A driver behavior detection method based on YOLO11n-LDS

Abstract

The application discloses a driver behavior detection method based on YOLO11n-LDS, and relates to the technical field of intelligent cockpit driving safety monitoring, and comprises the following steps: S1, collecting driver images and performing pretreatment; S2, inputting a feature map into a lossless spatial down-sampling module to generate a high-fidelity feature map; S3, inputting the high-fidelity feature map into a direction feature aggregation module to generate a multi-scale discriminant feature map; S4, inputting the multi-scale discriminant feature map into a spatial channel interactive attention module to generate a re-calibration feature map; and S5, performing feature fusion on the re-calibration feature map to generate a multi-scale fusion feature, inputting the multi-scale fusion feature into different scale detection heads, and outputting a driver behavior detection result. The driver behavior detection method based on YOLO11n-LDS solves the problems that the prior art has insufficient small-scale behavior recognition capability, limited expression of direction-sensitive features, and difficulty in balancing detection precision and calculation efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent cockpit driving safety monitoring technology, and in particular to a driver behavior detection method based on YOLO11n-LDS. Background Technology

[0002] With the rapid development of intelligent transportation systems and vehicle-mounted visual perception technology, deep learning-based methods for driver behavior recognition and fatigue detection have been widely researched and applied. These methods typically utilize in-vehicle cameras to capture real-time images of the driver's face, head, and hands. Through neural network models, they automatically identify and warn of distracting behaviors (such as using a mobile phone, drinking water, or operating the central control system) and fatigue states (such as yawning, closing eyes, or nodding), which is of great significance for improving driving safety.

[0003] Existing technologies mostly employ lightweight object detection networks (such as the YOLO series and SSD) to achieve real-time driver behavior recognition. These networks, through their compact backbone structure and efficient feature pyramids, ensure high detection speeds while meeting the computational resource constraints of in-vehicle embedded devices. However, in high-speed driving scenarios, unsafe driver behaviors often manifest as brief distractions or momentary fatigue reactions, such as quickly looking down at a phone, briefly closing the eyes, or yawning. These behaviors are characterized by subtle features, short duration, and significant changes in direction, making it difficult for detection models to consistently capture effective discriminative information. This can easily lead to false positives (misclassifying normal behavior as dangerous behavior) or false negatives (failing to identify genuinely dangerous behavior).

[0004] Meanwhile, traditional convolutional neural networks typically employ multiple stride convolutions or pooling operations for downsampling when processing image features to expand the receptive field and reduce computational cost. However, this repeated downsampling process leads to the gradual loss or blurring of key small-scale features in the image (such as the opening and closing of the eyes, subtle mouth movements, etc.). Furthermore, most existing convolutional structures are based on isotropic convolutional kernels, lacking explicit orientation modeling capabilities, making it difficult to effectively distinguish behaviors with similar textures but different orientations (e.g., looking left versus looking right, making a phone call versus drinking water, etc.). These limitations collectively result in decreased detection accuracy and insufficient system stability in complex driving environments (such as varying lighting, partial occlusion, and diverse poses). Summary of the Invention

[0005] The purpose of this invention is to provide a driver behavior detection method based on YOLO11n-LDS, which solves the problems of insufficient small-scale behavior recognition capability, limited expression of directional sensitive features, and difficulty in balancing detection accuracy and computational efficiency in existing technologies.

[0006] To achieve the above objectives, this invention provides a driver behavior detection method based on YOLO11n-LDS, comprising the following steps: S1. Acquire driver images and preprocess them to generate input feature maps and construct driver behavior datasets; S2. Input the input feature map into the lossless spatial downsampling module. Through a 4-branch parallel spatial rearrangement with a step size of 2, adjacent spatial pixels are reorganized into the channel dimension. While reducing the spatial resolution, the number of channels is expanded to 4 times, generating a high-fidelity feature map. S3. Input the high-fidelity feature map into the orientation feature aggregation module. The orientation feature aggregation module contains three parallel asymmetric branches. The outputs of each branch are spliced ​​by channels and connected by residuals to generate a multi-scale discriminative feature map. S4. Input the multi-scale discriminative feature map into the spatial channel interactive attention module. The spatial channel interactive attention module performs local average pooling and global average pooling simultaneously. Perform one-dimensional adaptive kernel convolution on the two pooling results respectively. After weighted fusion, multiply with the multi-scale discriminative feature map to generate a recalibrated feature map. S5. Perform multi-scale feature fusion on the recalibrated feature map, input the generated multi-scale fused features into different scale detection heads, and finally output the driver behavior detection results.

[0007] Preferably, when acquiring driver images in S1, the images are acquired from multiple angles from different perspectives of the driver using a camera.

[0008] Preferably, the preprocessing in S1 includes removing a large number of duplicate images, filtering blurry images, and labeling the effective images with behavior categories. The labeled driver behaviors include safe driving, making a phone call, yawning, and closing the eyes.

[0009] Preferably, the specific process for generating a high-fidelity feature map is as follows: S21. Spatial rearrangement of the input feature map is performed through 4 downsampling branches with a step size of 2, and the pixels of adjacent spatial positions are reorganized into the channel dimension, while expanding the number of channels to 4 times the original. S22. By performing convolution operations, the rearranged features are fused and compressed, spatial information is transformed into channel feature representation, and a high-fidelity feature map is output, thus completing the lossless transfer of spatial information to channel information.

[0010] Preferably, the specific process for generating multi-scale discriminative feature maps is as follows: S31. Input the high-fidelity feature map into the directional feature aggregation module, and generate shared features through 3×3 convolution processing; S32, the shared features are respectively passed through three parallel asymmetric branches, namely branch one Branch 2 and branch three Each branch outputs multi-scale spatial features in parallel, namely spatial feature one, spatial feature two, and spatial feature three; S33. Spatial feature one is processed by the activation function, then concatenated with spatial feature two and spatial feature three through a channel concatenation, and then subjected to a 1×1 convolution operation to generate a multi-scale concatenated feature map. S34. Multi-scale spliced ​​feature maps and high-fidelity feature maps are added element-wise through residual connections to generate multi-scale discriminative feature maps.

[0011] Preferably, the three branches in S32 are as follows: Branch 1 is a 3×3 convolution used to capture local detail information and output spatial feature 1; Branch 2 is a 5×5 convolution used to extract mesoscale context information and output spatial feature 2; Branch 3 is a 7×7 convolution used to extract spatial structure information and output spatial feature 3.

[0012] Preferably, the specific process for generating the recalibrated feature map is as follows: S41. Input the multi-scale discriminative feature map into the spatial channel interaction attention module, and simultaneously perform local average pooling and global average pooling to output local pooling features and global pooling features respectively; retain certain spatial structure information through local average pooling, and extract overall semantic information through global average pooling. S42. Perform one-dimensional convolution operations on the local pooling features and the global pooling features respectively to generate local attention features and global attention features respectively; S43. Weighted fusion of local attention features and global attention features, and then multiplying with the multi-scale discriminative feature map to generate a recalibrated feature map.

[0013] Preferably, the expression for generating the recalibrated feature map is: ; In the formula, Indicates the first Recalibrated feature maps at multiple scales; Feature map representing the input space channel interaction attention module; Indicates the feature map scale index; For element-wise multiplication; This represents a one-dimensional adaptive kernel convolution; Indicates global average pooling; This indicates local pooling.

[0014] Preferably, the specific process for generating multi-scale fusion features and outputting driver behavior detection results is as follows: S51. By performing upsampling, downsampling, and concatenation operations on recalibrated feature maps of different resolutions sequentially through multi-scale feature fusion, shallow detail information and deep semantic information are preserved to generate multi-scale fused features; the expression is: ; In the formula, This represents the multi-scale fusion features of the input detection head; Indicates the first Recalibrated feature maps at multiple scales; This indicates feature map splicing; S52. Input the multi-scale fused features into detectors of different scales, and each detector outputs the target category and bounding box position at the corresponding scale; the expression is: ; In the formula, Indicates the first Detection results at various scales; Indicates the operation of the detection head; This represents the multi-scale fusion features of the input detection head; S53. Perform post-processing on the output of all detection heads to obtain the final recognition result of the driver's behavior.

[0015] Preferably, the post-processing operation in S53 includes confidence filtering and non-maximum suppression (NMS), which removes redundant candidate boxes and maps the final detection result to the original image to achieve driver behavior and fatigue state recognition. S531. Perform confidence screening: Calculate the final detection confidence score for the behavior category. And set a confidence threshold. ;like Then retain the behavior category, if Then delete the behavior category; the expression for confidence filtering is: ; In the formula, This indicates the probability of a behavior category existing; Represents the probability of a behavior category; S532. When multiple bounding boxes exist for the same behavior category, calculate the intersection-union ratio of the bounding boxes. Set the non-maximum suppression threshold. Keep the bounding box of the highest score; if Then retain the bounding box with higher confidence; if This indicates that the targets are different behavior categories, so all bounding boxes are retained; the expression for calculating the intersection-union ratio of the bounding boxes is: ; In the formula, and These represent the first bounding box and the second bounding box, respectively. This represents the area of ​​the overlapping region between two bounding boxes; This represents the area of ​​the union region of the two bounding boxes.

[0016] Therefore, the driver behavior detection method based on YOLO11n-LDS described above has the following beneficial effects: (1) Significantly improves detection accuracy: This invention achieves excellent detection performance on a self-built driver behavior dataset through the collaborative design of a lossless spatial downsampling module (LSDConv), a directional feature aggregation module (DFA), and a spatial channel interaction attention module (SCIA). Experimental results show that the YOLO11-LDS model proposed in this invention achieves an accuracy of 94.27%, a recall of 92.05%, a mAP@0.5 of 97.65%, and a mAP@0.5:0.95 of 78.72%, all of which are superior to the baseline model and single-module variants, effectively reducing the false detection and false negative rates. (2) Enhance the ability to identify small-scale and direction-sensitive behaviors: The lossless spatial downsampling module transfers spatial information to the channel dimension in the early stage of feature extraction, avoiding the loss of key small-scale features in the traditional downsampling process; the direction feature aggregation module explicitly models the direction features of behavior through multi-scale parallel branches and residual connections, which can effectively distinguish behaviors with similar textures but different directions (such as making a phone call and drinking water, turning the head left and right, etc.), significantly improving the behavior discrimination ability in complex driving scenarios; (3) Keep the model lightweight and suitable for vehicle-mounted embedded deployment: The spatial channel interaction attention module adopts a combination of local and global pooling and one-dimensional convolution to achieve feature recalibration, which significantly improves the feature expression capability with low computational overhead; while improving the detection accuracy, the overall model maintains the lightweight characteristics of parameter quantity and computational quantity, which can meet the real-time requirements of vehicle-mounted equipment. (4) Enhance environmental robustness and generalization ability: This invention constructs a training set by collecting data from multiple angles and perspectives, and combines directional feature aggregation and attention mechanism to enable the model to have stronger adaptability to complex driving environments such as different lighting, posture, and occlusion; ablation experiments show that the performance is optimal when the three modules work together, which reflects the robustness improvement brought about by collaborative optimization. (5) Forming a collaborative processing flow of spatial information preservation, orientation information enhancement and channel attention optimization: This invention integrates lossless downsampling, orientation-aware modeling and channel interactive attention in the YOLO framework for the first time. Each module is passed layer by layer through feature maps, realizing a complete optimization chain from information fidelity to orientation enhancement and then to attention recalibration, which solves the problems of feature loss, insufficient orientation expression and difficulty in balancing computational efficiency in the prior art.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of a driver behavior detection method based on YOLO11n-LDS according to the present invention; Figure 2 This is a diagram of the overall architecture of the YOLO11n-LDS of this invention; Figure 3 This is a structural diagram of the lossless spatial downsampling module according to an embodiment of the present invention; Figure 4 This is a structural diagram of the directional feature aggregation module according to an embodiment of the present invention; Figure 5 This is a structural diagram of the spatial channel interaction attention module according to an embodiment of the present invention. Detailed Implementation

[0019] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0020] Please see Figures 1-2 A driver behavior detection method based on YOLO11n-LDS includes a lossless downsampling module, a direction-aware feature aggregation module, a spatial channel interaction attention module, and a detection head. The specific steps are as follows: S1. Collect driver images from multiple angles, including front, side and other perspectives, using mobile phone cameras to improve data diversity and model robustness. Preprocess the driver images to generate input feature maps and construct a driver behavior dataset containing various behavior categories such as safe driving, making phone calls, yawning, and closing eyes. The driver behavior dataset contains approximately 15,000 images. Preprocessing includes removing a large number of duplicate images, filtering blurry images, and labeling effective images with behavior categories, including driver behaviors such as safe driving, making a phone call, yawning, and closing eyes. S2, such as Figure 3 As shown, the input feature map enters the lossless spatial downsampling module, where spatial rearrangement is performed to generate a high-fidelity feature map; S21. Spatial rearrangement of the input feature map through 4 downsampling branches, with a step size of 2 for each downsampling branch, reorganizes the pixels of adjacent spatial positions into the channel dimension, and expands the number of channels to 4 times the original. S22. By performing convolution operations, the rearranged features are fused and compressed, spatial information is transformed into channel feature representation, and a high-fidelity feature map is output. This reduces spatial resolution without losing information, and completes the lossless transfer of spatial information to channel information. S3, such as Figure 4 As shown, the high-fidelity feature map is input into the directional feature aggregation module to enhance the ability to determine the direction of behavior and generate a multi-scale discriminative feature map. S31. Input the high-fidelity feature map into the directional feature aggregation module, and generate shared features through 3×3 convolution processing; S32, shared features are respectively through three branches, including branch one Branch 2 and branch three Each branch outputs multi-scale spatial features in parallel, including spatial feature one, spatial feature two, and spatial feature three. Branch 1 is a 3×3 convolution used to capture local detail information and outputs spatial feature 1; Branch 2 is a 5×5 multi-scale residual block used to extract mesoscale context information and outputs spatial feature 2; Branch 3 is a 7×7 multi-scale residual block used to extract spatial structure information over a larger range and outputs spatial feature 3. S33. Spatial feature one, after being processed by the activation function, is combined with spatial feature two and spatial feature three through channel concatenation and convolution to generate a multi-scale spliced ​​feature map. S34. Multi-scale spliced ​​feature maps and high-fidelity feature maps are combined through residual connections to generate multi-scale discriminative feature maps; S4, such as Figure 5As shown, the multi-scale discriminative feature map is input into the spatial channel interactive attention module for feature recalibration to generate a recalibrated feature map. S41. Input the multi-scale discriminative feature map into the spatial channel interaction attention module, and simultaneously perform local average pooling and global average pooling to output local pooling features and global pooling features respectively; retain certain spatial structure information through local average pooling, and extract overall semantic information through global average pooling. S42. Perform one-dimensional convolution operations on the local pooling features and the global pooling features respectively. convolution, Representing dimension, (representing adaptive convolutional kernel size), modeling the dependencies between channels, and generating local attention features and global attention features respectively; S43. Weighted fusion of local attention features and global attention features, and then multiplication with multi-scale discriminative feature map to generate recalibrated feature map, thereby achieving feature enhancement; The expression for generating the recalibrated feature map is: ; In the formula, Indicates the first Recalibrated feature maps at multiple scales; Feature map representing the input space channel interaction attention module; Indicates the feature map scale index; For element-wise multiplication; This represents a one-dimensional adaptive kernel convolution; Indicates global average pooling; This indicates local pooling; S5. Perform feature fusion on the recalibrated feature map to generate multi-scale fusion features, input them into different scale detection heads, and output the driver behavior detection results; S51. By performing upsampling, downsampling, and concatenation operations on recalibrated feature maps of different resolutions sequentially through multi-scale feature fusion, shallow detail information and deep semantic information are preserved to generate multi-scale fused features; the expression is: ; In the formula, This represents the multi-scale fusion features of the input detection head; Indicates the first Recalibrated feature maps at multiple scales; This indicates feature map splicing; S52. Input the multi-scale fused features into detectors of different scales, and each detector outputs the target category and bounding box position at the corresponding scale; the expression is: ; In the formula, Indicates the first Detection results at various scales; This indicates the operation of the detection head.

[0021] S53. Perform post-processing operations on the output of all detection heads. The post-processing operations include confidence filtering and non-maximum suppression (NMS). Redundant candidate boxes are removed through post-processing operations, and the final detection results are mapped to the original image to realize driver behavior and fatigue state recognition. S531. Perform confidence screening: Calculate the final detection confidence score for the behavior category. And set a confidence threshold. , It is generally set to 0.25; if Then retain the behavior category, if Then delete the behavior category; the expression for confidence filtering is: ; In the formula, This indicates the probability of a behavior category existing; for example, the probabilities of drinking water or smoking are 0.91 and 0.09, respectively. This represents the probability of a behavior category, such as 0.88 for drinking water or 0.12 for smoking; S532. When multiple bounding boxes exist for the same behavior category, calculate the intersection-union ratio of the bounding boxes. Set the non-maximum suppression threshold. , It is generally set to 0.5 to retain the bounding box of the highest score; if Then retain the bounding box with the higher confidence level. For example, if the confidence results of the first and second bounding boxes are respectively... , At that time, the result of its intersection and union ratio is If the non-maximum suppression is greater than the threshold of 0.5, then the confidence level is higher. The bounding box; if This indicates that the targets are different behavior categories, so all bounding boxes are retained; the expression for calculating the intersection-union ratio of the bounding boxes is: ; In the formula, and These represent the first bounding box and the second bounding box, respectively. This represents the area of ​​the overlapping region between two bounding boxes; This represents the area of ​​the union region of the two bounding boxes.

[0022] Experimental Verification 1 To verify the effectiveness of the YOLO11n-LDS (an improved YOLO11n model integrating lossless downsampling, directional feature aggregation, and attention mechanisms) of this invention, experiments were conducted using a driver behavior dataset. The results showed a precision of 94.27%, a recall of 92.05%, a mean accuracy (mAP@0.5, where IoU threshold is 0.5) of 97.65%, and a mean accuracy (mAP@0.5:0.95, where IoU threshold is between 0.5 and 0.95) of 78.72%. These experimental results demonstrate that this invention achieves high-precision driver behavior detection while maintaining a lightweight design.

[0023] Experimental Verification 2 To verify the contributions of each module, several variants were designed for ablation experiments, including YOLO11n (the YOLO11n baseline model), YOLO11n-LSDConv (a YOLO11n model incorporating a lossless spatial downsampling module), YOLO11n-DFA (a YOLO11n model incorporating a directional feature aggregation module), YOLO11n-SCIA (a YOLO11n model incorporating a spatial channel interaction attention module), YOLO11n-LSDConv+DFA (a YOLO11n model fusing lossless downsampling and directional feature aggregation modules), and YOLO11n-LDS (the model of this invention). The experimental results are shown in Table 1. The ablation experiments show that all three modules contribute positively to the final performance, and the combination of the three achieves the best detection accuracy, proving the effectiveness of the collaborative optimization design of this invention.

[0024] Table 1: Ablation Experiment Results

[0025] Experimental Verification 3 To verify the applicability of different attention mechanisms in driver behavior and fatigue detection tasks, comparative experiments were conducted using the Cross-Attention Fusion Module (CAFM), the Simple Parametric Attention Module (SimAM), the EfficientChannel, and the Spatial Channel Interaction Attention Module (SCIA) proposed in this invention. See Table 2. The results show that different attention mechanisms can enhance feature representation capabilities to some extent. CAFM achieved better recall, indicating its good response capability to the target region; SimAM performed better in precision, demonstrating its strong feature discrimination capability; EfficientChannel has lower structural complexity, but its overall detection performance is relatively limited. In contrast, the SCIA module proposed in this invention, by jointly modeling local spatial information and global channel dependencies, can more effectively enhance the representation capability of driver facial regions, eye states, and orientation-related behavioral features. In orientation-sensitive scenarios such as eye-closed detection, gaze deviation, and lateral behavior recognition, SCIA can more accurately focus on key regions while maintaining low computational overhead and good real-time detection efficiency, making it more suitable for fatigue and dangerous behavior detection tasks in complex driving environments.

[0026] Table 2

[0027] Therefore, this invention adopts a driver behavior detection method based on YOLO11n-LDS, which achieves excellent performance of 94.27% precision, 92.05% recall, and 97.65% mAP@0.5 on a self-built driver behavior dataset through the collaborative design of a lossless spatial downsampling module, a directional feature aggregation module, and a spatial channel interaction attention module. This significantly improves the recognition accuracy of small-scale and direction-sensitive behaviors, while maintaining the model's lightweight design to adapt to in-vehicle embedded deployment and enhancing robustness in complex driving environments. It forms a complete processing flow of spatial information preservation, directional information enhancement, and channel attention optimization, effectively solving the problems of feature loss, insufficient directional expression, and difficulty in balancing detection accuracy and computational efficiency in existing technologies.

[0028] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A driver behavior detection method based on YOLOv11n-LDS, characterized in that, Includes the following steps: S1. Acquire driver images and preprocess them to generate input feature maps and construct driver behavior datasets; S2. Input the input feature map into the lossless spatial downsampling module. Through a 4-branch parallel spatial rearrangement with a step size of 2, adjacent spatial pixels are reorganized into the channel dimension. While reducing the spatial resolution, the number of channels is expanded to 4 times, generating a high-fidelity feature map. S3. Input the high-fidelity feature map into the orientation feature aggregation module. The orientation feature aggregation module contains three parallel asymmetric branches. The outputs of each branch are spliced ​​by channels and connected by residuals to generate a multi-scale discriminative feature map. S4. Input the multi-scale discriminative feature map into the spatial channel interactive attention module. The spatial channel interactive attention module performs local average pooling and global average pooling simultaneously. Perform one-dimensional adaptive kernel convolution on the two pooling results respectively. After weighted fusion, multiply with the multi-scale discriminative feature map to generate a recalibrated feature map. S5. Perform multi-scale feature fusion on the recalibrated feature map, input the generated multi-scale fused features into different scale detection heads, and finally output the driver behavior detection results.

2. The driver behavior detection method based on YOLO11n-LDS according to claim 1, characterized in that: When capturing images of the driver in S1, the camera captures images from multiple angles from different perspectives of the driver.

3. The driver behavior detection method based on YOLO11n-LDS according to claim 2, characterized in that: Preprocessing in S1 includes removing a large number of duplicate images, filtering blurry images, and labeling effective images with behavioral categories. The labeled driver behaviors include safe driving, making a phone call, yawning, and closing eyes.

4. The driver behavior detection method based on YOLO11n-LDS according to claim 3, characterized in that, The specific process of generating high-fidelity feature maps is as follows: S21. Spatial rearrangement of the input feature map is performed through 4 downsampling branches with a step size of 2, and the pixels of adjacent spatial positions are reorganized into the channel dimension, while expanding the number of channels to 4 times the original. S22. By performing convolution operations, the rearranged features are fused and compressed, spatial information is transformed into channel feature representation, and a high-fidelity feature map is output, thus completing the lossless transfer of spatial information to channel information.

5. The driver behavior detection method based on YOLO11n-LDS according to claim 4, characterized in that, The specific process of generating multi-scale discriminative feature maps is as follows: S31. Input the high-fidelity feature map into the directional feature aggregation module, and generate shared features through 3×3 convolution processing; S32, the shared features are respectively passed through three parallel asymmetric branches, namely branch one Branch 2 and branch three Each branch outputs multi-scale spatial features in parallel, namely spatial feature one, spatial feature two, and spatial feature three; S33. Spatial feature one is processed by the activation function, then concatenated with spatial feature two and spatial feature three through a channel concatenation, and then subjected to a 1×1 convolution operation to generate a multi-scale concatenated feature map. S34. Multi-scale spliced ​​feature maps and high-fidelity feature maps are added element-wise through residual connections to generate multi-scale discriminative feature maps.

6. The driver behavior detection method based on YOLO11n-LDS according to claim 5, characterized in that, The three branches in S32 are as follows: Branch 1 is a 3×3 convolution used to capture local detail information and output spatial feature 1; Branch 2 is a 5×5 convolution used to extract mesoscale context information and output spatial feature 2; Branch 3 is a 7×7 convolution used to extract spatial structure information and output spatial feature 3.

7. The driver behavior detection method based on YOLO11n-LDS according to claim 6, characterized in that, The specific process of generating the recalibrated feature map is as follows: S41. Input the multi-scale discriminative feature map into the spatial channel interactive attention module, and simultaneously perform local average pooling and global average pooling, and output the local pooling feature and global pooling feature respectively. Spatial structure information is preserved by local average pooling, and overall semantic information is extracted by global average pooling. S42. Perform one-dimensional convolution operations on the local pooling features and the global pooling features respectively to generate local attention features and global attention features respectively; S43. Weighted fusion of local attention features and global attention features, and then multiplying with the multi-scale discriminative feature map to generate a recalibrated feature map.

8. The driver behavior detection method based on YOLO11n-LDS according to claim 7, characterized in that, The expression for generating the recalibrated feature map is: ； In the formula, Indicates the first Recalibrated feature maps at multiple scales; Feature map representing the input space channel interaction attention module; Indicates the feature map scale index; For element-wise multiplication; This represents a one-dimensional adaptive kernel convolution; Indicates global average pooling; This indicates local pooling.

9. The driver behavior detection method based on YOLO11n-LDS according to claim 8, characterized in that, The specific process of generating multi-scale fusion features and outputting driver behavior detection results is as follows: S51. By using a multi-scale feature fusion method, upsampling, downsampling, and stitching operations are performed sequentially on the recalibrated feature maps of different resolutions to retain shallow detail information and deep semantic information, thereby generating multi-scale fused features. The expression is: ； In the formula, This represents the multi-scale fusion features of the input detection head; Indicates the first Recalibrated feature maps at multiple scales; This indicates feature map splicing; S52. Input the multi-scale fused features into detectors of different scales. Each detector outputs the behavior category probability, behavior category confidence, and bounding box coordinates at the corresponding scale; the expression is: ； In the formula, Indicates the first Detection results at various scales; Indicates the operation of the detection head; This represents the multi-scale fusion features of the input detection head; S53. Perform post-processing on the output of all detection heads to obtain the final recognition result of the driver's behavior.

10. The driver behavior detection method based on YOLO11n-LDS according to claim 9, characterized in that: The post-processing operations in S53 include confidence filtering and non-maximum suppression (NMS). Redundant candidate boxes are removed through post-processing operations, and the final detection results are mapped to the original image to realize driver behavior and fatigue state recognition. S531. Perform confidence screening: Calculate the final detection confidence score for the behavior category. And set a confidence threshold. ;like Then retain the behavior category, if Then delete the behavior category; the expression for confidence filtering is: ； In the formula, This indicates the probability of a behavior category existing; Represents the probability of a behavior category; S532. When multiple bounding boxes exist for the same behavior category, calculate the intersection-union ratio of the bounding boxes. Set the non-maximum suppression threshold. Keep the bounding box of the highest score; if Then retain the bounding box with higher confidence; if This indicates that the targets are different behavior categories, so all bounding boxes are retained; the expression for calculating the intersection-union ratio of the bounding boxes is: ； In the formula, and These represent the first bounding box and the second bounding box, respectively. This represents the area of ​​the overlapping region between two bounding boxes; This represents the area of ​​the union region of the two bounding boxes.

Images