Unmanned vehicle accident judgment and responsibility judgment system and method based on roadside perception

Abstract

The invention discloses an unmanned vehicle accident judgment and responsibility judgment system and method based on roadside perception, and the system comprises a roadside perception module which is used for obtaining target scene image data, physical environment information and point cloud data in real time, the accident identification and evaluation module is used for judging whether an accident exists in the target scene and classifying the accident type, and through multi-dimensional analysis of the accident, the problems that the accident judgment response is not timely and the data dimension is not comprehensive in the unmanned vehicle accident handling process are solved. Besides, the invention provides an accident escape system and an accident judgment system based on the accident judgment system, an escape strategy is formulated by combining a steering wheel control technology, and accident liability is judged, so that the problems that an unmanned driving strategy cannot reasonably avoid obstacles and accident liability processing is complex are solved, the claim settlement time is shortened, and the safety is improved. And a whole-process solution from accident perception, intelligent decision-making to responsibility judgment is provided for the unmanned vehicle.

Description

Technical Field

[0001] The present invention relates to the technical field of driverless taxis, and in particular to a system and method for determining accident and liability of driverless vehicles based on roadside perception. Background Art

[0002] The autonomous driving technology of driverless vehicles is applied to a form of shared travel services, which represents an important development direction of future urban transportation. With the rapid development of 5G, big data, cloud computing and autonomous driving technology, driverless vehicles are gradually becoming a reality. Driverless vehicles can safely drive on the road without human drivers and provide pick-up and drop-off services for passengers. Among them, the L4 point-to-point autonomous driving system has matured, allowing the vehicle to completely take over the driving task under certain conditions, greatly reducing the demand for human drivers and improving the safety and efficiency of travel.

[0003] With the rapid development of driverless technology, road traffic safety issues have received increasing attention. In complex traffic scenarios, real-time detection and efficient processing of vehicle accidents have become one of the core challenges to ensure the safety of autonomous driving. Traditional accident detection systems mainly rely on on-board sensors (such as cameras, millimeter-wave radars, lidars, etc.) to obtain local environmental information and make accident judgments through on-board computing units. However, such methods have perception blind spots in scenarios such as extreme weather, sensor occlusion, or on-board equipment failure, resulting in a significant decrease in the accuracy and robustness of accident identification. In addition, it is difficult to fully reconstruct the three-dimensional information of the accident scene from a single on-board perspective, which limits the reliability of accident type classification and subsequent processing.

[0004] In recent years, roadside perception technology has been able to provide global environmental perception data for unmanned vehicles by deploying multimodal sensors (such as binocular cameras, lidar, millimeter-wave radar, etc.), becoming an important direction for improving accident identification capabilities. However, existing roadside perception-based systems still have the following problems:

[0005] Low efficiency in accident feature recognition: Traditional accident identification is difficult to meet real-time requirements when dealing with complex scenarios and large amounts of calculations. In addition, the accuracy of accident feature extraction is insufficient, which can easily lead to missed detection or false detection.

[0006] Single post-accident handling strategy: Existing escape systems mostly use static path planning methods, which do not consider the dynamic impact of real-time traffic information and vehicle damage status, and lack a collaborative rescue mechanism for vehicle body accidents, making it difficult to achieve efficient accident response.

[0007] Liability determination relies on manual work: The division of accident responsibility usually requires manual intervention. The lack of deep integration of the automated rule engine and the traffic regulations knowledge base results in inefficient and highly subjective liability determination.

[0008] To address the above issues, there is an urgent need for a comprehensive system that integrates high-precision roadside perception, efficient accident identification, dynamic path planning and automated responsibility determination to improve the safety and handling efficiency of driverless vehicles in accident scenarios. Summary of the invention

[0009] The purpose of the present invention is to provide a system and method for unmanned vehicle accident identification and responsibility judgment based on roadside perception. The present invention solves the problems existing in current unmanned vehicles in accident handling, such as untimely response to accident identification, incomplete data dimensions, inability to reasonably avoid obstacles, and complex handling of accident responsibility.

[0010] To achieve this purpose, the present invention designs an unmanned vehicle accident identification system based on roadside perception, which is characterized by comprising:

[0011] The roadside perception module is used to obtain target scene image data, target scene physical environment information and target scene point cloud data;

[0012] The accident recognition and assessment module is used to use the target recognition unit to extract the target scene accident features in the target scene image data, obtain the three-dimensional reconstructed driving scene according to the target scene image data and the target scene point cloud data, and compose the comprehensive accident information according to the target scene accident features, the three-dimensional reconstructed driving scene and the target scene physical environment information, and judge whether there is an accident in the target scene according to the comprehensive accident information; if there is an accident in the target scene, the accident type is classified according to the distance between the accident location and the vehicle body.

[0013] The beneficial effects of the present invention are:

[0014] 1) By integrating the data from binocular cameras, millimeter-wave radars and lidars on roadside equipment and on-board equipment, the image data, physical environment information and point cloud data of the target scene are obtained. Through the target recognition algorithm and 3D reconstruction algorithm, the accident characteristics and driving scenes are obtained, thereby obtaining comprehensive accident information, avoiding the problem of low accident recognition accuracy caused by single data recognition of accidents; using the comprehensive accident information to determine whether an accident has occurred in the target scene, and quickly identify the type of accident, immediately trigger the alarm mechanism to ensure the timeliness of emergency rescue and accident handling, ensure the accuracy and response speed of unmanned vehicle accident detection, reduce delays in accident handling, and improve the overall traffic safety level.

[0015] 2) The escape module formulates an escape strategy based on the accident type and accident information, covering two aspects: obstacle avoidance on the accident section and safe escape after the accident. It not only improves the emergency response capability of unmanned vehicles in complex environments, but also further enhances the safety and operation efficiency of the overall transportation system.

[0016] 3) Steering wheel angle control technology significantly improves the steering accuracy and response speed of unmanned vehicles in emergency situations, ensuring that the vehicle can quickly and accurately avoid obstacles, effectively reducing the risk of accidents. This technology optimizes the vehicle's driving path by adjusting the steering wheel angle in real time, improves the reliability and safety of the autonomous driving system, and enhances the smoothness of traffic flow and the efficiency of the overall transportation system.

[0017] 4) The claims process is fully automated. The present invention can automatically generate a liability determination letter and transmit the liability determination result to a third-party claims assessment point and an insurance company to conduct vehicle damage assessment, repair, and compensation. This can greatly reduce human resource costs, significantly shorten the claims time, effectively improve user satisfaction and trust, and enhance the user experience of unmanned vehicle services. At the same time, it also further reduces system maintenance and operating costs, bringing significant economic benefits to insurance companies and operators.

[0018] 5) Establishing an automated responsibility determination system can accurately divide accident responsibilities, effectively reduce subjective bias caused by human factors, and ensure the fairness and transparency of responsibility determination. At the same time, the comprehensive collection and analysis of multi-source data provides reliable evidence support, further improves the accuracy of responsibility determination, reduces disputes and controversies caused by unclear responsibilities, and ensures the fairness and transparency of responsibility division.

[0019] 6) Through real-time collection and automated analysis of multi-source data, the credibility and transparency of accident data are effectively improved, and the possibility of false reporting and exaggeration of losses is reduced. This not only protects the interests of insurance companies and reduces unnecessary compensation expenses, but also improves the overall reputation and credibility of the insurance industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of the unmanned vehicle accident identification system based on roadside perception of the present invention;

[0021] Figure 2 This is a flow chart of the unmanned vehicle responsibility judgment system based on roadside perception of the present invention;

[0022] Figure 3 This is a flow chart of the process of escaping from an accident and determining responsibility of an unmanned vehicle based on roadside perception in an embodiment of the present invention;

[0023] Figure 4 It is a YOLOv8n network framework diagram of the target recognition algorithm in an embodiment of the present invention;

[0024] Figure 5 This is a diagram showing the effect of point cloud reconstruction of a driving scene image in an embodiment of the present invention.

[0025] Figure 6 This is a rendering of point cloud data generated by a laser radar according to an embodiment of the present invention.

[0026] Figure 7 This is a diagram of the EPP-MVSNet network framework for point cloud reconstruction in an embodiment of the present invention;

[0027] Figure 8 This is a flowchart of accident identification in an embodiment of the present invention;

[0028] Figure 9 A light steering angle control network framework diagram of the steering wheel steering angle control in an embodiment of the present invention;

[0029] Figure 10 It is a flow chart of the escape strategy in the embodiment of the present invention;

[0030] Figure 11 4 is a flow chart of steering wheel angle control in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0032] Embodiment 1:

[0033] like Figures 1 - 11 An unmanned vehicle accident identification system based on roadside perception is shown, comprising:

[0034] The roadside perception module is used to obtain real-time road traffic information and image data of the target scene, physical environment information of the target scene and point cloud data of the target scene;

[0035] The main function of the roadside perception module is to obtain real-time image data of the target scene, point cloud data of the target scene, and physical environment information of the target scene. The setting of this module improves the system's environmental perception ability, enabling the system to fully and accurately understand the surrounding environment, providing rich data support for subsequent accident identification and assessment modules, ensuring that the system has high reliability and real-time performance in complex and changing road environments, thereby providing strong support for accident prevention and handling.

[0036] The accident identification and assessment module is used to extract the target scene accident features from the target scene image data using the target identification unit, obtain the three-dimensional reconstructed driving scene based on the target scene image data and the target scene point cloud data, and form the comprehensive accident information based on the target scene accident features, the three-dimensional reconstructed driving scene and the target scene physical environment information. According to the comprehensive accident information, it is judged whether there is an accident in the target scene; if there is an accident in the target scene, the accident type is classified according to the distance between the accident location and the vehicle body. If there is no accident in the target scene, it returns to the roadside perception module, and no escape strategy is formulated or responsibility is determined.

[0037] Comprehensive accident information refers to multi-dimensional data composed of target scene accident characteristics, target scene physical environment information and three-dimensional reconstructed driving scene.

[0038] The accident location and the distance of the vehicle are obtained by integrating the physical environment information of the target scene and the three-dimensional reconstructed driving scene.

[0039] The accident identification and assessment module extracts features from the target scene image data through the target recognition unit, extracts the accident features of the target scene, and combines the target scene physical environment information and the target scene point cloud data to form comprehensive accident information, and uses the comprehensive accident information to accurately identify and assess the accident. The setting of this module enables the system to quickly and accurately determine the occurrence of accidents and classify the types of accidents, which helps to improve the efficiency of accident handling and reduce the losses caused by accidents.

[0040] In the above technical solution, the roadside perception module includes roadside and vehicle-mounted equipment, and the roadside and vehicle-mounted equipment include: binocular camera, millimeter wave radar and laser radar, and multiple sensors are respectively provided on the roadside and vehicle-mounted equipment.

[0041] Binocular cameras are deployed on key roads and on-board equipment of vehicles to collect high-definition videos and images of both sides of the road and the surrounding environment, provide road image information and depth information, and obtain information such as vehicle appearance features, license plate numbers, pedestrian activities, and traffic signs. Among them, the position of the target scene in the image can be determined by the distance of the target scene relative to the binocular camera, and the depth information refers to the distance information of each object relative to the binocular camera, which can obtain the position of the vehicle.

[0042] Millimeter-wave radar is deployed on key roads and on-board equipment of unmanned vehicles to measure the moving speed of surrounding vehicles and objects, detect sudden deceleration or acceleration, accurately measure the distance between vehicles and obstacles, monitor the safe distance between workshops and road boundaries, and obtain the distance, speed and acceleration of the vehicle.

[0043] LiDAR generates high-precision 3D point cloud data, constructs a detailed 3D model of the road and the surrounding environment, and provides accurate position, shape, and size information of objects, assisting in the 3D reconstruction and liability determination of accident scenes.

[0044] At the same time, binocular cameras, millimeter-wave radars, and LiDAR are used for data collection, which can provide multi-dimensional and high-precision environmental perception for intelligent driving. Binocular cameras provide image data and depth information through stereo vision, millimeter-wave radars measure distance and speed under all-weather conditions, and LiDAR generates high-precision 3D point cloud data. This multi-sensor fusion solution can make up for the limitations of a single sensor, enhancing environmental adaptability and system robustness.

[0045] In the above technical solution, for the roadside perception module, the image data of the target scene is obtained from the image data of the binocular camera sensor through the timestamp alignment module and the spatial calibration module; the physical environment information of the target scene is obtained from the millimeter-wave radar data of the millimeter-wave radar sensor through the timestamp alignment module and the spatial calibration module; the point cloud data of the target scene is obtained from the 3D reconstruction of the image data of the binocular camera sensor and the LiDAR data of the LiDAR sensor through the timestamp alignment module and the spatial calibration module, where:

[0046] Timestamp alignment module, the timestamp alignment module uses the Global Navigation Satellite System clock time as the reference time and performs timestamp alignment through the PTP clock synchronization protocol;

[0047] Among them, PTP is the abbreviation of Precision Time Protocol.

[0048] Spatial calibration module, the spatial calibration module performs spatial calibration through the ICP algorithm.

[0049] The image data of the target scene, the physical environment information of the target scene, and the point cloud data of the target scene are spatially calibrated through the ICP algorithm.

[0050] Using the ICP algorithm to iteratively find the corresponding points of each point in the image data of the target scene on the physical environment information of the target scene and the point cloud data of the target scene, calculate the deviation between the two, and adjust the coordinate transformation parameters accordingly. Finally, all sensor data are represented in the same coordinate system, which helps to more accurately understand and predict the environmental conditions around the vehicle, improving the safety and efficiency of autonomous driving.

[0051] The specific process is as follows:

[0052] Stereo matching of the target scene image data obtained by a binocular camera can achieve point cloud reconstruction of the target scene. Denote the target scene point cloud data generated by the lidar as the target point cloud Q, and the point cloud data obtained after three-dimensional reconstruction of the target scene image data generated by the binocular camera as the source point cloud P. The ultimate goal is to minimize the sum of the squared distances between P and Q:

[0053]

[0054] where T0 is the rigid body transformation matrix, R is the rotation matrix, and t is the translation vector, that is p i is a point in the source point cloud P, q i is a point in the target point cloud Q, i is the number of the point in the point cloud, ranging from 1 to n, where n is the number of points in the target point cloud, and ||·|| 2 is the square of the Euclidean distance, used to measure the distance between the point T0p i in the transformed source point cloud and the corresponding point q i in the target point cloud.

[0055] For each point p i ∈P, find the nearest neighbor point q i in the target point cloud Q:

[0056]

[0057] where q is a point in the target point cloud Q, ||·|| represents the two-norm, and can also be expressed as ||·||2, which is the Euclidean distance between two points.

[0058] Use the least squares method to solve for the optimal rotation matrix R and translation vector t. First, calculate the centroids of P and Q:

[0059]

[0060] μ P is the centroid of P, and μ Q is the centroid of Q;

[0061] Calculate the point clouds after taking the centroids:

[0062] p′ i = p i - μ P , q′ i = q i - μ Q ;

[0063] p′ i and q′ i represent the corresponding points in the source point cloud and the target point cloud after taking the centroids;

[0064] Calculate the covariance matrix H:

[0065]

[0066] Where n represents the number in the point cloud, and here T represents the transpose operation;

[0067] Use singular value decomposition (SVD) to solve H = UΣV T , then the rotation matrix R is:

[0068] R = VU T ;

[0069] U represents the left singular matrix, which consists of orthogonal unit vectors that are the orthonormal basis of the column space of the original matrix H; V is the right singular matrix, also consisting of orthogonal unit vectors that are the orthonormal basis of the row space of the original matrix H.

[0070] The translation vector t is:

[0071] t = μ Q - Rμ P ;

[0072] Combine R and t into T0 and apply it to P new = T0P. If the change of T0 is less than the set threshold, or the number of iterations reaches the upper limit, stop the iteration, and the obtained P new is the point cloud data of the target scene after spatial calibration by the ICP algorithm. The calibration of the physical environment information of the target scene and the point cloud data of the target scene is similar to the above operations.

[0073] The timestamp alignment module and the spatial calibration module respectively ensure data synchronization and spatial consistency in the accident discrimination system. The timestamp alignment module uses the Precision Time Protocol (PTP) based on the Global Navigation Satellite System clock to ensure the time synchronization of all devices, thus guaranteeing the accuracy and real-time nature of data acquisition. The spatial calibration module realizes the alignment of spatial coordinates through the Iterative Closest Point (ICP) algorithm to ensure the spatial data consistency between different sensors. The settings of these two modules greatly improve the accuracy of data processing and the overall performance of the system, providing a reliable basis for subsequent data analysis and decision-making.

[0074] The PTP clock synchronization protocol ensures the accuracy and real-time nature of sensor data, environmental perception, and decision-making in intelligent driving by providing high-precision time synchronization, which is crucial for the precision and reliability of the system. At the same time, the ICP spatial registration algorithm can accurately register vehicle sensor data with a high-precision map to achieve precise vehicle positioning and autonomous driving. The combination of these two technologies, on the one hand, ensures the time consistency of data through the PTP protocol, and on the other hand, ensures the accuracy of spatial positioning through the ICP algorithm, jointly enhancing the perception and decision-making capabilities of the driving system in complex environments and enhancing driving safety and reliability.

[0075] In the above technical solution, the accident identification and evaluation module includes:

[0076] The target recognition algorithm training sub-module uses the labeled target scene image data for feature extraction training to obtain the target recognition unit;

[0077] The accident comprehensive information acquisition sub-module is used to obtain the accident comprehensive information according to the image data of the target scene, the physical environment information of the target scene, and the point cloud data of the target scene. The method is as follows:

[0078] Extract the accident features of the target scene image data according to the target recognition unit, and identify whether there are accident features in the image data of the target scene; the target recognition unit is a target recognition unit based on the lightweight YOLO series algorithm;

[0079] According to the image data and point cloud data of the target scene, perform 3D reconstruction on the driving scene through a 3D reconstruction algorithm to obtain a 3D reconstructed driving scene. The 3D reconstruction algorithm uses the MVSNet series algorithm;

[0080] The accident identification module is used to determine whether an accident has occurred in the target scene according to the accident comprehensive information. If it is determined that an accident has occurred, conduct an accident assessment on the accident;

[0081] The accident assessment module is used to classify the accident type of the target scene determined to have an accident according to the accident location of the target scene where the accident occurred by calculating the distance between the accident location and the vehicle body. The accident types are divided into road accidents and vehicle body accidents

[0082] In the above technical solution, in the accident identification and evaluation module, the method for obtaining the target recognition unit by training the target recognition algorithm is as follows:

[0083] Input the labeled target scene image data into the input layer, and the input layer preprocesses the labeled target scene image data. The preprocessing includes fixing the image size and normalization processing;

[0084] The preprocessed image data is subjected to feature extraction at different levels through the backbone network to obtain feature maps of different scales;

[0085] The neck network is used to fuse the feature maps of different scales through upsampling or downsampling fusion to obtain the fused features;

[0086] The detection head is used to predict the feature maps of different scales and the fused features to obtain the probability of an accident occurring in the target scene and the coordinates of the prediction box of the accident; The detection head includes a classification loss function and a localization loss function, and the classification loss function uses an improved cross-entropy loss:

[0087]

[0088] Among them, is the true label value of sample i″ on class c, taking values of 0 or 1, p i″ (c) is the predicted value, representing the probability that the model predicts that sample i" belongs to class c, classes represents all classes of accidents, referring to all classes of all target accidents;

[0089] The localization loss consists of two loss functions, CIoU and DFL. The CIoU loss function is:

[0090]

[0091] Among them, IoU represents the intersection over union between the target prediction box and the ground truth box, b, b gt represent the center points of the prediction box and the ground truth box respectively, and ρ represents the Euclidean distance between the two center points. c1 represents the diagonal length of the smallest circumscribed rectangle that can contain both the prediction box and the ground truth box. a represents the weight coefficient of v, and v represents the similarity of the aspect ratio, defined as:

[0092]

[0093] Among them, w gt and h gt are the width and height of the ground truth box respectively, and w and h represent the width and height of the prediction box respectively.

[0094] The DFL loss function is used for the continuous coordinate prediction of the bounding box, and optimizes the localization accuracy through discrete distribution modeling. Its formula is:

[0095]

[0096] Among them, i 1 is the index variable, representing the position in the sequence, is the predicted value of the network output at position i 1 of the sequence, For the predicted value of the network output at position i 1 which is +1, y is the true continuous coordinate value, where 1 is the value of the i-th reference point in the discretized distribution, 1 is the value of the (i + 1)-th reference point in the discretized distribution, and y, should satisfy

[0097] The loss is calculated using a classification loss function and a localization loss function, and is optimized through backpropagation to obtain the trained object detection algorithm, which is the object recognition unit.

[0098] In one embodiment, the object recognition algorithm uses the YOLOv8n network architecture, as Figure 4 shown. The model structure of YOLOv8n includes an input layer, a backbone network, a neck network, and a detection head.

[0099] The input layer first preprocesses the input labeled target scene image data, including adjusting the size of the image to fit the fixed input size required by the model, usually scaling it to a predetermined size, such as 640×640 pixels; and normalizes the pixel values, mapping them from the range [0,255] to the range [0,1] or [-1,1]. To improve the generalization ability of the model, the input layer also includes data augmentation operations, such as random cropping, flipping, rotation, color jitter, etc.

[0100] The backbone network is responsible for extracting features from the input image. It consists of convolutional layers, residual modules, etc., and is used to gradually extract the feature representation of the image from low-level to high-level. The convolutional layer extracts local features of the image by learning convolutional kernels. It can capture basic features such as edges and textures in the image. The residual block alleviates the gradient vanishing problem in the deep network by introducing residual connections, enabling the network to extract features more deeply.

[0101] The neck network is mainly composed of a Feature Pyramid Network (FPN) and a Path Aggregation Network (PAN). The FPN can fuse high-resolution and low-resolution feature maps by performing upsampling and downsampling operations on feature maps of different scales and then fusing them. The PAN further enhances feature fusion on the basis of the FPN. It improves the feature propagation efficiency by constructing bidirectional feature fusion paths.

[0102] The detection head is used to process the features extracted by the backbone network and the neck network to generate the final detection results. The detection head includes a classification loss and a localization loss. The classification loss function uses an improved cross-entropy loss:

[0103] ​​

[0104] Among them, is the true label value of sample i" on class c, taking values of 0 or 1, p i" (c) is the predicted value, representing the probability that the model predicts sample i″ belongs to class c. classes represents all classes of accidents, referring to all classes of all target accidents; the classes of classes include vehicle collision, rollover, hitting people, etc.

[0105] The localization loss consists of two loss functions, CIoU and DFL. The CIoU loss function is:

[0106]

[0107] Among them, IoU represents the intersection over union between the target prediction box and the ground truth box. b and b gt represent the center points of the prediction box and the ground truth box respectively, and ρ represents the Euclidean distance between the two center points. c1 represents the diagonal length of the smallest bounding rectangle that can contain both the prediction box and the ground truth box. a represents the weight coefficient of v, and v represents the similarity of the aspect ratio, defined as:

[0108]

[0109] Among them, w gt and h gt are the width and height of the ground truth box respectively, and w and h represent the width and height of the prediction box respectively.

[0110] The DFL loss function is used for the continuous coordinate prediction of the bounding box, and optimizes the localization accuracy through discrete distribution modeling. Its formula is:

[0111]

[0112] Among them, i 1 is the index variable, representing the position in the sequence, is the predicted value of the network output at position i 1 ; is the predicted value of the network output at position i 1 +1. y is the true continuous coordinate value, y i1 is the value of the i 1 -th fiducial point in the discrete distribution, is the value of the i 1 +1-th fiducial point in the discrete distribution. y and should satisfy

[0113] The detection head also includes non-maximum suppression operation. For the case where there are multiple detection boxes for the same target, the most reasonable detection box is selected through the non-maximum suppression algorithm.

[0114] The loss is calculated using a classification loss function and a localization loss function, and is optimized through backpropagation to obtain a trained object detection algorithm, which is the object recognition unit.

[0115] The YOLOv8n algorithm belongs to the lightweight algorithms in the YOLO series. It can significantly reduce the consumption of computing resources while ensuring detection accuracy, and is suitable for deployment on in-vehicle edge devices or low-power environments. Through an improved feature fusion mechanism and an Anchor-Free detection strategy, this algorithm effectively solves the problem of missed detection of small targets in complex traffic scenarios, and at the same time enhances the robustness to challenging environments such as occlusion and illumination changes. Its inference speed in milliseconds can process the input of multiple cameras in real time, providing key information such as accident recognition, obstacle detection, and traffic participant tracking for the autonomous driving system.

[0116] According to the image data and point cloud data of the target scene, the driving scene is reconstructed in three dimensions through a three-dimensional reconstruction algorithm to obtain a three-dimensionally reconstructed driving scene. The three-dimensional reconstruction of the image data is as Figure 5 shown, and the point cloud data generated by the lidar is as Figure 6 shown. The three-dimensional reconstruction algorithm uses the MVSNet series of algorithms;

[0117] In one embodiment, the three-dimensional reconstruction algorithm is the EPP-MVSNet algorithm. The network framework of EPP-MVSNet is as Figure 7 shown. The network adopts a multi-stage structure and performs depth prediction from coarse to fine. At each stage k, the depth map D k and the corresponding probability volume P k are inferred through four key steps: feature extraction, cost volume construction, regularization, and regression.

[0118] Given the input reference image I0 and source image I n (I0 represents the left-eye image captured by the binocular camera at the current moment, and I n represents the right-eye image captured by the binocular camera at the current moment and the binocular images captured by the binocular camera at the previous moment), pyramid feature extraction is applied to generate feature maps with coarse or fine spatial resolutions. The feature maps undergo a homography transformation to construct a feature volume.

[0119] The relevant cost volume construction for the coarse and fine stages is carried out using an epipolar aggregation module and an entropy-based improved strategy. In the epipolar aggregation module, the distribution of the original sampling points on the auxiliary view within the default depth hypothesis range is calculated in advance, and new sampling points are adaptively inserted according to its distribution interval. For each reference point p r , its corresponding source point p s at different depth hypotheses d mLower edge polar line discrete sampling. In the coarse stage, with the fixed depth hypothesis range, the interval between sampling source points can be reduced by increasing the number of hypotheses m. To reduce memory and computational costs, the features of the adaptive interval are aggregated onto the sampling points distributed along the polar line, thereby breaking the network efficiency limit brought about by the high-resolution cost volume:

[0120]

[0121] Among them, is expressed as the matching cost smoothed by window width α centered at the depth hypothesis at the m-th pixel point under the s-th view. cost is the original matching cost at the depth position x, is the initial depth estimate or candidate depth value of the m-th pixel point in the depth direction. α represents the integral range window width parameter, x represents the candidate depth value, and Ω(·) represents polar line aggregation. Discretize the equation to implement the polar line aggregation module.

[0122] For the entropy-based improvement strategy, on the basis of predicting the confidence in the last stage, narrow the hypothesis range and reduce redundancy. This module can adaptively adjust the depth hypothesis range of the next stage according to the depth prediction situation of the current stage. Given the probability volume P k , solve the current entropy:

[0123]

[0124] Among them, E represents entropy, k represents the current stage, M0 represents the number of depth hypotheses in the current stage, d represents the corresponding depth hypothesis, P represents the depth prediction probability, p represents the pixel position, is the m-th candidate depth value of pixel p under view k.

[0125] The greater the entropy, the lower the confidence in D k , and naturally the larger the required hypothesis depth range. The hypothesis depth range of the k + 1 stage is determined by the depth range r:

[0126]

[0127] Among them, r represents the depth range; λ represents the adjustment coefficient, controlling the influence intensity of entropy on resolution adjustment, is the average entropy of the depth probability distribution of pixel p under view k.

[0128] After calculating the entropy corresponding to each point on the depth map of the current stage, obtain the depth hypothesis range of the next stage according to the above formula. Since the confidence in the depth map is approximated by simply averaging the entropy of pixels, the adjustment coefficient λ is introduced to adjust the reduction factor of the hypothesis depth range.

[0129] The cost volume is regularized using a lightweight network composed of a pseudo-3D CNN. The cost volume adopts a weighted aggregation method, and soft argmin is used to regress the coarse depth from the cost volume. At the same time, to save computational effort, the visibility map as the weight is only generated in the rough stage, and the weight is reused in the subsequent stages through upsampling.

[0130] Finally, by regressing the regularized cost volume, a depth map and the corresponding probability map are generated for 3D reconstruction. Among them, the 3D reconstruction includes the coordinates of x, y, and z. The x and y coordinates can be obtained according to the calibration parameters of the image and the camera, while the z coordinate needs to be obtained according to the depth map, so as to obtain the 3D reconstructed driving scene.

[0131] The EPP-MVSNet algorithm efficiently generates high-precision point cloud data through multi-view stereo vision technology. The algorithm combines depth estimation and point cloud reconstruction, uses multi-camera perspective images to construct a dense depth map and directly outputs a 3D point cloud, solving the reconstruction accuracy problem in complex road scenes. Through adaptive depth hypothesis adjustment and cost volume optimization, the algorithm can establish a high-precision map and provide reliable 3D spatial information support for path planning.

[0132] Embodiment 2:

[0133] An accident escape system for an autonomous vehicle accident discrimination system based on roadside perception, and the roadside perception module is further used to obtain real-time road traffic information;

[0134] The accident escape system is used to formulate an escape strategy according to the accident type classification and real-time road traffic information, and send an execution instruction to the accident vehicle according to the escape strategy. The escape strategy is specifically as follows:

[0135] For road accidents, path planning is performed to guide the current vehicle into a safe area. The specific method is: using the 3D reconstructed driving scene combined with a random forest to infer the safe area, and through the enhanced Hybrid A* algorithm, planning the optimal path from the current position of the vehicle to the safe area;

[0136] For vehicle body accidents, it is judged whether the vehicle body accident affects the normal driving of the vehicle according to the damage condition of the vehicle. If it affects the normal driving of the vehicle, the accident location is sent to the vehicles within a set range nearby, so that the vehicles within the set range avoid the accident location, and the accident information and the accident location are sent to the rescue agency; if it does not affect the normal driving of the vehicle, path planning is performed to guide the current vehicle into a safe area, the accident information and the accident location are sent to the nearby rescue agency, and the real-time vehicle information is sent to the rescue agency;

[0137] The accident information includes: vehicle position, vehicle damage condition, and information about the people in the vehicle.

[0138] The method of using 3D reconstruction of the driving scene to infer the safe area in combination with a random forest and planning the optimal path from the current vehicle position to the safe area through the enhanced HybridA* algorithm is as follows:

[0139] Obtain the scene information matrix according to the 3D reconstruction of the driving scene, rasterize the scene map into candidate areas based on the scene information matrix, and predict the safety probability grid by grid through a random forest. The prediction formula is:

[0140]

[0141] where N is the number of decision trees, represents the voting result of the i″′-th tree for the grid (x, y), and P safe (x, y) is the safety probability of the grid (x, y).

[0142] Construct the objective function of path planning by combining non-holonomic constraints, obstacle perception distribution, and safe area guidance, obtain the optimal planned path, and incorporate the safety probability P output by the random forest safe into the objective function of path planning:

[0143] J = (1 - P safe )·(ω1L RS + ω2∑V(d0, d v ) + ω3D safe );

[0144] where ω1 is the weight coefficient for controlling the path length, ω2 is the weight coefficient for obstacle avoidance ability, ω3 is the weight coefficient for safe area guidance, P safe is the safety probability of the candidate area, L RS is the shortest path length of the Reeds-Shepp curve, and is a segment of the trajectory in L RS , and the type and direction of satisfy the vehicle kinematic constraints; V(d0, d v ) is the Voronoi field distance of the dynamic obstacle, and e represents the natural constant, d0 is the distance from the path point to the nearest obstacle, d v is the distance from the path point to the nearest Voronoi edge, and β is the parameter for controlling the potential energy decay rate; D safe is the Euclidean distance between the safe area target point and the centroid of the safe area.

[0145] Through the above method, the value of the objective function J of each candidate area can be calculated to guide the path planning algorithm to find the optimal path.

[0146] Random forests perform multi-dimensional modeling on scene information by integrating multiple decision trees and output the probability distribution of the safe area. The safe area needs to meet the requirements of being far from the accident point (to avoid secondary collisions), having sufficient roadside space (such as emergency stopping belts, hard shoulders), no dynamic obstacles (such as oncoming vehicles, pedestrians), and conforming to vehicle kinematic constraints (such as turning radius, slope requirements), etc. Different types of data need to be encoded with different methods. For numerical features, such as slope and curvature, they can be directly input into the model; for categorical features, such as obstacle types, One-Hot encoding is used; for spatio-temporal features, such as obstacle trajectory prediction, temporal features are extracted through a sliding window.

[0147] The model input includes a scene information matrix containing the above features. Key features are selected through the Gini coefficient or permutation importance; cost-sensitive learning is adopted to solve the problem of sample imbalance between safe and non-safe areas in accident data; and K-Fold cross-validation is used to prevent overfitting and adjust the depth and number of decision trees. The map is rasterized into candidate areas, and the safe probability is predicted grid by grid through the model:

[0148]

[0149] where N is the number of decision trees, represents the voting result of the i-th tree for the grid (x, y), and P safe (x, y) is the safe probability of the grid (x, y). For example, if the distance to the right guardrail is ≥ 3m and there are no dynamic vehicles within 200m ahead, the model marks the right hard shoulder as a high-safe area. Combining the signal light status and pedestrian trajectory prediction, a right-turn lane that is far from the crosswalk and unoccupied is selected as the safe area.

[0150] The enhanced Hybrid A* algorithm uses a 3D state space (x, y, θ) to describe the vehicle pose. Each node generates child nodes with multiple steering angles, and the reachability is verified through the vehicle kinematic model. Combining non-holonomic constraints (the shortest path length of the Reeds-Shepp curve: where each segment of the trajectory The type and its direction need to meet the vehicle kinematic constraints, and the total path needs to belong to one of the 48 combinations of Reeds-Shepp), the obstacle perception distribution (the Voronoi field distance of dynamic obstacles: where e is the natural constant, d0 is the distance from the path point to the nearest obstacle, d v is the distance from the path point to the nearest Voronoi edge, and β is a parameter controlling the potential energy decay rate.) and the safe area guidance (the Euclidean distance between the target point and the centroid of the safe area: where (x g , y g ) are the coordinates of the target point, (xc ,y c ) is the coordinate of the center of mass of the safe area) to build a heuristic objective function and dynamically select the optimal path direction. Combined with the above model, the safety probability P output by the random forest is safe Incorporate the objective function of path planning:

[0151] J=(1-P safe )·(ω1L RS +ω2∑V(d0,d v )+ω3D safe );

[0152] Where ω1 is the weight coefficient of the control path length, ω2 is the weight coefficient of the obstacle avoidance capability, ω3 is the weight coefficient of the safety area guidance, P safe is the safety probability of the candidate area;

[0153] The discrete path points generated by the enhanced Hybrid A* algorithm are smoothed by fitting the B-spline curve to minimize the curvature change rate. Path deviation is allowed in areas with high safety probability, and the gradient descent method is introduced to perform nonlinear optimization on the original trajectory. According to the objective function defined above and through conjugate gradient interpolation, a continuous smooth path is generated to reduce redundant steering and dynamically adjust the position of the trajectory points in real time. For example, after the vehicle detects an accident ahead, the random forest marks the right shoulder as a safe area, and the enhanced Hybrid A* plans a path to change lanes to the right and slow down to stop, and obtains path characteristics, such as a maximum curvature of ≤0.2rad / m, to avoid loss of control due to sharp turns.

[0154] By integrating real-time scene data with the random forest algorithm, safe areas can be accurately identified, and the enhanced HybridA* algorithm combined with vehicle dynamics constraints can be used to generate obstacle avoidance paths. In complex accident scenarios, the optimal driving trajectory that takes into account both safety and efficiency can be quickly planned, effectively improving the intelligent driving system's autonomous escape capability and accident response reliability in emergency situations, and reducing the risk of secondary accidents.

[0155] like Figure 8As shown in the figure, when the roadside sensing device identifies an accident on the road ahead, the accident type of the current accident is obtained through the accident assessment module, and it is divided into a road accident or a body accident: if it is a road accident, the enhanced HybridA* algorithm is used to evaluate the safety and feasibility of different escape paths for unmanned vehicles on this road section, select the best path and generate specific operation instructions; at the same time, the Internet of Vehicles communication protocol, combined with the roadside equipment, monitors the safe driving of the vehicle; if it is a body accident, the accident information of the vehicle itself is obtained through the roadside equipment and the on-board equipment to further determine whether it affects the normal driving of the vehicle. If a minor accident does not affect driving, the machine learning model is used to infer the safe area, and the enhanced Hybrid A* algorithm is used to plan the optimal path to the safe area, control the sending of instructions to let the vehicle go to the safe area, and transmit the accident information to the unmanned vehicles and rescue agencies in the nearby area in real time; if a major accident occurs, causing the vehicle to be unable to drive normally, the accident information is directly transmitted to the unmanned vehicles and rescue agencies in the nearby area to wait for assistance.

[0156] In the above technical solution, the accident escape system uses a steering wheel control technology in the process of path planning to guide the current vehicle into a safe area. The steering wheel control technology is used to control the vehicle to travel according to the optimal path to enter the safe area using an angle control network. The angle control network includes a perception module and a decision module, including:

[0157] The perception module is responsible for processing the input actual driving images, extracting and integrating visual features through the trained optimal model of the steering angle control network, and generating the predicted value of the steering wheel angle with the integrated visual features;

[0158] The corner control network includes an initial layer, stages 1 to 4, and an NCP module;

[0159] Among them, the initial layer is used to initially extract multi-scale features and contains two parallel convolutional layers;

[0160] Stages 1 to 4 are used to gradually extract multi-scale perceptual features. They include four stages. The multi-scale features extracted by the initial layer are input into stage 1 and transmitted downward layer by layer. The resolution of each stage is 1 / 2 of the previous layer. Each stage includes an upper branch inverse residual block, a lower branch inverse residual block, upsampling, downsampling, and superposition fusion operations.

[0161] The NCP module is used to make decisions based on the features extracted from stages 1 to 4 to obtain the steering wheel angle prediction. It includes two branches, upper and lower. Each branch contains four layers of neurons, namely perception neurons, intermediate neurons, command neurons, and motor neurons.

[0162] Among them, the sensory neurons are used to receive feature inputs and preliminarily encode information; the intermediate neurons are used to extract abstract features based on the encoded information; the command neurons are used to generate control commands using the abstract features; the motor neurons are used to output the final predicted value of the steering wheel angle;

[0163] The decision-making module is used to generate corresponding control commands based on the predicted value of the steering wheel angle.

[0164] The driving images are obtained through the binocular cameras on the vehicle.

[0165] In one embodiment, as Figure 9 shown, the lightweight angle control network consists of an initial layer, stages 1-4, and an NCP module. Among them, the initial layer and stages 1-4 constitute the perception feature extraction, and the NCP module makes decisions based on the feature extraction results. The initial layer contains two 3×3 convolutional layers with a stride of 2, which are used for downsampling and outputting two feature maps with different resolutions and numbers of channels. Stages 1-4 include 5×5 inverted residual blocks (IRBs) in the upper branch, 3×3 IRBs in the lower branch, upsampling, downsampling, and additive fusion operations. The IRBs in each stage extract features in the upper and lower branches respectively and perform fusion. DW Conv N×N represents depthwise separable convolution. The downsampling operation is shown by the orange dashed box, and the upsampling operation is shown by the dark green dashed box. Each stage uses a 2×2 convolution with a stride of 2 for downsampling. The NCP module also contains an upper branch and a lower branch. In each branch, the NCP consists of sensory neurons, intermediate neurons, command neurons, and motor neurons. NCP M means that the NCP has M neurons, and the number of neurons in each layer and its detailed structure are shown in the figure.

[0166] The perception module of the multi-layer neural network of the lightweight angle control network can efficiently extract and integrate visual features. Its design is lightweight, and the operation and processing are rapid and timely, which can quickly and accurately process the driving images. It can improve the perception and response speed of the driving system to the road conditions, generate accurate predicted values of the steering wheel angle, and the decision-making module then converts it into accurate control commands, thereby improving the vehicle's ability to cope with complex environments and better ensuring driving safety.

[0167] In the above technical solution, the method for obtaining the optimal model of the trained corner control network is: using the dataset composed of historical road images and real angles to train the angle control network, and using the mean square error as the loss function. The formula is:

[0168]

[0169] where p 0 represents the steering wheel angle predicted by the network, r 0 represents the actual angle during driving, m0 represents the number of images or steering wheel angles, represents the predicted value of the i′th image or steering wheel angle, represents the true value of the i′th image or steering wheel angle, is the square of the L2 norm;

[0170] During the training process, the gradient descent method is used to minimize the MSE and update the parameters to obtain the optimal model of the angle control network.

[0171] In one embodiment, if Figure 10 As shown in the figure, the optimal path generated by the escape strategy is used to control the driving of the unmanned vehicle, in which the core control of the steering wheel angle is realized by the lightweight corner control network. The lightweight corner control network uses computer vision technology to perform deep feature extraction on the input image and identify key features in the road, including lane lines, obstacles, and road curvature. The nonlinear mapping capability of the angle prediction model is used to directly generate the predicted value of the steering wheel angle. To ensure the accuracy of the steering wheel angle prediction, the system designs an end-to-end error feedback and optimization mechanism. In the training stage, the lightweight corner control network receives training data containing the driver's perspective image and the corresponding real steering wheel angle, generates the predicted steering wheel angle through forward propagation, and calculates the error (such as mean square error, MSE) after comparing it with the real value. Through the back propagation algorithm, the error is used to optimize the model parameters so that the network gradually improves the prediction accuracy. The training stage is carried out in multiple rounds of iterations until the lightweight corner control network reaches the optimal state and the trained model parameters are saved. The lightweight corner control network does not require a large number of parameters, can be easily implemented in vehicle deployment, and meets real-time requirements. During the testing phase, the system receives real-time road images captured by the camera and inputs them into the trained lightweight steering angle control network model. The model generates the target steering wheel angle signal based on the input image and evaluates and verifies it. At the same time, the final steering angle signal is transmitted to the vehicle's steering wheel control module, driving the steering wheel motor to make real-time adjustments to ensure the speed and accuracy of the vehicle's steering decision.

[0172] Embodiment 3:

[0173] An accident responsibility judgment system for unmanned vehicle accident identification based on roadside perception, which judges the accident responsibility of vehicles involved in body accidents. The accident responsibility judgment is based on the comprehensive accident information before and at the current moment of the accident combined with traffic regulations and historical traffic accident responsibility judgment rules, and the accident responsibility is divided to generate the accident responsibility judgment result;

[0174] The method for dividing accident liability and generating an accident liability judgment result is as follows: Through time series analysis, analyze the acceleration and speed changes of the vehicle to identify abnormal behaviors before and after the accident; use the trajectory backtracking algorithm to reconstruct the driving path before and after the accident, combine the traffic regulation knowledge base and abnormal behaviors, and use a rule engine to determine whether there are any violations in the vehicle's behavior. Determine the liability of the vehicle with violations in the accident as one of unilateral liability, joint liability, or no liability, determine the liability ratio of each party with joint liability, and generate an accident liability judgment result based on the liability division.

[0175] The accident liability judgment system realizes the automatic division of accident liability by comprehensively analyzing the image data of the target scene, the physical environment information of the target scene, and the point cloud data of the target scene at the previous moment and the current moment, in combination with accident type classification and relevant traffic regulations. Through time series analysis and trajectory backtracking algorithm, it judges whether there are any violations in the vehicle's behavior, reduces the subjective deviation caused by human intervention, ensures the fairness, transparency of the judgment process, enhances interpretability, and improves the claims settlement efficiency.

[0176] In the above technical solution, a communication module is further included, and the communication module is used to transmit the accident liability judgment result to a third-party claims settlement and loss assessment point and an insurance company for vehicle loss assessment, repair, and compensation.

[0177] Embodiment 4:

[0178] A method for discriminating accidents of driverless vehicles based on roadside perception includes the following steps:

[0179] Obtain the target scene image data, the target scene physical environment information, and the target scene point cloud data;

[0180] Use the target recognition unit to extract the target scene accident features in the target scene image data, obtain a three-dimensional reconstructed driving scene based on the target scene image data and the target scene point cloud data, form accident comprehensive information according to the target scene accident features, the three-dimensional reconstructed driving scene, and the target scene physical environment information, and judge whether there is an accident in the target scene according to the accident comprehensive information; if there is an accident in the target scene, classify the accident type according to the distance between the accident location and the vehicle body.

[0181] Embodiment 5:

[0182] A computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0183] The content not described in detail in this specification belongs to the prior art well-known to those skilled in the art. Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0184] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or combinations of blocks.

[0185] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or combinations of blocks.

[0186] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or combinations of blocks.

[0187] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than limit the scope of its protection. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that after reading the present invention, various changes, modifications or equivalent replacements can still be made to the specific implementation manners of the invention. However, these changes, modifications or equivalent replacements are all within the scope of protection of the claims pending for the invention.

[0188] The content not described in detail in this specification belongs to the prior art well known to those skilled in the art.

Claims

1. An accident discrimination system for driverless vehicles based on roadside perception, characterized in that, Including: The roadside perception module is used to obtain target scene image data, target scene physical environment information, and target scene point cloud data; The accident recognition and evaluation module is used to extract the target scene accident features in the target scene image data by using the target recognition unit, obtain the three-dimensional reconstructed driving scene according to the target scene image data and the target scene point cloud data, form the accident comprehensive information according to the target scene accident features, the three-dimensional reconstructed driving scene, and the target scene physical environment information, and judge whether there is an accident in the target scene according to the accident comprehensive information; if there is an accident in the target scene, the accident is classified into accident types according to the distance between the accident location and the vehicle body.

2. The accident discrimination system for driverless vehicles based on roadside perception according to claim 1, characterized in that, For the roadside perception module, the image data of the target scene is obtained from the image data of the binocular camera sensor through the timestamp alignment module and the spatial calibration module; the physical environment information of the target scene is obtained from the millimeter wave radar data of the millimeter wave radar sensor through the timestamp alignment module and the spatial calibration module; the point cloud data of the target scene is obtained from the three-dimensional reconstruction of the image data of the binocular camera sensor and the lidar data of the lidar sensor through the timestamp alignment module and the spatial calibration module, where: Timestamp alignment module, the timestamp alignment module performs timestamp alignment through the PTP clock synchronization protocol; Spatial calibration module, the spatial calibration module performs spatial calibration through the ICP algorithm.

3. The accident discrimination system for driverless vehicles based on roadside perception according to claim 1, wherein For the accident recognition and evaluation module, including: The target recognition algorithm training sub-module uses the labeled target scene image data to perform feature extraction training to obtain the target recognition unit; The accident comprehensive information acquisition sub-module is used to obtain the accident comprehensive information according to the target scene image data, the target scene physical environment information, and the target scene point cloud data. The method is as follows: Extract the accident features of the target scene image data according to the target recognition unit, and identify whether there are accident features in the target scene image data; the target recognition unit is a target recognition unit based on the lightweight YOLO series algorithm; According to the target scene image data and point cloud data, perform three-dimensional reconstruction of the driving scene through the three-dimensional reconstruction algorithm to obtain the three-dimensional reconstructed driving scene, and the three-dimensional reconstruction algorithm uses the MVSNet series algorithm; The accident recognition module is used to judge whether an accident occurs in the target scene according to the accident comprehensive information. If it is judged as an accident, the accident is evaluated; The accident evaluation module is used to classify the accident types of the target scene judged to have an accident according to the accident location of the target scene where the accident occurs by calculating the distance between the accident location and the vehicle body. The accident types are divided into road accidents and vehicle body accidents.

4. The accident discrimination system for driverless vehicles based on roadside perception according to claim 3, wherein In the accident recognition and evaluation module, the method for obtaining the target recognition unit by training the target recognition algorithm is: Input the labeled target scene image data into the input layer, and the input layer preprocesses the labeled target scene image data. The preprocessing includes fixing the image size and normalization processing; perform feature extraction at different levels on the preprocessed image data through the backbone network to obtain feature maps of different scales; Feature fusion of feature maps of different scales is performed through upsampling or downsampling fusion using a neck network to obtain fused features. A detection head is used to predict the feature maps of different scales and the fused features to obtain the probability of an accident occurring in the target scenario and the coordinates of the predicted bounding box of the accident. The detection head includes a classification loss function and a localization loss function. The classification loss function uses an improved cross-entropy loss: Among them, is the true label value of sample i″ on class c, taking values of 0 or 1, p i″ (c) is the predicted value, representing the probability that the model predicts sample i″ belongs to class c; classes represents all classes of accidents; The localization loss consists of two loss functions, CIoU and DFL. The CIoU loss function is: Among them, IoU represents the intersection over union between the target prediction box and the ground truth box. b and b gt represent the center points of the prediction box and the ground truth box respectively, ρ represents the Euclidean distance between the two center points, c1 represents the diagonal length of the smallest bounding rectangle that can contain both the prediction box and the ground truth box, a represents the weight coefficient of v, and v represents the similarity of the aspect ratio, which is defined as: Among them, w gt and h gt are the width and height of the ground truth box respectively, and w and h represent the width and height of the predicted box respectively; The DFL loss function is used for continuous coordinate prediction of the bounding box, and optimizes the localization accuracy through discretized distribution modeling. Its formula is: where i 1 is an index variable representing a position in the sequence, is the predicted value of the network output at position i 1 , is the predicted value of the network output at position i 1 +1, y is the true continuous coordinate value, is the value of the i 1 -th reference point in the discretized distribution, is the value of the (i 1 +1)-th reference point in the discretized distribution, y, should satisfy The classification loss function and the localization loss function are used for loss calculation and optimized through backpropagation to obtain the trained object detection algorithm. The trained object detection algorithm is the object recognition unit.

5. An accident escape system for the driverless vehicle accident discrimination system according to claim 1, characterized in that, The roadside perception module is also used to obtain real-time road traffic information. The accident escape system is used to formulate an escape strategy based on accident type classification and real-time road traffic information, and send an execution instruction to the accident vehicle according to the escape strategy. The escape strategy is specifically: For road accidents, path planning is performed to guide the current vehicle into a safe area. The specific method is: using 3D reconstruction of the driving scene combined with a random forest to infer the safe area, and planning the optimal path from the current position of the vehicle to the safe area through the enhanced Hybrid A* algorithm. For vehicle body accidents, it is judged whether the vehicle body accident affects the normal driving of the vehicle according to the damage condition of the vehicle. If it affects the normal driving of the vehicle, the accident location is sent to the vehicles within a set range nearby, so that the vehicles within the set range avoid the accident location, and the accident information and accident location are sent to the rescue agency. If it does not affect the normal driving of the vehicle, path planning is performed to guide the current vehicle into a safe area, the accident information and accident location are sent to the nearby rescue agency, and the real-time information of the vehicle is sent to the rescue agency. The accident information includes: vehicle location, vehicle damage condition, and information about the people in the vehicle. The method of using 3D reconstruction of the driving scene combined with a random forest to infer the safe area and planning the optimal path from the current position of the vehicle to the safe area through the enhanced Hybrid A* algorithm is: Obtain the scene information matrix according to the 3D reconstruction of the driving scene, rasterize the scene map into candidate areas according to the scene information matrix, and predict the safety probability grid by grid through a random forest. The prediction formula is: where N is the number of decision trees, represents the voting result of the i″′-th tree for the grid (x, y), and P safe (x, y) is the safety probability of the grid (x, y); Construct the objective function of path planning by combining non - holonomic constraints, obstacle perception distribution, and safety region guidance, obtain the optimal planned path, and incorporate the safety probability P output by the random forest safe into the objective function of path planning: J = (1 - P safe ) · (ω1L RS + ω2ΣV(d0, d v ) + ω3D safe ); Among them, ω1 is the weight coefficient for controlling the path length, ω2 is the weight coefficient for obstacle avoidance ability, ω3 is the weight coefficient for safety area guidance, P safe is the safety probability of the candidate area, L RS is the shortest path length of the Reeds-Shepp curve, and is a section of the trajectory in L RS , and the type and direction of satisfy the vehicle kinematic constraints; V(d0, d v ) is the Voronoi field distance of the dynamic obstacle, and e is the natural constant, d0 is the distance from the path point to the nearest obstacle, d v is the distance from the path point to the nearest Voronoi edge, β is the parameter for controlling the potential energy decay rate; D safe is the Euclidean distance between the safety area target point and the centroid of the safety area; Calculate the value of the objective function J for each candidate area to guide the path planning algorithm to find the optimal path.

6. The accident escape system according to claim 5, characterized in that During the process of the accident escape system guiding the current vehicle into a safe area through path planning, a steering wheel control technology is used. The steering wheel control technology is used to control the vehicle to drive according to the optimal path into the safe area using an angle control network. The angle control network includes a perception module and a decision module, including: The perception module is responsible for processing the input driving images, extracting and integrating visual features through the trained optimal model of the angle control network, and generating a predicted value of the steering wheel angle from the integrated visual features. The angle control network includes an initial layer, stages 1 to 4, and an NCP module. Among them, the initial layer is used to initially extract multi-scale features and includes two parallel convolutional layers; Stages 1 to 4 are used to gradually extract multi-scale perception features. It includes 4 stages. The multi-scale features extracted by the initial layer are input into Stage 1 and passed down layer by layer. The resolution of each stage is 1 / 2 of the previous layer. Each stage includes an upper-branch inverted residual block, a lower-branch inverted residual block, upsampling, downsampling, and superposition fusion operations; The NCP module is used to make decisions on the features extracted in Stages 1 to 4 to obtain the steering wheel angle prediction. It includes two branches, the upper and the lower. Each branch contains four layers of neurons, namely, perception neurons, intermediate neurons, instruction neurons, and motion neurons; Among them, the perception neurons are used to receive feature inputs and initially encode information; the intermediate neurons are used to extract abstract features based on the encoded information; the instruction neurons are used to generate control instructions using the abstract features; the motion neurons are used to output the final steering wheel angle prediction value; The decision module is used to generate corresponding control instructions based on the predicted value of the steering wheel angle.

7. The accident escape system according to claim 6, wherein The method for obtaining the optimal model of the trained corner control network is as follows: Use the dataset composed of historical road images and true angles to train the angle control network. Use the mean square error as the loss function, and the formula is: Among them, p 0 represents the steering wheel angle predicted by the network, r 0 represents the actual angle during driving, m0 represents the number of images or steering wheel angles, represents the predicted value of the i'-th image or steering wheel angle, represents the true value of the i'-th image or steering wheel angle, is the square of the L2 norm; During the training process, the gradient descent method is used to minimize the MSE and update the parameters, so as to obtain the optimal model of the angle control network.

8. An accident liability judgment system for accident discrimination according to claim 1, characterized in that, For the vehicle that has a vehicle body accident, determine the accident liability. The accident liability judgment is based on the accident comprehensive information at the previous moment and the current moment before the accident, combined with traffic regulations and historical traffic accident liability judgment rules, for comparison, divide the accident liability, and generate the accident liability judgment result; The method for dividing the accident liability and generating the accident liability judgment result is as follows: Through time series analysis, analyze the acceleration and speed changes of the vehicle to identify abnormal behaviors before and after the accident; Use the trajectory backtracking algorithm to reconstruct the driving path before and after the accident. Combine the traffic regulation knowledge base and abnormal behaviors, and use the rule engine to judge whether there are any violations in the vehicle behavior. Determine the liability of the vehicle with violations in the accident as one of unilateral liability, joint liability, or no liability, determine the liability ratio of each party with joint liability, and generate the accident liability judgment result according to the liability division.

9. A method for judging an accident of a driverless vehicle based on roadside perception, characterized in that, It includes the following steps: Obtain the target scene image data, target scene physical environment information, and target scene point cloud data; Use the target recognition unit to extract the target scene accident features in the target scene image data. According to the target scene image data and the target scene point cloud data, obtain the three-dimensional reconstructed driving scene. According to the target scene accident features, the three-dimensional reconstructed driving scene, and the target scene physical environment information, form the accident comprehensive information, and judge whether there is an accident in the target scene according to the accident comprehensive information; If there is an accident in the target scene, classify the accident type according to the distance between the accident location and the vehicle body.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method described in Claim 9.

Images