Network (BiFPN) for optimizing feature fusion and weighted processing. Its structure example is as follows Figure 4 As shown, the specific ones are:

[0051] S3-1: The Bidirectional Feature Pyramid Network (BiFPN) is a highly efficient feature fusion architecture that overcomes the limitations of the traditional FPN's unidirectional information flow through a bidirectionally connected, cross-scale feature fusion network. BiFPN forms a closed-loop structure through top-down and bottom-up cross-connection paths, enabling more efficient information flow and feature interaction. This not only improves the fusion capabilities of multi-scale features but also enhances the model's expressiveness.

[0052] S3-2: Embedding the BiFPN module into the YOLOv11n model significantly improves the network's ability to weightedly fuse features at different levels. Each fusion module weights the input features based on their importance, automatically learning and determining the weights for each layer. This allows for more intelligent utilization of multi-scale feature information and improves feature extraction. This weighted fusion mechanism is crucial for multi-dimensional feature optimization, particularly in object detection tasks in complex backgrounds. It effectively improves model performance and reduces interference from irrelevant background information.

[0053] S4: Research on the global local spatial attention GLSA module for improving model robustness and accuracy. Its structure example is as follows Figure 5 As shown, the specific ones are:

[0054] S4-1: The GLSA attention mechanism is a global-local spatial feature aggregation module that adaptively assigns feature weights at both global and local scales, thereby enhancing the model's focus on key driver behaviors. By integrating global and local features, GLSA can highlight the driver's behavioral characteristics in complex backgrounds while suppressing interference from irrelevant background information, significantly improving the model's detection accuracy.

[0055] S4-2: Combining the GLSA attention mechanism with the BiFPN mechanism overcomes BiFPN's limited receptive field in complex backgrounds or with stacked objects, improving the model's ability to recognize specific target features. The BiFPN-GLSA fusion YOLOv11 model performs well in complex object detection tasks while maintaining lightweight, high real-time performance, and enhanced robustness. The GLSA mechanism effectively enhances global feature extraction, improves focus on regional features of the driver, and reduces interference from background noise, further optimizing the recognition of local details.

[0056] S5: Study the SIoU loss function for enhancing positioning accuracy and robustness, which is specifically:

[0057] S5-1: The calculation formula of SIoU loss is:

[0058] The four components of the loss function are distance cost Δ, shape cost Ω, and overlap area cost (IoU). SIoU incorporates the vector angle information between the predicted and true bounding boxes, allowing for improved positioning accuracy and robustness by taking into account the scale and angle of the bounding box when calculating IoU. When the area of ​​the target box is correlated with the SIoU value, this ensures a more stable training process.

[0059] S5-2: During training, network performance is optimized by minimizing the SIoU loss to achieve more accurate object localization. The SIoU loss function consists of four functions: angle cost, distance cost, shape cost, and overlap area cost. It aims to control the loss in each aspect, thereby improving the accuracy of box-to-box matching. This loss function helps predict boxes quickly adjust during optimization, speeding up training and enhancing robustness.

[0060] S6: Build the YOLOv11n-ECA model for driver distraction recognition based on optimization methods. Analyze the specific improvements: Based on the original YOLOv11 model, the Backbone layer embeds the Efficient Channel Attention (ECA) mechanism to improve the network's feature extraction capabilities. Adding the ECA module after the convolutional layer uses an efficient channel attention mechanism to enhance the network's focus on important features while reducing computational complexity. Compared to the traditional SE module, ECA does not rely on global pooling operations, improving computational efficiency. This results in the YOLOv11n-ECA model for driver distraction recognition.

[0061] S7: Constructing a YOLOv11n-BiFPN-GLAS model for driver distraction recognition based on optimization methods. The specific improvements are analyzed as follows: Based on the original YOLOv11 model, the Neck layer introduces BiFPN (Bidirectional Feature Pyramid Network) and GLSA (Global Local Attention Spatial) to significantly enhance multi-scale feature fusion and spatial attention mechanisms. The GLSA module follows the Backbone output and performs global and local attention adjustments on the feature map before feature fusion. The BiFPN is located between the feature fusion layers, immediately following the Backbone feature output and the GLSA module. Its function is to efficiently fuse features from different scales through a bidirectional information flow (top-down and bottom-up), thereby constructing the YOLOv11n-BiFPN-GLAS network for driver distraction recognition.

[0062] S8: Build an optimization-based model for driver distraction recognition, YOLOv11n-SIoU. Analyze the specific improvements: Based on the original YOLOv11 model, SIoU Loss (Scalable Intersection over Union Loss) is introduced to further enhance the model's positioning accuracy and robustness. Compared to the traditional IoU loss function, the SIoU loss function incorporates the angle between the predicted and true bounding boxes, enabling the model to consider the scale, angle, and shape of the bounding box in addition to calculating the IoU. This improves the accuracy and stability of bounding box regression in object detection, thereby constructing a YOLOv11n-SIoU network for driving posture recognition.

[0063] S9: All improved methods are integrated to construct an optimized YOLO fusion network model, YOLOv11n-Driver. The specific method is to fuse the improved ECA attention mechanism, Bidirectional Feature Pyramid Network (BiFPN), GLSA attention mechanism, and SIoU loss function to form the YOLOv11n-Driver fusion model for driver distraction behavior recognition. The YOLOv11n-Driver fusion model workflow is as follows: the input image first passes through the Backbone. Backbone's improved attention mechanism can focus on key image areas, such as the driver's face and hands, suppress irrelevant information, and enhance feature expression. The processed features are then fed into the Neck. The Neck has an improved BiFPN-GLSA structure with bottom-up and top-down connections, which fuses features of different scales, preserves details, and improves the ability to detect objects of different sizes. Simultaneously, BiFPN uses a weighted strategy to assign weights based on feature layer importance to optimize the fusion effect. The multi-scale features are then output and fed into the detection head, which includes a feature enhancement module to enhance features related to driver behavior. At the same time, the SIoU loss function is used to further optimize the model, optimize the center point positioning capability of the anchor frame, accelerate the convergence and training speed of the network, accurately identify the driver's gestures, posture and other behaviors, and output the detection results.

[0064] To validate the superiority of each fusion model, this case study conducted comparative experiments. YOLO models were selected for comparative experiments on the G20 driver distraction behavior dataset. YOLOv5n, YOLOv8n, and YOLOv11n models were constructed for driver distraction behavior recognition. Based on the YOLOv11 model, which achieved the best overall detection performance, the YOLOv11n model was optimized using the proposed ECA attention module, Bidirectional Feature Pyramid Network (BiFPN), GLSA attention mechanism, and SIoU loss function. These models were then constructed as YOLOv11n-ECA, YOLOv11n-BiFPN-GLAS, and YOLOv11n-SIoU models for driver distraction behavior recognition. The overall experimental results for all models are shown in Table 2. As can be seen from the table, while each model achieved good training results on both the training and validation sets, YOLOv5n's detection accuracy for some behaviors was low, and its generalization ability was poor. The combined results of the three models show particularly poor recall for In-Vehicle Ingestion Behavior, indicating limited recognition capabilities for this behavior and requiring further optimization. YOLOv11 performed the best of the three models, achieving a recognition rate of 0.963. YOLOv8 achieved a moderate recognition rate, while YOLOv5 achieved a recognition rate of 0.943. While the overall accuracy rates are similar, small inaccuracies can impact traffic safety, so the model with the highest accuracy must be selected for optimization. The model with the highest recognition accuracy was YOLOv11n, achieving over 90% accuracy for all five categories of driver distraction. However, classification errors were primarily concentrated in In-Vehicle Ingestion Behavior, which is highly similar to other distractions such as Head-Down Behavior. Furthermore, the overall recall loss is reflected in a small number of images not detected, significantly impacting the model's average recognition accuracy. Therefore, further dataset and model optimization are needed.

[0065] When the ECA attention mechanism is used alone to optimize the Backbone network structure, the model's precision P increases by 3.4%, the recall R increases by 1.4%, and the average accuracy increases by 1.0%. The improvement in precision indicates that it effectively reduces the number of false positives and allows the model to identify more important features. The slight increase in recall R also shows that the attention module may not directly solve the problem of missed detection, and its optimization effect is relatively limited, but it can still improve the classification accuracy.

[0066] When BiFPN-GLAS is used alone to replace Neck to optimize the PANet structure and other modules, the model's precision P is improved by 3.1%, the recall rate R is improved by 0.4%, and the average accuracy is improved by 1.2%. This shows that the bidirectional fusion of the BiFPN module can effectively enhance the small target detection effect. At the same time, the spatial attention optimization of GLSA can reduce the occurrence of false recognition, but it also increases the amount of calculation. In the performance of small data sets, it may not be possible to fully reflect the enhanced effect of the BiFPN module, which is related to the data set situation in this article.

[0067] When only SIoU is used to optimize the original loss function, the overall model's computational complexity does not change much. The model's precision P increases by 0.3%, the recall R increases by 1.8%, and the average accuracy increases by 0.7%. Since the change in the loss function module does not change the structure of the model, it has almost no effect on the model's computational complexity. It only speeds up the network's convergence and training speed, improving detection accuracy. At the same time, due to the characteristics of the SIoU loss function, it adds the calculation of angle costs and can optimize the center point positioning through its dynamic distance loss without increasing the amount of computation. This is why we can see a significant improvement in the recall R, effectively reducing the problem of partial occlusion and unrecognition, and solving the problem of missed detection of some targets. However, it may also slightly affect the probability of missed detection, resulting in a limited improvement in accuracy.

[0068] Table 2 Comparative experimental results of three selected YOLO series models and a single optimization model

[0069]

[0070] In order to illustrate the optimization effect of the above optimization method on the YOLOv11 model, an ablation experiment was also conducted. The confusion matrix results of the YOLOv11n-Driver fusion model are as follows: Figure 6The overall experimental results are shown in Table 3. As can be seen, the introduction of these three improvements leads to significant improvements in precision, recall, F1-Score, mAP, and Tims compared to the original model. Further analysis of Table 3 shows that the simultaneous introduction of the ECA attention mechanism and the BiFPN-GLSA module, along with the improved loss function, significantly improves model performance, with precision, recall, and mAP increasing by 2.1%, 4.9%, and 2.0%, respectively, while maintaining stable FPS. This synergistic effect stems from the complementary functions between the modules: the ECA-optimized channel features provide cleaner input for BiFPN-GLSA, reducing noise interference with multi-scale fusion. The spatial attention and bidirectional fusion mechanisms of BiFPN-GLSA enhance feature discrimination in complex scenes, laying the foundation for accurate SIoU regression. SIoU, through optimization of the loss function, further amplifies the localization potential of multi-scale features, particularly for detecting small and occluded objects. This effectively improves the model's detection accuracy and compensates for some previously missed and false detections. The optimal balance between model size and accuracy was achieved, and the expected experimental results were achieved. In summary, the fusion model has good detection performance in driving distraction behavior detection and meets the improvement target requirements.

[0071] Table 3 Ablation experiment results

[0072]

[0073] The technical means disclosed in the solutions of the present invention are not limited to those disclosed in the above-mentioned embodiments, but also include technical solutions composed of any combination of the above-mentioned technical features. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A driver distraction behavior recognition method based on an optimized YOLO fusion network model, characterized by: The steps include: S1: Construct a standard image dataset G20 of vehicle driver distraction behavior in highway scenes; S2: Research on the ECA attention mechanism to enhance the network's ability to extract key features; S3: Research on Bidirectional Feature Pyramid Network (BiFPN) for optimizing feature fusion and weighted processing; S4: Study the global local spatial attention (GLSA) module to improve model robustness and accuracy; S5: Study the SIoU loss function for enhancing positioning accuracy and robustness; S6: Build an optimization-based driver distraction behavior recognition model, YOLOv11n-ECA, and analyze its improved advantages and comprehensive detection performance. S7: Build an optimization-based driver distraction recognition model, YOLOv11n-BiFPN-GLAS, and analyze its improved advantages and comprehensive detection performance. S8: Build an optimization-based driver distraction recognition model, YOLOv11n-SIoU, and analyze its improved advantages and comprehensive detection performance. S9: The above-mentioned improved methods are integrated based on the fusion rules to construct the optimized fusion network YOLOv11n-Driver and perform driver distraction behavior recognition.

2. The method for identifying driver distracted behavior based on the optimized YOLO fusion network model according to claim 1, characterized in that: The specific method of constructing the vehicle driver distraction behavior dataset in the highway scene in step S1 is as follows: S1-1: Image data is collected from surveillance cameras along the G20 Qingyin Expressway. All images maintain consistent resolution. This dataset covers six different driver behavior categories, including head-on-side turning, eating in the car, smoking, head-down, mobile phone use, and normal driving. After data cleaning and secondary review, invalid or unclassifiable images were deleted to construct a dataset of driver distraction behaviors. The dataset was randomly divided into training, validation, and test sets in a ratio of 7:2:

1.

3. The method for identifying driver distraction based on the optimized YOLO fusion network model according to claim 1, characterized in that: The specific contents of the ECA attention mechanism used to enhance the network's key feature extraction capability in step S2 are as follows: S2-1: The ECA attention mechanism improves the model's ability to identify target regions by optimizing cross-channel information interaction. The ECA mechanism focuses only on the interaction between each channel and its k neighboring channels, enabling ECA to improve the model's feature extraction capabilities and enhance the model's discriminability without incurring additional computational costs. S2-2: The ECA attention mechanism is embedded into the YOLOv11 model head to improve the model's performance in complex backgrounds. In particular, it enhances the model's ability to focus on the driver's key action areas and important features, which play a decisive role in identifying driver distraction. At the same time, the ECA mechanism can suppress interference from complex background areas or irrelevant objects, improving recognition capabilities.

4. The method for identifying driver distracted behavior based on an optimized YOLO fusion network model according to claim 1, characterized in that: The specific contents of the bidirectional feature pyramid network BiFPN used to optimize feature fusion and weighted processing in step S3 are as follows: S3-1: The Bidirectional Feature Pyramid Network (BiFPN) is a network that fuses cross-scale features through bidirectional connections. BiFPN forms a closed-loop structure through top-down and bottom-up cross-connection paths, enabling information flow and feature interaction to improve the fusion of multi-scale features and enhance the model's expressiveness. S3-2: The BiFPN module is embedded in the YOLOv11n model to improve the network's ability to weightedly fuse features at different levels. Each fusion module performs weighted processing based on the importance of the input features, automatically learning and determining the weights of features at each layer, intelligently utilizing multi-scale feature information to improve feature extraction capabilities. In the task of object detection in complex backgrounds, the performance of the model is improved and the interference of irrelevant background information is reduced.

5. The method for identifying driver distraction based on the optimized YOLO fusion network model according to claim 1, characterized in that: The specific contents of the global local spatial attention (GLSA) module used to improve the robustness and accuracy of the model in step S4 are as follows: S4-1: The GLSA attention mechanism adaptively allocates feature weights at both global and local scales, enhancing the model's focus on key driver behaviors. By integrating global and local features, GLSA can highlight driver behavior characteristics in complex environments while suppressing interference from irrelevant background information, improving the model's detection accuracy. S4-2: Combining the GLSA attention mechanism with the BiFPN mechanism can overcome the limitation of BiFPN's limited receptive field in complex backgrounds or when objects are stacked, and improve the model's ability to recognize specific target features. The YOLOv11 model after BiFPN-GLSA fusion performs well in complex target detection tasks while maintaining lightweight and high real-time performance, and has stronger robustness. The GLSA mechanism can enhance the extraction of global features, improve the focus on the driver's regional features, reduce the interference of background noise, and optimize the recognition ability of local details.

6. The method for identifying driver distraction based on an optimized YOLO fusion network model according to claim 1, characterized in that: The specific content of the SIoU loss function used to enhance positioning accuracy and robustness in step S5 is: S5-1: The calculation formula of SIoU loss is: Among them, the distance cost Δ, shape cost Ω and overlap area cost IoU are the four components of the loss function. SIoU can consider the scale and angle of the bounding box when calculating IoU by adding the vector angle information between the predicted box and the real box, thereby improving the positioning accuracy and robustness. When the area relationship of the target box is correlated with the SIoU value, it can ensure a more stable training process. S5-2: During the training process, the network performance is optimized by minimizing the SIoU loss to achieve more accurate target positioning. The SIoU loss function consists of four functions: angle cost, distance cost, shape cost, and overlapping area cost. It aims to control the loss in each aspect, thereby improving the matching accuracy between boxes. This loss function can help the predicted box to adjust quickly during the optimization process, improve training speed, and enhance robustness.

7. The method for identifying driver distraction based on an optimized YOLO fusion network model according to claim 1, characterized in that: In step S6, a YOLOv11n-ECA model for driver distraction recognition based on an optimization method is constructed. The specific improvement method is analyzed as follows: based on the original YOLOv11 model, the Backbone part improves the network's feature extraction capability by embedding the ECA mechanism; an ECA module is added after the convolutional layer, and the network's focus on important features is enhanced through an efficient channel attention mechanism, while reducing computational complexity; ECA does not rely on global pooling operations, which improves computational efficiency, thereby constructing a YOLOv11n-ECA model for driver distraction recognition.

8. The method for identifying driver distraction based on an optimized YOLO fusion network model according to claim 1, characterized in that: In step S7, a YOLOv11n-BiFPN-GLAS model for driver distraction recognition based on an optimization method is constructed. The specific improvement method is analyzed as follows: Based on the original YOLOv11 model, BiFPN and GLSA are introduced in the Neck part to significantly enhance the fusion of multi-scale features and the spatial attention mechanism; the GLSA module is immediately after the Backbone output and performs global and local attention adjustments on the feature map before feature fusion; BiFPN is located between the feature fusion layers, immediately after the Backbone feature output and the GLSA module, and its function is to efficiently fuse features from different scales through a bidirectional information flow, thereby constructing a YOLOv11n-BiFPN-GLAS network for driver distraction recognition.

9. The method for identifying driver distracted behavior based on an optimized YOLO fusion network model according to claim 1, characterized in that: In step S8, a YOLOv11n-SIoU model for driver distraction behavior recognition based on an optimization method is constructed. The specific improvement method is analyzed as follows: based on the original YOLOv11 model, SIoU Loss is introduced to improve the positioning accuracy and robustness of the model; the SIoU loss function introduces the angle information between the predicted box and the true box, enabling the model to consider the scale, angle, and shape of the bounding box on the basis of calculating the IoU, thereby improving the accuracy and stability of the box regression in object detection, thereby constructing a YOLOv11n-SIoU network for driving posture behavior recognition.

10. The method for identifying driver distracted behavior based on an optimized YOLO fusion network model according to claim 1, characterized in that: In step S9, all the improved methods are integrated based on the fusion rules to construct the optimized YOLO fusion network model YOLOv11n-Driver. The specific method is: the ECA attention mechanism, the bidirectional feature pyramid network BiFPN, the GLSA attention mechanism and the improved method of the SIoU loss function are integrated to form the YOLOv11n-Driver driver distraction behavior recognition fusion model; the specific content of the fusion rule is: the YOLOv11n-Driver fusion model workflow is as follows: the input image first passes through Backbone, Backbone improves the attention mechanism, can focus on the key areas of the image, suppress The processed features enter the Neck part, which has an improved BiFPN-GLSA structure with bottom-up and top-down connections, fusing features of different scales, retaining details, and improving the ability to detect targets of different sizes. At the same time, BiFPN uses a weighted strategy to assign weights according to the importance of the feature layer to optimize the fusion effect. Then, the multi-scale features are output and input into the detection head. The detection head has a feature enhancement module to strengthen the features related to driver behavior. At the same time, the SIoU loss function is used to further optimize the model, optimize the center point positioning capability of the anchor frame, accelerate the convergence and training speed of the network, accurately identify the driver's gestures and posture behaviors, and output the detection results.

Images