Automatic driving YOLO target detection method and system coping with light influence

Abstract

The invention relates to an automatic driving YOLO target detection method and system for coping with light influence, and the method comprises the steps: employing a vehicle-mounted camera to shoot the actual condition images of streets and roads on a plurality of streets, obtaining the vehicle actual condition images of different streets and different roads in different time periods, and forming an image data set; marking the image data set by using a LabelImg tool, and marking a vehicle target in the image, including a bounding box and category information of a vehicle; constructing an automatic driving YOLO target detection network model coping with light influence; and training an automatic driving YOLO target detection network model coping with the light influence by using the obtained image data set, wherein the trained automatic driving YOLO target detection network model coping with the light influence is used for coping with vehicle target detection under the light influence. Accurate vehicle distance adjustment and safe driving decisions can be provided for the automatic driving vehicle, and driving safety and efficiency under various light conditions are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle target detection and processing, and in particular to an autonomous driving YOLO target detection method and system for coping with light influence. Background Art

[0002] In the field of vehicle target detection and processing technology, accurate identification and positioning of vehicle targets are crucial for autonomous driving, intelligent transportation systems and other fields. In autonomous driving systems, it is a key task to identify road conditions through accurate target detection to achieve safe driving of autonomous vehicles. Especially after detecting other vehicles, the system needs to make fast and accurate decisions based on real-time traffic conditions and the impact of surrounding vehicles, so as to achieve automatic obstacle avoidance of autonomous vehicles by automatically controlling vehicle speed and braking. With the development of technology, image-based vehicle target detection technology has made significant progress, but it still faces many challenges, especially in environments with complex light changes. Light conditions have a direct impact on the performance of vehicle target detection systems. Poor lighting environments such as night, cloudy days, strong light, etc. will cause image quality to deteriorate, which in turn affects the accuracy and robustness of target detection.

[0003] Traditional vehicle target detection methods mainly rely on manual feature extraction and classic machine learning algorithms. These methods can achieve certain results under certain conditions, but often perform poorly when dealing with complex lighting changes. For example, traditional methods based on features such as edge detection and corner detection are prone to failure under lighting changes, resulting in inaccurate feature extraction, which in turn affects detection performance.

[0004] With the development of deep learning technology, vehicle target detection methods based on convolutional neural networks (CNN) and You Only Look Once (YOLO) series network models have attracted widespread attention due to their powerful feature learning capabilities. These methods can automatically learn features from images and have a certain adaptability to lighting changes. However, existing methods still have the problem of performance degradation when facing extreme lighting conditions, such as strong light, shadow, and weak light. Although existing research methods have made some progress in dealing with the influence of light, there are still limitations: First, image preprocessing methods may not be able to fully restore the detailed information of the image, especially under extreme lighting conditions. Second, multi-exposure image fusion technology requires additional hardware support and is difficult to implement in application scenarios with high real-time requirements. Finally, although deep learning methods perform well in feature learning, training these models requires preprocessing of color brightness, contrast, etc. in advance, which takes a long time and increases processing costs, and the generalization ability of existing target detection models still needs to be improved.

[0005] Therefore, developing an autonomous driving YOLO target detection method and system that can effectively cope with the influence of light, realizing high-precision target detection under the influence of light during the driving process of the autonomous driving vehicle, and completing the effective identification of the vehicle condition during the autonomous driving, thereby accurately identifying the vehicle target in the direction of travel and automatically avoiding obstacles. This has important practical significance and application value. Summary of the invention

[0006] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a YOLO target detection method and system for autonomous driving that copes with the influence of light, and solve the above-mentioned obvious defects in the prior art. It can maintain high accuracy and robustness under various lighting conditions, and at the same time has the ability of real-time processing to meet the needs of autonomous driving in identifying vehicles and realizing automatic obstacle avoidance, and improve the performance of vehicle target detection in complex lighting environments.

[0007] In order to solve the above technical problems, the technical solution of the present invention is as follows: In a first aspect, the present invention provides an autonomous driving YOLO target detection method for coping with light influence, the detection method comprising the following steps: Step 1: Get the dataset Use vehicle-mounted cameras to capture the actual vehicle conditions on several streets, obtain real-time images of vehicles on different roads in different streets and at different time periods, and form an image dataset; the images in the image dataset include images under sufficient lighting conditions and images under weak lighting conditions; use the LabelImg tool to annotate the image dataset and identify the vehicle targets in the image, including the vehicle's bounding box and category information; Step 2: Build an autonomous driving YOLO target detection network model that copes with light effects The YOLO target detection network model for autonomous driving that copes with light influence includes three parts: the backbone, the neck, and the output. The main part is composed of a convolution layer Conv at the head, a middle part and a SPPF module at the tail in series from top to bottom, wherein the middle part includes four groups of serial structures composed of ghost convolution GhostConv and CA-C3Ghost modules; the convolution kernel size of the convolution layer Conv at the head is 2; The neck part consists of two columns of bottom-up structures. The first column is composed of the CA-C3Ghost module, the depthwise separable dilated convolution DSDC module, the upsample Upsample, the Concat splicing operation, the CA-C3Ghost module, the depthwise separable dilated convolution DSDC module, the upsample Upsample, and the Concat splicing operation, and the second column is composed of the CA-C3Ghost module, the Concat splicing operation, the depthwise separable dilated convolution DSDC module, the CA-C3Ghost module, the Concat splicing operation, the depthwise separable dilated convolution DSDC module, and the CA-C3Ghost module. The outputs of the SPPF module in the trunk are connected to the two CA-C3Ghost modules at the bottom of the two columns in the neck part; the outputs of the third CA-C3Ghost module and the second CA-C3Ghost module in the trunk are connected to the two Concat splicing operations in the first column of the neck part; the outputs of the two depthwise separable dilated convolutional DSDC modules in the first column of the neck part are connected to the two Concat splicing operations in the second column; at the same time, the output of the Concat splicing operation at the top of the first column is input to the CA-C3Ghost module at the top of the second column; In the output part, the outputs of the three CA-C3Ghost modules in the second column of the neck part are processed by three convolutional layers Conv with a convolution kernel of 1, which can identify moving vehicle targets of different sizes in an image and display the category information and bounding box of the detected target in the image; The CA-C3Ghost module includes a channel attention CA module, and the output of the channel attention CA module is divided into two branches, one branch is processed by a convolution Conv operation with a convolution kernel size of 1 and three series ghost convolution GhostConv operations to obtain a first output; the other branch directly obtains a second output by a convolution Conv operation with a convolution kernel size of 1; after the first output and the second output are spliced, a convolution Conv operation with a convolution kernel size of 1 is performed to obtain the output of the CA-C3Ghost module; This completes the construction of the YOLO target detection network model for autonomous driving that copes with the influence of light. Step 3: Use the image data set obtained in step 1 to train the YOLO target detection network model for autonomous driving that copes with the influence of light, and use the trained YOLO target detection network model for autonomous driving that copes with the influence of light to detect vehicle targets under the influence of light.

[0008] Furthermore, in step 3, the balanced focus distribution loss L is used during the training process. BFD As a loss constraint, the balanced focal distribution loss LBFD It is expressed as:

[0009]

[0010]

[0011]

[0012] Among them, IoU represents the intersection-over-union score between the predicted bounding box and the actual bounding box; γ represents a hyperparameter that adjusts the importance of the IoU term, and the default value is γ=2.0; α and β represent hyperparameters that control the importance of width and height, and the default values ​​are α=2.0 and β=0.25; δ represents a hyperparameter that dynamically adjusts IoU, and the default value is δ=0.5; max(0,R(b)+R(w)+R(h)) is a regularization term used to penalize the bounding box regression error, R(b) is the bounding box regularization term, R(w) is the width regularization term, and R(h) is the height regularization term; ρ() represents a function that calculates the Euclidean distance between two elements; b, Represent the predicted bounding box center and the true bounding box center, w, Denote the predicted bounding box width and the true bounding box width, h, denote the predicted bounding box height and the true bounding box height respectively; c, c w 、c h Represents the hyperparameters that manage the influence of the center point, width, and height. The default values ​​are c=1.0, c w =1.25, c h =1.25.

[0013] Furthermore, the depthwise separable dilated convolution DSDC module includes three parallel dilated convolutions, and the dilated convolution kernel sizes are set to 1, 2, and 3 respectively. The output results of the three are spliced ​​and divided into two branches, which are processed by a depthwise separable convolution and a pointwise convolution respectively. The result after the depthwise separable convolution processing is used as the input of the pointwise convolution on the one hand, and added with the output of the pointwise convolution on the other hand to obtain the output of the depthwise separable dilated volume DSDC module.

[0014] Furthermore, the image size in the image dataset is 640×640.

[0015] Furthermore, when starting training, the parameters of network initialization are set as follows: the epoch of the training network is set to 650, the optimizer adopts the stochastic gradient descent SGD optimizer, and the initial learning rate of SGD is set to 0.01, and the gradient weight decay parameter is set to 5e-5; the training is stopped when the error of the balanced focal distribution loss change is within ±1e-5.

[0016] In a second aspect, the present invention provides an autonomous driving YOLO target detection system that copes with light influence, wherein the system performs the steps of the detection method, including: Binocular camera module, used to obtain real-time images of road conditions, including images in sufficient light and images in weak light; The YOLO target detection network model for autonomous driving that copes with the influence of light is used to detect targets on moving vehicles in real time; The vehicle distance calculation module is used to calculate the vehicle distance between the vehicle and the vehicle identified in the traveling direction according to the target detection result; The vehicle driving feedback module is used to adjust the vehicle driving state according to the vehicle distance calculated by the vehicle distance calculation module.

[0017] Furthermore, the safe distance range is set to [s2, s3], and the dangerous distance threshold is set to s1, s1 Furthermore, the safety distance range is not less than 5 meters and not more than 15 meters, and the danger distance threshold is 3 meters.

[0018] Furthermore, the average pixel accuracy mAP of the YOLO target detection network model for autonomous driving that copes with the influence of light is greater than 59%, the frame rate per second FPS is greater than 80, and the number of parameters is controlled between 3-5 million.

[0019] Compared with the prior art, the present invention has the following beneficial effects: The Self-Driving Cars YOLO Counteract Light Influence Object Detection Network model (SD-YOLO-CLNet) in this invention has shown significant advantages in many aspects: In the present invention, SD-YOLO-CLNet can detect the position of the vehicle in the direction of travel of the autonomous vehicle in real time in response to the influence of light. After detecting the vehicle, SD-YOLO-CLNet can provide accurate distance adjustment and safe driving decisions for the autonomous vehicle, ensuring driving safety and efficiency under various light conditions. Based on the autonomous vehicle system, adding this model can realize efficient detection function, timely control the safe distance and avoid obstacles, effectively avoid collisions and optimize driving efficiency.

[0020] The present invention accomplishes the task of target detection by taking into account the influence of light. By combining the Internet of Vehicles technology, the system enables the autonomous driving vehicle to make optimal and safe autonomous driving decisions in complex and high-speed traffic environments.

[0021] In the model of the present invention, the CA-C3Ghost module and the phantom convolution GhostConv are used multiple times, which greatly reduces the computational load to enhance the speed of target detection, and can also maintain performance similar to the original convolution through effective feature map expansion. The CA-C3Ghost module uses the phantom convolution GhostConv and adds the channel attention Coordinate Attention (CA) module to enhance the model's attention to feature information by combining spatial position information. Two pooling layers are used in the CA module to aggregate information in different spatial directions respectively to generate direction-aware and position-aware attention maps. The introduction of the CA-C3Ghost module enables the model to have the ability to extract features under the influence of special conditions, especially under the influence of light, according to the image differences caused by sufficient light and weak light. The model emphasizes important features more and does not suppress unimportant features. It can effectively extract all pixel feature information in the image when the light is insufficient, improve the quality of feature extraction, and improve the robustness of target detection.

[0022] In terms of loss function, the present invention proposes a balanced focal distribution loss Balanced Focal Distribution Loss to solve the problem of category imbalance. By dynamically adjusting the loss, the model pays more attention to difficult-to-classify samples during training and gives higher weights to difficult-to-classify samples, which helps to improve the processing performance of the model in the data set.

[0023] In summary, the SD-YOLO-CLNet of the present invention deepens the scale of feature extraction in vehicle target detection under the influence of light, realizes deeper pyramid-shaped feature fusion and extraction while reducing the amount of calculation, and can better cope with target detection tasks under the influence of light. It not only improves the accuracy of target detection, but also achieves lightweight and enhances its applicability in autonomous driving vehicles. It takes into account accuracy, calculation effect and parameter minimization, and can run efficiently on resource-constrained devices. It shows excellent performance in the field of autonomous driving target detection and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of the structure of SD-YOLO-CLNet according to an embodiment of the present invention.

[0025] Figure 2 Schematic diagram of the structure of GhostConv.

[0026] Figure 3 Schematic diagram of the structure of the Depthwise-Separable Dilated Convolutions (DSDC) module.

[0027] Figure 4 The figure is a schematic diagram of the structure of a CA-C3Ghost module according to an embodiment of the present invention.

[0028] Figure 5 A schematic diagram of a driving process flow of an autonomous driving vehicle using SD-YOLO-CLNet in one embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to more clearly describe the technical problems, technical solutions and advantages of the present invention, the following will be described in detail with reference to the drawings and embodiments. It should be noted that these embodiments are only used to illustrate the principles and application scope of the present invention and should not be regarded as limiting the present invention.

[0030] In the description of this specification, the specific features, structures or characteristics described may be combined in an appropriate manner in any one or more embodiments.

[0031] Figure 3 The structural diagram of the depthwise separable dilated convolution DSDC module includes three parallel dilated convolutions. In order to complete the multi-level dilation, the dilated convolution kernel sizes are set to 1, 2, and 3 respectively. The output results of the three are concatenated and divided into two branches, which are processed by a depthwise separable convolution and a pointwise convolution respectively. The result of the depthwise separable convolution is used as the input of the pointwise convolution on the one hand, and added with the output of the pointwise convolution on the other hand to obtain the output of the depthwise separable dilated convolution DSDC module. Depthwise-Separable Dilated Convolutions (DSDC) combines depthwise separable convolutions and dilated convolutions. Figure 3 The dotted box in the middle adds holes in the convolution kernel to increase the receptive field. DSDC combines the two to solve the problem of difficulty in extracting image information caused by light to the greatest extent. It expands the receptive field and reduces the number of convolution calculations to make the model lighter, while ensuring the decrease of calculation parameters and maintaining a balance of accuracy.

[0032] Embodiment 1: The YOLO object detection method for autonomous driving against light influence in this embodiment uses a YOLO object detection network model for autonomous driving against light influence (Self-Driving Cars YOLO Counteract Light InfluenceObject Detection Network), and includes the following steps: Step 1: Get the dataset Use a car camera to shoot the actual traffic conditions on several streets, select these streets for shooting at different time periods, obtain the actual conditions of vehicles on different roads in different streets, and collect them as image datasets under well-lit conditions and weak-lit conditions (weak light includes shadows and low light conditions) during the day and at night. In this embodiment, the image dataset totals 8,500 images, including 4,000 well-lit images and 4,500 weak-lit images. The acquired image dataset is annotated using the LabelImg tool to identify the vehicle targets in the image, including the vehicle's bounding box and category information. The image dataset is randomly divided into a training set and a test set in a ratio of 7:3. When inputting the network model, in order to ensure the consistency of the image size, all images are set to 640×640.

[0033] Step 2: Build an autonomous driving YOLO target detection network model that copes with light effects The structure diagram of the autonomous driving YOLO target detection network model that copes with the influence of light is as follows: Figure 1 As shown, it consists of three main parts: the trunk, the neck and the output.

[0034] The main part is composed of the convolution layer Conv at the head, the middle part, and the Spatial Pyramid Pooling – Fast (SPPF) module at the tail from top to bottom. The middle part includes four groups of serial structures composed of ghost convolution GhostConv and CA-C3Ghost modules. The structure of ghost convolution GhostConv can be found in Figure 2 The structure of the CA-C3Ghost module can be found in Figure 4 .

[0035] The input image is first processed by the backbone part, and the convolution layer Conv with a convolution kernel size of 2 at the head of the backbone part performs initial feature extraction and expands its number of channels, extracting the input image of size 640×640×3 (height, width, number of channels) into an initial feature map of size 320×320×32 Next, the first set of ghost convolution GhostConv and CA-C3Ghost modules are used to extract deep features from the initial feature map, and the output feature map with a size of 160×160×64 is obtained. ; Then, the second set of ghost convolution GhostConv and CA-C3Ghost modules are used to map the feature map Perform feature extraction again to obtain a feature map of size 80×80×128 ; The next step is to use the third group of ghost convolution GhostConv and CA-C3Ghost modules to perform feature map Perform the same feature extraction operation to obtain a feature map of size 40×40×256 ; The last set of ghost convolution GhostConv and CA-C3Ghost modules for feature maps Perform feature extraction to obtain a feature map of size 20×20×512 Finally, the SPPF module performs Perform spatial pyramid pooling and channel concatenation of the pooling results to obtain a feature map of size 20×20×512 containing multi-scale information Used as partial input to the next stage.

[0036] The neck part consists of two columns of bottom-up structures, the first column of which is a characteristic graph. As input, a CA-C3Ghost module and a depth-separable dilated convolution DSDC module are used for feature extraction to obtain a feature map of size 20×20×256. , after an upsampling process, and then combined with the feature map The size of the splicing is Feature map ; Then the feature map As input, a CA-C3Ghost module and a depth-separable dilated convolution DSDC module are used for feature extraction to obtain a feature map of size 40×40×128. , after an upsampling process, and then combined with the feature map The feature map of size 80×80×256 is obtained by splicing The second column is the feature map As input, a CA-C3Ghost module is used for feature extraction to obtain a feature map of size 20×20×512. , feature map With feature map The feature map is concatenated to get a size of 20×20×256 ; Next, the feature map A feature map of size 40×40×256 is obtained by extracting features through a depthwise separable dilated convolution DSDC module and a CA-C3Ghost module. , the feature map With feature map The feature map of size 80×80×128 is obtained by splicing ; Final feature map Feature extraction is performed through a depth-wise separable dilated convolution DSDC module and a CA-C3Ghost module, where the CA-C3Ghost module uses feature maps The result of the DSDC module at the top of the second column is taken as input to obtain a feature map of size 80×80×128. .

[0037] The output part includes three convolution layers Conv with a convolution kernel size of 1. From top to bottom, the three convolution layers Conv are respectively based on feature maps , feature map , feature map As input, by processing feature maps of three different scales, it is possible to identify moving vehicle targets of different sizes in an image and display the category information and bounding box of the detected target in the image.

[0038] Balanced Focal Distribution Loss aims to improve the performance of the target detection model when dealing with class imbalance problems, especially in bounding box regression. It combines the advantages of Focal Loss and Efficient Intersection over Union (EIoU) Loss to improve the model's attention to difficult samples and optimize the regression accuracy of bounding boxes. BFD It can be expressed by the following formula:

[0039]

[0040]

[0041]

[0042] Among them, IoU represents the intersection-over-union score between the predicted bounding box and the actual bounding box; γ represents the hyperparameter for adjusting the importance of the IoU term, and the default value is γ=2.0; α and β represent the hyperparameters for controlling the importance of width and height, and the default values ​​are α=2.0 and β=0.25; δ represents the hyperparameter for dynamically adjusting IoU, and the default value is δ=0.5; max(0,R(b)+R(w)+R(h)) is the regularization term used to penalize the bounding box regression error, R(b) is the bounding box regularization term, R(w) is the width regularization term, and R(h) is the height regularization term; ρ() represents the function for calculating the Euclidean distance between two elements; b, Represent the predicted bounding box center and the true bounding box center, w, Denote the predicted bounding box width and the true bounding box width, h, denote the predicted bounding box height and the true bounding box height respectively; c, c w 、c h Represents the hyperparameters that manage the influence of the center point, width, and height. The default values ​​are c=1.0, c w =1.25, c h =1.25.

[0043] Step 3: Use the image data set obtained in step 1 to train the YOLO target detection network model for autonomous driving that copes with the influence of light, and use the trained YOLO target detection network model for autonomous driving that copes with the influence of light to detect vehicle targets under the influence of light.

[0044] Embodiment 2: The YOLO target detection method for autonomous driving under the influence of light in this embodiment implements vehicle target detection under the influence of light to achieve automatic obstacle avoidance for the autonomous driving vehicle through the following steps: 1. Data collection and preparation stage 1.1 Data Collection Equipment: Use a vehicle-mounted camera to capture images of actual road conditions, and capture images of well-lit and dimly-lit conditions on several different streets and at different time periods. Dimly-lit conditions include shadows and low light conditions, but do not include severe weather conditions such as heavy fog, snow, wind, and sand.

[0045] Operation: Take images of road conditions and vehicles for each street, ensuring that images under different light conditions and multiple groups of vehicle target images are included. For example, in this embodiment, the target is the image of the vehicle condition under different light conditions on the road, lock the target and shoot the target road with the vehicle-mounted camera at different angles.

[0046] 1.2 Data Processing Dataset acquisition: The collected image data is preliminarily processed, including removing obvious noise and outliers in the image, and cropping and resizing the image to ensure the consistency and standardization of the image dataset. The collected images are divided into training set and test set in a 7:3 ratio.

[0047] Data augmentation: Perform data augmentation operations such as rotation, scaling, and flipping on the training set (excluding adjustments such as brightness) to improve the robustness and generalization ability of the model.

[0048] 2. Model training phase 2.1 Model Construction Construct SD-YOLO-CLNet, which includes three parts: trunk, neck and output. The specific structure is the same as Example 1.

[0049] 2.2 SD-YOLO-CLNet Training When starting training, first complete the parameter setting for network initialization: the epoch of the training network is set to 650, the optimizer uses the stochastic gradient descent SGD optimizer, and the initial learning rate of SGD is set to 0.01, and the gradient weight decay parameter is set to 5e-5.

[0050] During the training process, the training set images are input into SD-YOLO-CLNet, and the images are read according to the training set storage path; When the overall network objective loss function L BFD When there is no longer a significant decrease (error ±1e-5), the network model training is considered to be stable and the training process is completed.

[0051] Loss function: Balanced Focal Distribution Loss is used to optimize model parameters to improve target detection performance.

[0052] When conducting network testing, the prepared test set images are input, and the network weights trained in the training set are imported to obtain the test structure, thus completing the model target detection purpose.

[0053] 3. Processing and analysis stage 3.1 Model Deployment Deployment: In this embodiment, two cameras with a resolution of 1920×1080, a frame rate of 30FPS, and a field of view FOV (H × V× D) of 69.4° × 42.5° × 77° (±3°) are used as the left camera and the right camera respectively to form a binocular camera module. The two cameras need to be installed in parallel and their optical axes must be aligned to ensure the accuracy of stereoscopic vision. Then the left camera and the right camera are calibrated to obtain the internal and external parameters, distortion coefficients, and homography matrices of the two cameras. According to the calibration results, the two cameras are corrected so that the images obtained by the two cameras after correction are located in the same plane and parallel to each other. Then, the disparity map conversion is performed, and the images obtained by the two cameras are converted into depth maps through the reprojection matrix to obtain the coordinates in the three-dimensional space, completing the preparation of the binocular camera module.

[0054] The trained SD-YOLO-CLNet is deployed to the target detection system of the autonomous vehicle for real-time vehicle detection during driving. When the autonomous vehicle is driving on urban roads at a speed of less than 70 km / h, its target detection system detects the real-time vehicle image transmitted by the on-board camera. After identifying the vehicle target, the recognition result is transmitted to the distance calculation module, which calculates the distance between the vehicle and the vehicle identified in the direction of travel. Then the vehicle driving feedback module adjusts the vehicle driving state according to different vehicle distances.

[0055] 3.2 Results Analysis Model quality evaluation: The target detection performance is evaluated by mAP (average pixel accuracy) and FPS (frames per second) of target detection, as well as the number of parameters.

[0056] On a dataset of road conditions under different light conditions captured by the same on-board camera (8,500 images), the SD-YOLO-CLNet of the present invention was used to compare with the existing methods. The comparison results show that compared with the existing excellent target detection methods such as FasterRCNN, CenterNet, YOLOv5s, and YOLOv7, the target detection method of the present invention has lower computational complexity and better performance. The comparison results of different models after training for vehicle target detection processing are shown in Table 1.

[0057]

[0058] The results shown in Table 1 show that in the target detection processing of the road vehicle image dataset taken by the vehicle camera under different light conditions, the mAP of SD-YOLO-CLNet is higher than that of the existing YOLOv7, YOLOv5s, CenterNet and FasterRCNN, and it still has a higher frame rate per second FPS while ensuring that the number of parameters is lower. Compared with the above-mentioned similar target detection methods, the difference is significant, which proves that the present invention has superior application performance in vehicle target detection under the influence of light.

[0059] Embodiment 3: This embodiment deals with the influence of light on the autonomous driving YOLO target detection system (see Figure 5 ), the system comprising: Binocular camera module, used to obtain real-time images of road conditions, including images in sufficient light and images in weak light; The YOLO object detection network model for autonomous driving that copes with light effects (SD-YOLO-CLNet) is used to detect objects in real time on moving vehicles; The vehicle distance calculation module is used to calculate the vehicle distance between the vehicle and the vehicle identified in the traveling direction according to the target detection result; The vehicle driving feedback module is used to adjust the vehicle driving state according to the vehicle distance calculated by the vehicle distance calculation module.

[0060] The safe distance range is set to no less than 5 meters and less than 15 meters, and the dangerous distance threshold is set to 3 meters; if the vehicle distance is greater than or equal to 3 meters and less than 5 meters, it is considered a dangerous distance, that is, there is a possibility of collision, and the vehicle braking unit will slow down to maintain a safe distance; if the vehicle distance is greater than or equal to 5 meters and less than 15 meters, it is considered a safe distance, and the vehicle can maintain normal driving without speed adjustment; if the vehicle distance is not less than 15 meters, the vehicle can accelerate appropriately to control the vehicle distance within the safe distance to optimize driving efficiency.

[0061] Embodiment 4: Apply SD-YOLO-CLNet to traffic flow monitoring, especially the ability to detect and count vehicle targets under different lighting conditions.

[0062] Deployment: Deploy the trained SD-YOLO-CLNet into the urban traffic monitoring system for real-time vehicle flow detection and counting.

[0063] Quality evaluation: Calculate the mAP (average pixel accuracy) and FPS (frames per second) of target detection, as well as the parameters Parameters, to evaluate the target detection performance.

[0064] Performance improvement: Based on feedback from actual applications, optimize the network model and improve the processing efficiency of the system.

[0065] User feedback: Collect user feedback on the object detection performance to further adjust and improve system settings.

[0066] This example verifies the ability of SD-YOLO-CLNet to detect and count vehicle targets under different lighting conditions, providing a new solution for urban traffic flow monitoring.

[0067] The present invention aims to solve the existing defects of the current vehicle target detection technology, namely, insufficient target detection accuracy and slow processing speed under the influence of light. The SD-YOLO-CLNet in the present invention realizes vehicle target detection under the influence of light, which can improve recognition efficiency and save computing resources, thereby realizing timely safety distance control and improving driving efficiency through the driving feedback of the autonomous driving vehicle.

[0068] Therefore, the technical solution of the present invention improves the target detection efficiency and avoids the influence caused by light, which is of great significance to the target detection efficiency and safety of autonomous driving vehicles.

[0069] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A YOLO target detection method for autonomous driving that copes with light influence, characterized in that: The detection method comprises the following steps: Step 1: Get the data set; Use vehicle-mounted cameras to capture the actual vehicle conditions on several streets, obtain real-time vehicle images on different roads in different streets and at different time periods, and form an image dataset; use the LabelImg tool to annotate the image dataset and identify the vehicle targets in the image, including the vehicle's bounding box and category information; Step 2: Build an autonomous driving YOLO target detection network model that copes with light effects; The YOLO target detection network model for autonomous driving that copes with light influence includes three parts: the backbone, the neck, and the output. The main part is composed of a convolution layer Conv at the head, a middle part and a SPPF module at the tail in series from top to bottom, wherein the middle part includes four groups of serial structures composed of ghost convolution GhostConv and CA-C3Ghost modules; the convolution kernel size of the convolution layer Conv at the head is 2; The neck part consists of two columns of bottom-up structures. The first column is composed of the CA-C3Ghost module, the depthwise separable dilated convolution DSDC module, the upsample Upsample, the Concat splicing operation, the CA-C3Ghost module, the depthwise separable dilated convolution DSDC module, the upsample Upsample, and the Concat splicing operation, and the second column is composed of the CA-C3Ghost module, the Concat splicing operation, the depthwise separable dilated convolution DSDC module, the CA-C3Ghost module, the Concat splicing operation, the depthwise separable dilated convolution DSDC module, and the CA-C3Ghost module. The outputs of the SPPF module in the trunk are connected to the two CA-C3Ghost modules at the bottom of the two columns in the neck part; the outputs of the third CA-C3Ghost module and the second CA-C3Ghost module in the trunk are connected to the two Concat splicing operations in the first column of the neck part; the outputs of the two depthwise separable dilated convolutional DSDC modules in the first column of the neck part are connected to the two Concat splicing operations in the second column; at the same time, the output of the Concat splicing operation at the top of the first column is input to the CA-C3Ghost module at the top of the second column; In the output part, the outputs of the three CA-C3Ghost modules in the second column of the neck part are processed by three convolutional layers Conv with a convolution kernel of 1, which can identify moving vehicle targets of different sizes in an image and display the category information and bounding box of the detected target in the image; The CA-C3Ghost module includes a channel attention CA module, and the output of the channel attention CA module is divided into two branches, one branch is processed by a convolution Conv operation with a convolution kernel size of 1 and three series ghost convolution GhostConv operations to obtain a first output; the other branch directly obtains a second output by a convolution Conv operation with a convolution kernel size of 1; after the first output and the second output are spliced, a convolution Conv operation with a convolution kernel size of 1 is performed to obtain the output of the CA-C3Ghost module; This completes the construction of the YOLO target detection network model for autonomous driving that copes with the influence of light. Step 3: Use the image data set obtained in step 1 to train the YOLO target detection network model for autonomous driving that copes with the influence of light, and use the trained YOLO target detection network model for autonomous driving that copes with the influence of light to detect vehicle targets under the influence of light.

2. The detection method according to claim 1, characterized in that: In step 3, the balanced focus distribution loss is used during training. L BFD As a loss constraint, the balanced focal distribution loss L BFD It is expressed as: ， ， ， ， in, IoU represents the intersection-over-union score between the predicted bounding box and the actual bounding box; γ represents the adjustment IoU The importance of the item is a hyperparameter; α and β represent the hyperparameters controlling the width and height importance; δ represents the dynamic adjustment IoU Hyperparameters of max(0, R (b)+ R (w)+ R (h)) is a regularization term used to penalize the bounding box regression error. R (b) is the bounding box regularization term, R (w) is the width regularization term, R (h) is the height regularization term; ρ() represents the function for calculating the Euclidean distance between two elements; b、 Represent the predicted bounding box center and the true bounding box center respectively, w、 Represent the predicted bounding box width and the true bounding box width, respectively. h、 Represent the predicted bounding box height and the true bounding box height respectively; c , c w 、c h Represents the hyperparameters that manage the influence of center point, width, and height.

3. The detection method according to claim 2, characterized in that: γ=2.0, α=2.0, β=0.25, δ=0.5, c =1.0 、 c w =1.25 、c h =1.

25.

4. The detection method according to claim 1, characterized in that: The depthwise separable dilated convolution DSDC module includes three parallel dilated convolutions, and the dilated convolution kernel sizes are set to 1, 2, and 3 respectively. The output results of the three are spliced ​​and divided into two branches, which are processed by a depthwise separable convolution and a pointwise convolution respectively. The result after the depthwise separable convolution processing is used as the input of the pointwise convolution on the one hand, and is added to the output of the pointwise convolution on the other hand to obtain the output of the depthwise separable dilated volume DSDC module.

5. The detection method according to claim 1, characterized in that: The image size in the image dataset is 640×640.

6. The detection method according to claim 1, characterized in that: When starting training, the parameters of network initialization are set as follows: the epoch of the training network is set to 650, the optimizer uses the stochastic gradient descent SGD optimizer, and the initial learning rate of SGD is set to 0.01, and the gradient weight decay parameter is set to 5e-5; the training is stopped when the error of the balanced focal distribution loss change is within ±1e-5.

7. An autonomous driving YOLO target detection system that copes with light influence, characterized in that: The system performs the steps of any one of the detection methods of claims 1 to 6, including: Binocular camera module, used to obtain real-time images of road conditions, including images in sufficient light and images in weak light; The YOLO target detection network model for autonomous driving that copes with the influence of light is used to detect targets on moving vehicles in real time; The vehicle distance calculation module is used to calculate the vehicle distance between the vehicle and the vehicle identified in the traveling direction according to the target detection result; The vehicle driving feedback module is used to adjust the vehicle driving state according to the vehicle distance calculated by the vehicle distance calculation module.

8. The system according to claim 7, characterized in that The average pixel accuracy mAP of the autonomous driving YOLO target detection network model that copes with the influence of light is greater than 59%, the frame rate per second FPS is greater than 80, and the number of parameters is controlled between 3-5 million.

9. The system according to claim 7, characterized in that Set the safe distance range as [s2, s3), set the dangerous distance threshold as s1, where s1 < s2 < s3, and the unit is meters; if the vehicle distance is less than s2, it is regarded as a dangerous distance, and there is a potential possibility of collision. The vehicle braking unit will decelerate to maintain the safe distance; if the vehicle distance is within the safe distance range [s2, s3), it is regarded as a safe distance, and the normal driving of the vehicle is maintained without speed adjustment; if the vehicle distance is not less than s3, the vehicle can accelerate and keep the vehicle distance within the safe distance.

10. The system according to claim 9, characterized in that The safe distance range is not less than 5 meters and less than 15 meters, and the dangerous distance threshold is 3 meters.

Images