A method for constructing a bad driving behavior detection model, a detection method and system

Abstract

The application discloses a bad driving behavior detection model, method and system, and belongs to the technical field of intelligent transportation and computer vision. The model is based on YOLOv11, the backbone network adopts a ghost convolution and a variable kernel double-branch module to extract features, combines fast spatial pyramid pooling and a partial spatial attention module to enhance multi-scale context expression, a coordinate attention module is introduced into a feature fusion network to strengthen key area position perception, and a three-scale decoupling structure is adopted in a detection head to separate a boundary box regression and a category prediction task. The application effectively reduces model parameter quantity and calculation cost, improves detection accuracy and robustness in complex scenes such as black night and occlusion, and is suitable for real-time driver behavior monitoring and early warning of vehicle embedded devices.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation, vehicle safety assistance, driver behavior monitoring, and computer vision target detection, specifically to the YOLOv11-Ghost-CA behavior detection method and system. Background Technology

[0002] Poor driving behavior is a significant contributing factor to road traffic accidents. When drivers engage in distracted activities such as making or receiving phone calls, sending messages, drinking beverages, adjusting in-vehicle devices, retrieving items from the back seat, fixing their hair, applying makeup, or conversing with other passengers, their attention shifts from the road environment to these other actions, easily reducing vehicle control and risk response capabilities. Therefore, how to accurately identify poor driving behavior in real time and provide timely warnings is a crucial issue for intelligent transportation systems and in-vehicle driver assistance systems.

[0003] With the development of deep learning and computer vision technologies, image-based driving behavior detection methods are increasingly being applied to driver monitoring scenarios. The YOLO (You Only Look Once) series of models are single-stage object detection models, possessing advantages such as compact structure, fast detection speed, and strong end-to-end output capabilities, making them suitable for real-time detection tasks. YOLOv11 further optimizes the backbone network, feature fusion network, and detection head structure based on YOLOv8, achieving strong real-time inference capabilities while maintaining high accuracy. It is widely used in tasks such as industrial defect detection, intelligent transportation, and driver behavior monitoring. However, the complete YOLOv11 network structure still faces the following three problems when deployed on automotive embedded devices: First, the backbone network has a large number of conventional convolutional and bottleneck modules, resulting in a large number of model parameters and computational load, which may lead to excessive inference latency or excessive storage consumption in the vehicle terminal; Second, the feature fusion network has a relatively simple fusion method for features at different scales and fails to provide sufficient spatial location enhancement for key areas such as driver hand movements and head posture; Third, the detection head couples target location prediction and category prediction in the same branch, and the optimization objectives of the two tasks are prone to mutual interference, affecting the final detection accuracy and convergence speed.

[0004] Existing lightweight object detection techniques typically reduce resource consumption by decreasing the number of network layers, replacing lightweight convolutions, or compressing model parameters. However, model compression alone may lead to a decrease in feature representation capabilities. Existing attention mechanisms can enhance feature representation in key regions, but directly adding attention modules may introduce additional computation or cause excessive sensitivity to noise in complex scenes. Furthermore, existing detection heads often employ a coupled design at the P3, P4, and P5 scales, failing to decouple the object localization task from the category judgment task, thus hindering the full utilization of multi-scale feature discrimination capabilities. Therefore, a method for detecting undesirable driving behavior that balances model lightweighting, feature representation capabilities, and robustness to complex scenes is needed, enabling it to operate in in-vehicle embedded environments while maintaining high detection accuracy in dark and occluded scenarios. Summary of the Invention

[0005] To address the problems of existing driving behavior detection models, such as large parameter count, difficulty in embedded deployment, impact on accuracy due to coupled detection head design, and unstable detection performance in dark or occluded scenarios, this invention provides a method for constructing a bad driving behavior detection model, a detection method, and a system.

[0006] The method for constructing a bad driving behavior detection model provided in this invention is based on the YOLOv11 network and includes a backbone network, a feature fusion network, and a decoupled detection head.

[0007] The backbone network input is a ghost convolution module, which is sequentially connected to four sets of alternating ghost convolution modules and variable kernel dual-branch modules to extract features at different depths. Finally, it is connected to a fast spatial pyramid pooling module and a partial spatial attention module to obtain deep features containing multi-scale context and key regions. The second set of alternating ghost convolution modules and variable kernel dual-branch modules extracts the basic features of the driver image, and the third set extracts the complex features of the driver image.

[0008] In the feature fusion network, the deep features are upsampled and then concatenated with complex features. The resulting fusion features are then passed through a variable kernel dual-branch module to form a first fusion feature. This first fusion feature is then concatenated with basic features through an upsampled module and passed through a variable kernel dual-branch module to obtain a first output feature. This first output feature is then concatenated with the first fusion feature through a coordinate attention module and a ghost convolution module, and then passed through a variable kernel dual-branch module to obtain a second output feature. The second output feature is then concatenated with the input deep features through a coordinate attention module and a ghost convolution module, and then passed through a variable kernel dual-branch module to obtain a third output feature.

[0009] The decoupled detection head includes three-scale bounding box regression branches and category prediction branches, which predict the target location and the category of bad driving behavior based on the three output features of the feature fusion network.

[0010] Preferably, the ghost convolution module includes: performing a basic convolution on the input feature map to generate an intrinsic feature map with a small number of core channels; performing an identity mapping on the intrinsic feature map; and performing a depthwise convolution and / or linear transformation on the intrinsic feature map to generate a ghost feature map; concatenating the identity mapping result and the ghost feature map with the intrinsic feature map along the channel dimension to obtain an output feature map; the linear transformation includes depthwise separable convolution, channel-wise convolution, pixel convolution, or a combination thereof.

[0011] Preferably, the variable kernel dual-branch module includes: performing a 1×1 convolution on the input feature map and then dividing it into a bypass branch and a deep feature extraction branch; the deep feature extraction branch selects a bottleneck module or a variable kernel module according to the c3k parameter to perform N feature extractions; the output of the bypass and the output of the deep feature extraction branch are concatenated, and then fused by a 1×1 convolution before output; where N is an integer greater than or equal to 1; the bypass is used to retain the original input information.

[0012] Preferably, the fast spatial pyramid pooling module includes: performing a 1×1 convolution on the input feature map and then dividing it into an original scale branch and three sequential 5×5 max pooling branches; concatenating the output of the original scale branch and the output of the three pooling branches into four paths, and then merging them through a 1×1 convolution before outputting the result; gradually expanding the receptive field through three consecutive 5×5 max pooling operations and aggregating contextual information at different scales.

[0013] Preferably, the partial spatial attention module includes: after performing convolutional streaming on the input feature map, it is divided into a bypass and a partial spatial attention (PSA) block branch; the PSA block branch contains N serial PSA blocks, and each PSA block sequentially performs spatial attention calculation, residual summation, feedforward convolution, and residual summation; the bypass output and the PSA block branch output are concatenated, and then output after convolutional fusion; where N is an integer greater than or equal to 1; the bypass is used to retain channel features that have not entered the PSA attention calculation.

[0014] Preferably, the decoupled detection head includes detection branches at three scales: P3, P4, and P5. Each scale detection branch includes a bounding box regression branch and a class prediction branch. The bounding box regression branch sequentially performs convolution, depthwise convolution, and 1×1 convolution to output the x and y coordinates, width, and height of the target bounding box. The class prediction branch sequentially performs convolution, depthwise convolution, convolution, and 1×1 convolution to output the probability of the bad driving behavior category. The three scales, P3, P4, and P5, are responsible for detecting small-scale, medium-scale, and large-scale targets, respectively.

[0015] The present invention also provides a method for detecting unsafe driving behavior using the above-mentioned detection model, comprising the following steps:

[0016] Step 1: Obtain driver images or in-vehicle video frames and construct a dataset of bad driving behaviors that includes safe driving categories and at least one category of bad driving behaviors;

[0017] Step 2: Preprocess and divide the dataset of poor driving behavior to obtain a training set and a validation set, and generate at least one complex scenario validation set based on the validation set;

[0018] Step 3: Construct the aforementioned unsafe driving behavior detection model;

[0019] Step 4: Train the bad driving behavior detection model using the training set, and evaluate the trained model using the validation set and the complex scenario validation set;

[0020] Step 5: Input the driver image or in-vehicle video frame to be detected into the trained bad driving behavior detection model, and output the driving behavior category, confidence score and target location information;

[0021] Step 6: Based on the driving behavior category and confidence level, determine whether the driver has engaged in undesirable driving behavior, and generate a prompt or warning message when the warning conditions are met.

[0022] Preferably, the category of undesirable driving behavior includes at least one of the following: typing with the right hand, making a phone call with the right hand, typing with the left hand, making a phone call with the left hand, adjusting the radio, drinking a beverage, taking something from the back of the car, fixing hair and applying makeup, and speaking to other passengers.

[0023] Preferably, the complex scene verification set in step two includes a nighttime scene verification set and an occlusion scene verification set; the nighttime scene verification set is obtained by adjusting the brightness of the verification image to 0.3 times the original brightness; the occlusion scene verification set is obtained by generating an occlusion region with random position, random size and / or random color on the verification image, wherein the occlusion region is a rectangular region and the length and width of the rectangular region are randomly determined within the range of 50 pixels to 100 pixels.

[0024] The present invention also provides a system for detecting undesirable driving behavior using the above-described detection model, comprising:

[0025] Image acquisition module, used to acquire driver images or in-vehicle video frames;

[0026] The data processing module is used to construct a dataset of poor driving behavior, and to annotate, divide, and generate a validation set for complex scenarios from the dataset.

[0027] The model building module is used to build the aforementioned unsafe driving behavior detection model;

[0028] The model training module is used to train the bad driving behavior detection model and evaluate the trained model.

[0029] The behavior detection module is used to input the image to be detected into the trained bad driving behavior detection model and output the driving behavior category, confidence level and target location information.

[0030] The warning module is used to generate prompts or warning information based on the driving behavior category and confidence level.

[0031] The beneficial effects of this invention are:

[0032] 1. This invention splits the YOLOv11 network into three distinct segments: a backbone network, a feature fusion network, and a decoupled detection head. In the backbone network, ghost convolution modules, variable kernel dual-branch modules, fast spatial pyramid pooling modules, and partial spatial attention modules are introduced in sequence, making the network hierarchy clear and the module responsibilities well-defined, which facilitates module-level adjustment and optimization for different driving behavior scenarios.

[0033] 2. This invention uses a ghost convolution module to replace the conventional convolution in the backbone network. It generates eigenmaps through a small number of basic convolutions, and then generates ghost feature maps through identity mapping and depthwise convolution. This reduces redundant computation, lowers model size and computational resource requirements, and makes the model more suitable for deployment on in-vehicle embedded devices, edge computing terminals and intelligent traffic monitoring devices.

[0034] 3. This invention combines a variable kernel dual-branch module and a coordinate attention module in the feature fusion network. By using position encoding in the horizontal and vertical directions, it enhances the model's perception of key action areas, enabling the model to better distinguish between bad driving behaviors with similar local actions, such as making a phone call, typing, drinking a beverage, or taking items from the back seat.

[0035] 4. This invention introduces a partial spatial attention module into the backbone network. Through a dual-branch structure, the model performs PSA attention calculations only on a portion of the channels, while the remaining channels are retained through a bypass. This not only enhances the model's ability to focus on key regions but also avoids the additional overhead caused by introducing attention calculations on all channels.

[0036] 5. The present invention designs the detection head as a decoupled structure with three scales: P3, P4, and P5. The bounding box regression task and the category prediction task are completed on two independent branches. The two tasks are optimized independently and do not interfere with each other, which can improve the detection accuracy and convergence speed of the model on multi-scale targets.

[0037] 6. This invention verifies the model's recognition capability in complex in-vehicle scenarios by constructing a blackout scene verification set and an occlusion scene verification set, making the model training and evaluation process closer to the in-vehicle application environment.

[0038] 7. In one experimental embodiment, the YOLOv11-Ghost-CA model achieved an accuracy of 99.809%, a recall of 99.775%, and an mAP50-95 of 99.499% in normal scenarios, with a model size of 8.77MB. In simulated dark scenarios, the accuracy reached 99%, and in simulated occlusion scenarios, the accuracy reached 96.8%, indicating that this method can reduce the model size while maintaining high accuracy and improving recognition stability in complex scenarios. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall process of the YOLOv11-Ghost-CA behavior detection method of the present invention.

[0040] Figure 2 This is a schematic diagram of the overall structure of the YOLOv11-Ghost-CA bad driving behavior detection model of the present invention.

[0041] Figure 3 This is a schematic diagram of the ghost convolution module of the present invention.

[0042] Figure 4 This is a schematic diagram of the variable core dual-branch module of the present invention.

[0043] Figure 5 This is a schematic diagram of the structure of the fast spatial pyramid pooling module of the present invention.

[0044] Figure 6 This is a schematic diagram of the structure of part of the spatial attention module of the present invention.

[0045] Figure 7 This is a schematic diagram of the coordinate attention module of the present invention.

[0046] Figure 8 This is a schematic diagram of the decoupling detection head of the present invention.

[0047] Figure 9 This is a schematic diagram illustrating the process of constructing a dataset of poor driving behavior and simulating dark and occluded scenarios according to the present invention.

[0048] Figure 10 This is a schematic diagram of the structure of a system for detecting poor driving behavior. Detailed Implementation

[0049] The present invention will be further described below with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0050] Example 1

[0051] This embodiment constructs a dataset for detecting unsafe driving behavior based on the publicly available Kaggle State Farm Distracted Driver Detection dataset. The dataset is labeled with categories including safe driving and nine common unsafe driving behaviors: c0 safe driving, c1 typing with the right hand, c2 making a phone call with the right hand, c3 typing with the left hand, c4 making a phone call with the left hand, c5 adjusting the radio, c6 drinking a beverage, c7 reaching for items in the back seat, c8 fixing hair and applying makeup, and c9 speaking to other passengers. The dataset is divided into training and validation sets according to a preset ratio, and a data.yaml configuration file is generated.

[0052] Based on the validation set, two types of complex scene validation sets are constructed: the first is a nighttime scene validation set, where the brightness of the validation image is reduced to 0.3 times the original brightness; the second is an occlusion scene validation set, where rectangular occlusion regions with lengths and widths of 50 to 100 pixels are randomly generated on the validation image, and the colors of the rectangles are randomly generated. Through these methods, the robustness of the model under low-light and partial occlusion conditions can be validated without re-collecting a large amount of real nighttime and occlusion data.

[0053] Example 2

[0054] This embodiment constructs a YOLOv11-Ghost-CA model, the overall network consisting of three segments: a backbone network, a feature fusion network, and a decoupled detection head. The backbone network starts from the input image.

[0055] The backbone network input is a ghost convolution module, which is sequentially connected to four sets of alternating ghost convolution modules and variable kernel dual-branch modules to extract features at different depths. Finally, it is connected to a fast spatial pyramid pooling module and a partial spatial attention module to obtain deep features containing multi-scale context and key regions. The second set of alternating ghost convolution modules and variable kernel dual-branch modules extracts the basic features of the driver image, and the third set extracts the complex features of the driver image.

[0056] In the feature fusion network, the deep features are upsampled and then concatenated with complex features. The resulting fusion features are then passed through a variable kernel dual-branch module to form a first fusion feature. This first fusion feature is then concatenated with basic features through an upsampled module and passed through a variable kernel dual-branch module to obtain a first output feature. This first output feature is then concatenated with the first fusion feature through a coordinate attention module and a ghost convolution module, and then passed through a variable kernel dual-branch module to obtain a second output feature. The second output feature is then concatenated with the input deep features through a coordinate attention module and a ghost convolution module, and then passed through a variable kernel dual-branch module to obtain a third output feature.

[0057] The decoupled detection head includes bounding box regression branches and category prediction branches at three scales: P3, P4, and P5. These branches predict the target location and the category of bad driving behavior based on the three output features of the feature fusion network.

[0058] In the backbone network, the ghost convolution module employs a two-stage feature generation approach: the first stage uses a small number of basic convolutions to generate representative intrinsic feature maps; the second stage performs an identity mapping on the intrinsic feature maps to preserve core features, and performs depthwise convolutions and / or linear transformations on the remaining channels to generate ghost feature maps; finally, the identity mapping result and the ghost feature maps are concatenated with the intrinsic feature maps along the channel dimension as the output. The linear transformations include depthwise separable convolutions, channel-wise convolutions, pixel convolutions, or combinations thereof.

[0059] The variable kernel dual-branch module adopts a 1×1 convolutional flow structure: the input feature map is divided into a bypass and a deep feature extraction branch after 1×1 convolution; the deep feature extraction branch selects either a bottleneck module (c3k=False) or a variable kernel C3k module (c3k=True) according to the c3k parameter to perform N feature extractions; the bypass output and the deep feature extraction branch output are concatenated, and then fused by 1×1 convolution before output; where N is an integer greater than or equal to 1; the bypass is used to preserve the original input information and avoid the loss of key details in deep computation.

[0060] The fast spatial pyramid pooling module adopts a serial pooling structure: the input feature map is divided into an original scale branch and three serial 5×5 max pooling branches after 1×1 convolution; the output of the original scale branch and the output of the three pooling branches are concatenated in four ways, and then fused by 1×1 convolution before output; the receptive field is gradually expanded by three consecutive 5×5 max pooling, and contextual information of different scales is aggregated, so that the model has a stable feature response on targets of different scales.

[0061] The spatial attention module adopts a CSP (Cross Stage Partial) dual-branch + PSA block serial stacking structure: the input feature map is divided into bypass and PSA block branches after convolution and streaming; the PSA block branch contains N serial PSA blocks, and each PSA block performs spatial attention calculation, residual summation, feedforward convolution and residual summation in sequence; the bypass output and the PSA block branch output are concatenated and then fused by convolution before output; where N is an integer greater than or equal to 1; the bypass is used to retain channel features that have not entered the PSA attention calculation, so that the model can enhance the attention to key regions without significantly increasing the amount of computation.

[0062] In the feature fusion network, the coordinate attention module decomposes the two-dimensional global pooling into one-dimensional global pooling in both horizontal and vertical directions. The pooling results in the horizontal and vertical directions are concatenated, processed through a shared 1×1 convolution and activation function, and then split into two branches. These branches are then processed through 1×1 convolutions and a sigmoid function to obtain horizontal and vertical attention weights, respectively. Finally, these two attention weights are applied to the input feature map to obtain the enhanced output feature map. The coordinate attention module is placed after the P3 and P4 output layers to enhance the expressive power of small-to-medium scale driver action features and local key region features.

[0063] Example 3

[0064] This embodiment conducts ablation experiments on YOLOv11, YOLOv11-Ghost, YOLOv11-CA, and YOLOv11-Ghost-CA. Evaluation metrics include precision, recall, mAP50-95, and model size. The model scale is set to n, the number of training epochs is set to 200, and the input image size is set to 320.

[0065] In typical scenarios, the YOLOv11-Ghost-CA model achieves an accuracy of 99.809%, a recall of 99.775%, and an mAP50-95 of 99.499%, with a model size of 8.77MB. Compared to the original YOLOv11 model, YOLOv11-Ghost-CA demonstrates improved detection accuracy while reducing the model size from 9.97MB to 8.77MB. This indicates that the combination of ghost convolution and coordinate attention mechanisms can improve model recognition capabilities while reducing model size.

[0066] In a simulated dark environment, the original YOLOv11 model achieved a precision of 80.1%, a recall of 80.6%, and an mAP50-95 of 89.5%; the YOLOv11-Ghost-CA model achieved a precision of 99%, a recall of 98.5%, and an mAP50-95 of 99.5%. These results demonstrate that the model described in this invention maintains high detection stability under low-light conditions.

[0067] In simulated occlusion scenarios, the original YOLOv11 model achieved a precision of 88.2%, a recall of 86.6%, and an mAP50-95 score of 82.7%; the YOLOv11-Ghost-CA model achieved a precision of 96.8%, a recall of 97.3%, and an mAP50-95 score of 93.1%. These results demonstrate that the key local features preserved by the ghost convolution and the spatial location representation enhanced by the CA module work synergistically to improve the model's recognition ability under partial occlusion conditions.

[0068] Example 4

[0069] This embodiment deploys the trained YOLOv11-Ghost-CA model on an in-vehicle embedded terminal. An in-vehicle camera captures driver images in real time. The terminal resizes and normalizes the images before inputting them into the detection model. The detection model sequentially extracts multi-level features through the ghost convolution module, variable kernel dual-branch module, fast spatial pyramid pooling module, and partial spatial attention module in the backbone network. Then, it fuses features at different scales through a feature fusion network. Finally, it outputs bounding box regression results and the probability of undesirable driving behavior categories through P3, P4, and P5 three-scale decoupled detection heads. The model outputs the behavior category and confidence score. When multiple consecutive frames show the driver making a phone call, typing, drinking a beverage, adjusting the radio, taking items from the back seat, fixing their hair, or speaking to passengers, and the confidence score exceeds a preset threshold, the terminal provides a reminder via voice broadcast, in-vehicle pop-up window, or beeping, while simultaneously recording the event time, behavior category, and confidence score. This deployment method enables real-time monitoring of undesirable driver behavior without relying on high-performance servers.

[0070] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for constructing a detection model for unsafe driving behavior, characterized in that, The detection model is based on the YOLOv11 network and includes a backbone network, a feature fusion network, and a decoupled detection head. The backbone network input is a ghost convolution module, which is sequentially connected to four sets of alternating ghost convolution modules and variable kernel dual-branch modules to extract features at different depths. Finally, it is connected to a fast spatial pyramid pooling module and a partial spatial attention module to obtain deep features containing multi-scale context and key regions. The second set of alternating ghost convolution modules and variable kernel dual-branch modules extracts the basic features of the driver image, and the third set extracts the complex features of the driver image. The partial spatial attention module includes: after performing convolutional stream on the input feature map, it is divided into a bypass and a partial spatial attention block branch; the partial spatial attention block branch contains N serial partial spatial attention blocks, and each partial spatial attention block sequentially performs spatial attention calculation, residual summation, feedforward convolution, and residual summation; the bypass output and the partial spatial attention block branch output are concatenated, and then fused by convolution before output; where N is an integer greater than or equal to 1; the bypass is used to retain channel features that did not enter the partial spatial attention calculation; In the feature fusion network, the deep features are upsampled and then concatenated with complex features. The resulting fusion features are then passed through a variable kernel dual-branch module to form a first fusion feature. This first fusion feature is then concatenated with basic features through an upsampled module and passed through a variable kernel dual-branch module to obtain a first output feature. This first output feature is then concatenated with the first fusion feature through a coordinate attention module and a ghost convolution module, and then passed through a variable kernel dual-branch module to obtain a second output feature. The second output feature is then concatenated with the input deep features through a coordinate attention module and a ghost convolution module, and then passed through a variable kernel dual-branch module to obtain a third output feature. The decoupled detection head includes three-scale bounding box regression branches and category prediction branches, which predict the target location and the category of bad driving behavior based on the three output features of the feature fusion network.

2. The method for constructing a detection model for poor driving behavior according to claim 1, characterized in that, The ghost convolution module includes: performing basic convolution on the input feature map to generate an intrinsic feature map with a small number of core channels; performing an identity mapping on the intrinsic feature map; and performing depthwise convolution and / or linear transformation on the intrinsic feature map to generate a ghost feature map; concatenating the identity mapping result and the ghost feature map with the intrinsic feature map along the channel dimension to obtain an output feature map; the linear transformation includes depthwise separable convolution, channel-wise convolution, pixel convolution, or a combination thereof.

3. The method for constructing a detection model for undesirable driving behavior according to claim 1, characterized in that, The variable kernel dual-branch module includes: performing a 1×1 convolution on the input feature map and then dividing it into a bypass and a deep feature extraction branch; the deep feature extraction branch selects a bottleneck module or a variable kernel module according to the c3k parameter to perform N feature extractions; the output of the bypass and the output of the deep feature extraction branch are concatenated, and then fused by a 1×1 convolution before output; where N is an integer greater than or equal to 1; the bypass is used to retain the original input information.

4. The method for constructing a detection model for undesirable driving behavior according to claim 1, characterized in that, The fast spatial pyramid pooling module includes: performing a 1×1 convolution on the input feature map and then dividing it into an original scale branch and three sequential 5×5 max pooling branches; concatenating the output of the original scale branch and the output of the three pooling branches into four paths, and then merging them through a 1×1 convolution before outputting the result; gradually expanding the receptive field through three consecutive 5×5 max pooling operations and aggregating contextual information at different scales.

5. The method for constructing a detection model for poor driving behavior according to claim 1, characterized in that, The decoupled detection head includes detection branches at three scales: P3, P4, and P5. Each scale's detection branch includes a bounding box regression branch and a class prediction branch. The bounding box regression branch sequentially performs convolution, depthwise convolution, and 1×1 convolution to output the x and y coordinates, width, and height of the target bounding box. The class prediction branch sequentially performs convolution, depthwise convolution, convolution, and 1×1 convolution to output the probability of the category of poor driving behavior. The three scales, P3, P4, and P5, are responsible for detecting small-scale, medium-scale, and large-scale targets, respectively.

6. A method for detecting unsafe driving behavior using the detection model established according to claim 1, characterized in that, Includes the following steps: Step 1: Obtain driver images or in-vehicle video frames and construct a dataset of bad driving behaviors that includes safe driving categories and at least one category of bad driving behaviors; Step 2: Preprocess and divide the dataset of poor driving behavior to obtain a training set and a validation set, and generate at least one complex scenario validation set based on the validation set; Step 3: Construct the aforementioned unsafe driving behavior detection model; Step 4: Train the bad driving behavior detection model using the training set, and evaluate the trained model using the validation set and the complex scenario validation set; Step 5: Input the driver image or in-vehicle video frame to be detected into the trained bad driving behavior detection model, and output the driving behavior category, confidence score and target location information; Step 6: Based on the driving behavior category and confidence level, determine whether the driver has engaged in undesirable driving behavior, and generate a prompt or warning message when the warning conditions are met.

7. The detection method according to claim 6, characterized in that, The categories of undesirable driving behaviors include at least one of the following: typing with the right hand, making a phone call with the right hand, typing with the left hand, making a phone call with the left hand, adjusting the radio, drinking beverages, taking things from the back of the car, fixing hair and applying makeup, and speaking to other passengers.

8. The detection method according to claim 6, characterized in that, The complex scene verification set mentioned in step two includes a nighttime scene verification set and an occlusion scene verification set. The nighttime scene verification set is obtained by adjusting the brightness of the verification image to 0.3 times the original brightness. The occlusion scene verification set is obtained by generating an occlusion region with random position, random size and / or random color on the verification image. The occlusion region is a rectangular region, and the length and width of the rectangular region are randomly determined within the range of 50 pixels to 100 pixels.

9. A system for detecting unsafe driving behavior based on the detection model established in claim 1, characterized in that, include: Image acquisition module, used to acquire driver images or in-vehicle video frames; The data processing module is used to construct a dataset of poor driving behavior, and to annotate, divide, and generate a validation set for complex scenarios from the dataset. The model building module is used to build the aforementioned unsafe driving behavior detection model; The model training module is used to train the bad driving behavior detection model and evaluate the trained model. The behavior detection module is used to input the image to be detected into the trained bad driving behavior detection model and output the driving behavior category, confidence level and target location information. The warning module is used to generate prompts or warning information based on the driving behavior category and confidence level.

Images