Digital reconstruction method of road traffic accident
Patent Information
- Application Number
- CN202610412628.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-31
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies for investigating road traffic accidents suffer from problems such as road closures affecting traffic efficiency, difficulty in obtaining information on damage at the scene, and difficulty in obtaining vehicle operation information. Furthermore, they lack comprehensive utilization of vehicle operation information prior to accidents.
The system uses drone aerial photography and point cloud scanning to acquire accident scene data, combines vehicle-road cooperative system and Transformer model to perform panoramic segmentation, extract traffic participant information, simulate the accident process in collision simulation software, use intelligent vehicle unit to acquire vehicle operating status, and adjust model parameters to reproduce the accident.
It has improved the automation and intelligence of accident scene investigation, reduced the closure time, fully reproduced the accident process, and clarified responsibilities.
Smart Images

Figure CN122289278A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent transportation technology, and in particular to a method for digitally reproducing road traffic accidents. Background Technology
[0002] With the significant increase in the number of motor vehicles, coupled with factors such as road traffic conditions, weather conditions, and distracted driving, road traffic accidents occur frequently. Current methods for investigating road traffic accidents mainly involve road closures and on-site measurements and photography. These methods have the following shortcomings: First, they require road closures, which affects traffic efficiency; second, conducting investigations at the accident scene can easily damage information about the accident; and third, after a collision, it is difficult to obtain vehicle movement information, which is detrimental to the subsequent determination of liability and insurance claims.
[0003] Chinese invention patent publication number CN112446958A discloses a road traffic accident auxiliary processing method based on laser point cloud. The method uses a high-precision three-dimensional laser scanner to scan the accident scene and establish a three-dimensional model of the accident scene based on the acquired image data and three-dimensional point cloud data. It also measures the collision marks on the vehicle body to estimate the instantaneous speed of the car before the collision, thereby shortening the traffic congestion time.
[0004] Chinese invention patent publication CN120124451A discloses an intelligent reconstruction method for road traffic accidents based on automated Pc-Crash simulation and MLP algorithm. The method involves creating a road model in Pc-Crash simulation software based on an overview map of the accident scene taken at the accident site, and creating vehicle models in the same software based on vehicle information obtained from the accident scene investigation. Then, an accident simulation dataset (such as vehicle collision speed, angular velocity, etc.) is designed, and a multilayer perceptron neural network model is constructed to simulate and output results such as vehicle tire friction marks and the stopping positions of accident participants. The method is considered reliable when the average positional error between the simulation and the actual survey results for vehicle and pedestrian stopping positions is less than 5%.
[0005] Chinese invention patent publication number CN120495503A proposes a three-dimensional reconstruction system for road traffic accidents. This system collects environmental and evidence data from the scene of a road traffic accident, establishes a three-dimensional scene model of the accident scene, and then reconstructs the accident process to obtain the operational data of traffic participants at the accident scene and three-dimensional animation.
[0006] Chinese invention patent publication number CN107194989A proposes a three-dimensional reconstruction system and method for traffic accident scenes based on drone aerial photography. The method sets the flight route and waypoints of the drone to collect images of the road traffic accident scene from different altitudes and angles, and then uses multi-view stereo vision processing technology to establish a three-dimensional traffic accident scene model.
[0007] Currently, drone aerial photography and lidar scanning technologies are being applied to road traffic accident scene investigation. Digital reconstruction of road traffic accidents requires not only road, traffic, and environmental information, but also vehicle operational information prior to the accident. However, existing technologies do not adequately utilize vehicle operational information prior to accidents, often extracting accident information from single roadside cameras / radar or drones, with limited development of digital reconstruction methods from a vehicle-road cooperative perspective. Summary of the Invention
[0008] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a method for digital reconstruction of road traffic accidents.
[0009] Therefore, the present invention adopts the following technical solution:
[0010] A method for digitally reconstructing road traffic accidents includes the following steps:
[0011] S1, to take aerial photos and point cloud scans of the accident site to obtain aerial images and point cloud data;
[0012] S2, preprocessing the aerial images obtained in S1;
[0013] S3 uses a Transformer-based panoramic segmentation model for traffic accident scenes to perform panoramic segmentation on the aerial images obtained in S2, extracting the masks, types, and number of traffic participants at the accident scene.
[0014] S4 projects the mask of traffic participants obtained in S3 onto the three-dimensional coordinate system of the point cloud data obtained in S1 to help determine the point set contained in each traffic participant, thereby calculating the relative spatial position of each traffic participant at the accident scene.
[0015] S5, obtain comprehensive vehicle and road information:
[0016] Based on the communication status of the vehicle-road cooperative system, information on road traffic flow, traffic signal control, and road channelization design is determined through roadside units;
[0017] If all the vehicles involved in the accident are intelligent vehicles, the operating status information of the vehicles involved in the accident will be obtained directly from the on-board unit installed on the vehicle to determine the cause of the accident and to assign responsibility.
[0018] If all the vehicles involved in the accident are traditional vehicles and are not equipped with an on-board unit, proceed to step S61; if one of the vehicles involved in the accident is an intelligent vehicle and the other is a traditional vehicle, obtain the operating status information of the intelligent vehicle through the on-board unit and proceed to step S62.
[0019] S6, Traffic Accident Simulation Analysis:
[0020] S61, in the traffic accident collision simulation software, establish an accident collision mathematical model, and continuously adjust the speed and heading angle of the accident vehicles in the model to perform simulation optimization until the absolute error between the simulated distance after the accident vehicle collision and the relative spatial distance obtained in S4 is found. 5%, completing the reconstruction of road traffic accidents;
[0021] S62, in traffic accident collision simulation software, establishes a mathematical model of accident collision to simulate the process of a traffic accident. By continuously adjusting the speed and heading angle of the traditional vehicles in the model, simulation optimization is performed until the absolute error between the simulated distance after the collision and the relative spatial distance obtained in S4 is found. 5%, completing the reconstruction of road traffic accidents.
[0022] The specific operation of step S2 above is as follows:
[0023] Gamma correction is used to improve the quality of aerial images obtained from S1. The correction formula is as follows:
[0024]
[0025] in, Aerial images; These are corrected aerial images; To correct parameters, if the aerial image has a high exposure, If the aerial image has low exposure, .
[0026] The specific operation of step S3 above is as follows:
[0027] S31, Build a panoramic segmentation model of a traffic accident scene based on Transformer:
[0028] The Transformer-based panoramic segmentation model for traffic accident scenes includes a feature enhancement backbone, a pixel decoder, and a Transformer decoder, which are used to process the three-channel images of the input model. The feature enhancement backbone is obtained by introducing a balanced pooling attention module to improve the original backbone of the Mask2Former model.
[0029] S32, training the panoramic segmentation model of the traffic accident scene built in S31:
[0030] A weighted loss function is constructed. Based on the weighted loss function, virtual accident scene data containing extreme weather is generated using an open-source autonomous driving simulation platform. Noise simulating rain, snow, obstruction, and fog is injected into the input. The panoramic segmentation performance of the traffic accident scene panoramic segmentation model in complex environments is optimized by using adversarial training to complete the training.
[0031] S33, use the trained panoramic segmentation model of the traffic accident scene to perform panoramic segmentation on the preprocessed aerial image obtained in S2, and then extract the mask, type and number of traffic participants at the accident scene, including accident vehicles, non-motorized vehicles and pedestrians.
[0032] In step S31 above:
[0033] The process of processing the three-channel image of the input model is as follows:
[0034] (1) The feature enhancement backbone processes the input three-channel image to obtain four multi-scale feature maps. The feature enhancement backbone includes a Swin-Tiny network and four balanced pooling attention modules, wherein:
[0035] The Swin-Tiny network includes a Patch Partition module and four hierarchical stages (Stage 1 to Stage 4). First, the Patch Partition module processes the input image, converting the number of channels to 48 and reducing the width and height to match the input image. Then, feature maps are output through the concatenated Stages 1 through Stage 4. ;
[0036] The balanced pooling attention module is used to process the feature maps respectively. Feature extraction is performed to obtain feature maps. ;
[0037] Further analysis of feature maps Multi-scale fusion is performed to obtain new multi-scale features. Specifically: for ; for The upsampling results and Weighted summation; for The upsampling results and Weighted summation; for The upsampling results and Weighted summation;
[0038] (2) Feature map output by the feature enhancement backbone Feature maps obtained after Conv4d processing The length and width remain the same, but the number of channels becomes 256;
[0039] (3) The pixel decoder processes the input feature map Multi-scale fusion is performed to obtain feature maps. With feature map :
[0040] For feature maps A 6-layer multi-scale deformable attention transformer is applied for processing; the width, height, and number of channels of the output features remain unchanged after processing. Then, the feature map is obtained through upsampling, weighted summation, and convolution. Feature map Compared with the upsampled feature map Perform a weighted sum, then... Convolution and Convolution processing yields feature maps ;
[0041] (4) The Transformer decoder has 9 stacked Transformer decoder layers, initialized with a set of learnable query vectors, including Each query receives feature maps from every three decoder layers. As a pixel embedding, it receives the query vector output from the previous decoder layer, processes it through Masked Cross-Attention, Self-Attention, and FFN, and outputs the updated query vector; each query has a corresponding mask image and corresponding category;
[0042] (5) The final query vector output by the Transformer decoder is fed into two fully connected layers. One layer is used for classification to obtain the corresponding category and category probability. The other layer is mapped to the mask embedding vector and the feature map. Perform a dot product to obtain the mask image.
[0043] In step S31 above:
[0044] Stages 1 through 4 contain stacked 2, 2, 6, and 2 Swing Transformer blocks, respectively. In Stage 1, the Linear Embedding layer processes the input feature map, increasing the number of channels. The output feature map is obtained, and then features are extracted through the Swing Transformer block to form the output of Stage 1. Stages 2 through 4 use the PatchMerging layer to downsample the input feature maps, reducing their width and height to their original values. The number of channels is expanded to twice the input.
[0045] In step S31 above:
[0046] The balanced pooling attention module processes the input features. Average pooling and max pooling are performed using horizontal and vertical pooling kernels respectively, generating four one-dimensional feature maps. Then, a fused one-dimensional horizontal feature is obtained by performing a one-dimensional convolution with a kernel size of 3 and weighted summation. and one-dimensional vertical features Then the two are connected along the spatial dimension to form ,pass Convolution compresses the number of channels, then splits them along the spatial dimension into and ,pass Convolution and spatial broadcasting operations respectively connect the two to the initial input features. Having the same number of channels and spatial dimensions, then applying them... Activation function generates attention weights and Finally, the original input features With weight , Element-wise multiplication yields the enhanced output features. .
[0047] In step S32 above:
[0048] The formula for the weighted loss function is as follows:
[0049] ;
[0050] in: For classification loss, the corresponding weights ; For mask loss; For boundary-aware loss, the corresponding weights The value range is 0.2-0.5;
[0051] The formula for the boundary-aware loss function is:
[0052] ;
[0053] in:
[0054] ;
[0055] ;
[0056] ;
[0057] ;
[0058] expansion With corrosion All operations use structural elements;
[0059] It is the continuous probability mask of the model's output prediction. Refers to the actual mask;
[0060] It should be a small positive number to prevent the denominator from being 0.
[0061] In step S5 above:
[0062] The vehicle operation status information refers to the vehicle's operation status information 10 to 20 seconds before the traffic accident.
[0063] The operational status information includes position, speed, acceleration, heading angle, and braking.
[0064] The specific operation of step S61 above is as follows:
[0065] Based on the type of the accident vehicle and the road traffic flow, traffic signal control, and road channelization design information obtained from S5, a mathematical model of the accident collision is established by setting vehicle type, road, weather, and traffic signal information in the traffic accident collision simulation software.
[0066] The specific operation of step S62 above is as follows:
[0067] Based on the type of vehicle involved in the accident, the road traffic flow obtained from S5, traffic signal control, road channelization design information, and the operating status information of the intelligent vehicle, a mathematical model of the accident collision is established in the traffic accident collision simulation software by setting vehicle type, road, weather, and traffic signal information.
[0068] Compared with the prior art, the present invention has the following beneficial effects:
[0069] 1. This invention proposes a panoramic segmentation model of traffic accident scene based on Transformer, which overcomes the problems of "unseen", "unclear", and "incomplete" targets at road traffic accident scene, thereby improving the automation and intelligence level of accident scene investigation.
[0070] 2. This invention leverages the advantages of UAVs—their maneuverability, flexibility, and wide field of view—to reduce the time required for road closures to acquire information; and through airborne lidar scanning and image enhancement processing, it solves the problem of missing accident scene investigation information under low-light conditions (such as dusk, dawn, and cloudy days).
[0071] 2. This invention incorporates information from a vehicle-road cooperative system, comprehensively utilizing road-traffic-environment-vehicle information to more fully reconstruct the occurrence process of traffic accidents, thereby clarifying accident liability; Attached Figure Description
[0072] Figure 1 This is a flowchart of the digital reconstruction method for road traffic accidents in the embodiment;
[0073] Figure 2 This is a structural diagram of the panoramic segmentation model of a traffic accident scene based on Transformer in the embodiment.
[0074] Figure 3 for Figure 2 Structure diagram of the feature-enhanced backbone;
[0075] Figure 4 for Figure 3 Structure diagram of the balanced pooling attention module;
[0076] Figure 5 A schematic diagram illustrating a virtual accident scenario incorporating extreme weather, generated using CARLA;
[0077] Figure 6 The above is a schematic diagram of the panoramic segmentation result of the target at the traffic accident scene in the embodiment. (a) is the original image of the accident scene, (b) is the segmentation result of Mask2Former, and (c) is the segmentation result of the panoramic segmentation model of the traffic accident scene based on Transformer.
[0078] Figures 7-9 These are schematic diagrams illustrating the changes in the vehicle's trajectory before, during, and after a collision. Detailed Implementation
[0079] The technical solution of the invention will be clearly and completely described below with reference to the accompanying drawings and embodiments. Obviously, the following embodiments are only some embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0080] Example
[0081] This invention provides a method for digital reconstruction of road traffic accidents, such as... Figure 1 As shown, it includes the following steps:
[0082] S1 uses an airborne camera and airborne lidar to conduct aerial photography and point cloud scanning of the accident site:
[0083] The drone, using the center of the accident site as its origin, flies around the site from different angles and altitudes, employing onboard cameras and LiDAR to conduct aerial photography and point cloud scanning. The aerial images and point cloud data are transmitted to the accident information processing platform in real time. The drone flies at an altitude of 5-50 meters, taking pictures and scanning the point cloud at angles of 30°-90°.
[0084] S2, preprocessing the aerial images obtained in S1:
[0085] When lighting conditions are poor during UAV aerial surveys (such as at night, early morning, or when the light is too strong or too weak), image enhancement is used to preprocess the aerial images received by the accident information processing platform in order to extract image features of the accident scene. The specific operation of image enhancement is as follows:
[0086] Gamma correction is used to improve the quality of aerial images. The correction formula is as follows:
[0087]
[0088] in, The aerial image; These are corrected aerial images; To correct parameters, when the aerial image is overexposed, Ensure that the brightness value of the output image is lower than that of the input image; this is especially important when the drone's aerial image has low exposure (dark scenes). This increases the brightness of the output image.
[0089] S3, Aerial Image Panoramic Segmentation: Using a Transformer-based panoramic segmentation model for traffic accident scenes, the types and quantities of traffic participants (such as accident vehicles, non-motorized vehicles, pedestrians, etc.) at the accident scene are extracted. The specific operation is as follows:
[0090] S31, Build a panoramic segmentation model of a traffic accident scene based on Transformer.
[0091] like Figure 2As shown, the Transformer-based panoramic segmentation model for traffic accident scenes in this invention is an improvement on the Mask2Former model: a balanced pooling attention (BPA) module is introduced to improve the original backbone of the Mask2Former model to form a feature-enhanced backbone, thereby enhancing the model's adaptability in panoramic segmentation of accident scenes.
[0092] In this invention, the lightweight Swin-Tiny network in the Swin Transformer architecture is used as the original backbone of the Mask2Former model.
[0093] The improved Transformer-based panoramic segmentation model for traffic accident scenes includes: a feature-enhanced backbone, a pixel decoder, and a Transformer decoder. Among them:
[0094] (1) The feature enhancement backbone is used to enhance the three-channel image of the input model ( The process yields four multi-scale feature maps. Specifically, such as Figure 3 As shown, the feature enhancement backbone includes a Swin-Tiny network and four Balanced Pooling Attention (BPA) modules. Wherein:
[0095] The Swin-Tiny network includes a Patch Partition module and four hierarchical stages (Stage 1 to Stage 4). First, the Patch Partition module processes the input image, converting the number of channels to 48 and reducing the width and height to match the input image. Then, through the concatenated Stages 1 through Stage 4, multi-scale feature maps are output respectively. Stages 1 through 4 contain stacked 2, 2, 6, and 2 Swin Transformer blocks, respectively. In Stage 1, the LinearEmbedding layer processes the input feature map, increasing the number of channels. The resulting output feature map is then processed by a SwingTransformer block for feature extraction, forming the output of Stage 1. ( Stages 2 through 4 use Patch Merging layers to downsample the input feature maps, reducing their dimensions to their original size. The number of channels is expanded to twice the input.
[0096] like Figure 4As shown, the Balanced Pooling Attention (BPA) module is used to process the feature maps respectively. Feature extraction is performed to obtain new multi-scale feature maps. The core objective is to enhance the ability to extract detailed features without losing overall features. The BPA module processes the input features... Using horizontal and vertical The directional pooling kernels are used for average pooling and max pooling to generate four one-dimensional feature maps. These feature information are processed through one-dimensional convolutions with a kernel size of 3 to achieve information exchange across channels, and then weighted and summed using learnable parameters to obtain the fused one-dimensional horizontal feature. and one-dimensional vertical features Then the two are connected along the spatial dimension to form ,pass Convolution (scaling ratio) Compress the number of channels, then... Split along spatial dimensions and ,pass Convolution and spatial broadcasting operations make it compatible with the initial input features Having the same number of channels and spatial dimensions, then applying them... Activation function generates attention weights and Finally, the original input features With weight , Element-wise multiplication yields the enhanced output features. .
[0097] While preserving the independence of features at each scale, further refine the feature maps. Multi-scale fusion is performed to obtain new multi-scale features. Specifically, That is ; for The upsampling results and Weighted summation; for The upsampling results and Weighted summation; for The upsampling results and The weighted sum.
[0098] The feature enhancement backbone outputs multi-scale feature maps. Feature maps obtained after Conv4d processing The length and width remain the same, but the number of channels becomes 256.
[0099] (2) The pixel decoder is used to process the input feature map Multi-scale fusion is performed to obtain feature maps. With feature map ,like Figure 2 As shown, for the feature map A 6-layer multi-scale deformable attention Transformer (MSDeformAttn) is applied for processing; the width, height, and number of channels of the output feature remain unchanged after processing. Then, the feature map is obtained through upsampling, weighted summation, and convolution. Feature map Compared with the upsampled feature map Perform a weighted sum, then... Convolution and Convolution processing yields feature maps .
[0100] (3) The Transformer decoder has 9 stacked Transformer decoder layers, initialized with a set of learnable query features. Each query has a corresponding target mask and category. The ultimate goal is to predict the target mask and category for each query. Each of the three decoder layers receives feature maps. As pixel embedding (specifically, first layer receiving) Second layer reception Third layer reception Fourth layer reception Other layers receive feature maps in the same manner, and receive a set of learnable query vectors (queries features) output from the previous decoder layer. Each query is processed using MaskedCross-Attention, Self-Attention, and FFN to output an updated query vector. Each query has a corresponding mask image and a corresponding category.
[0101] The final query vector output from the Transformer decoder is fed into two fully connected layers. One layer performs classification to obtain the corresponding class and class probability, while the other layer maps it to a mask embedding vector and the feature map. Perform a dot product to obtain the mask image.
[0102] The model's results can be visualized as an overlay of all mask images and their corresponding categories.
[0103] S32 trains the Transformer-based panoramic segmentation model of traffic accident scenes built in S31.
[0104] A boundary-aware loss is introduced based on the original loss function to construct a weighted loss function. This weighted loss function is a weighted sum of the classification loss, masking loss, and boundary-aware loss.
[0105] ;
[0106] in: For classification loss, corresponding weights ; For mask loss; The introduced boundary-aware loss corresponds to the weights. The value range is 0.2-0.5.
[0107] The formula for the boundary-aware loss function is:
[0108] ;
[0109] in:
[0110] ;
[0111] ;
[0112] ;
[0113] ;
[0114] Both dilation and erosion operations use structural elements (typically 3×3 convolution kernels).
[0115] It is the continuous probability mask of the model's output prediction. Refers to the actual mask;
[0116] It is a very small positive number to prevent the denominator from being 0 and to stabilize the training values.
[0117] The boundary-aware loss is applied to the mask contour to assist in boundary supervision. By emphasizing the consistency between the predicted mask and the real mask in the boundary region, it guides the model to segment the object contour more accurately.
[0118] Based on the above weighted loss function , adopt as Figure 5 The CARLA or UE5 example shown generates virtual accident scene data containing extreme weather (such as nighttime collisions and snow rollovers), and injects simulated rain, snow, occlusion, and fog noise into the input. It uses adversarial training to optimize the panoramic segmentation performance of the Transformer-based traffic accident scene panoramic segmentation model in complex environments.
[0119] The panoramic segmentation results of traffic accident scenes based on the Transformer model of this invention are compared with those of the Mask2Former model. The results are as follows: Figure 6 As shown, the panoramic segmentation model of the traffic accident scene of the present invention can more comprehensively identify traffic participants, especially accident vehicles.
[0120] S33 uses a trained Transformer-based panoramic segmentation model for traffic accident scenes to perform panoramic segmentation on the preprocessed aerial images obtained in S2, thereby extracting the masks, types, and quantities of each type of traffic participants (such as accident vehicles, non-motorized vehicles, pedestrians, etc.) at the accident scene.
[0121] S4, Spatial location calculation of traffic participants:
[0122] The masks corresponding to traffic participants obtained in S3 are projected onto the three-dimensional coordinate system of the point cloud data obtained in S1 based on the intrinsic and extrinsic parameters of the airborne camera. This helps to determine the set of points contained in each traffic participant, thereby more accurately calculating the relative spatial position of each traffic participant at the accident scene.
[0123] S5, obtain comprehensive vehicle and road information:
[0124] Based on the communication status of the vehicle-road cooperative system, information such as road traffic flow, traffic signal control, and road channelization design is determined through the roadside unit (RSU).
[0125] If all the vehicles involved in the accident are intelligent vehicles equipped with an On Board Unit (OBU), the OBU will directly provide the vehicle's operating status information (such as position, speed, acceleration, heading angle, braking, etc.) for 10 to 20 seconds before the accident. Professionals can directly use the operating status information to determine the cause of the accident and assign responsibility.
[0126] If the vehicles involved in the accident are all conventional vehicles and are not equipped with onboard units, proceed to step S61;
[0127] If one of the vehicles involved in the accident is an intelligent vehicle and the other is a traditional vehicle, the vehicle's operating status information for 10 to 20 seconds before the traffic accident is obtained through the on-board unit, and then step S62 is executed.
[0128] S6, Traffic Accident Simulation Analysis:
[0129] S61, based on the type of vehicle involved in the accident and the road traffic flow, traffic signal control, and road channelization design information obtained from S5, establishes a mathematical model of the accident collision in the traffic accident collision simulation software by setting vehicle type, road, weather, and traffic signal information, such as... Figures 7-9 As shown, the process of simulating a traffic accident is optimized by continuously adjusting the speed and heading angle of the vehicles involved in the accident in the model, until the absolute error between the simulated distance after the collision and the relative spatial distance obtained by S4 is calculated. If the score is 5%, the road traffic accident reconstruction process is considered to meet the requirements and the road traffic accident reconstruction is considered complete.
[0130] Professionals can use simulations to optimize the speed and heading angle of the two vehicles involved in the accident, thereby identifying the cause of the accident and determining liability.
[0131] S62, based on the type of the accident vehicle, road traffic flow, traffic signal control, road channelization design information obtained from S5, and the operating status information of the intelligent vehicle, establishes a mathematical model of accident collision in the traffic accident collision simulation software by setting vehicle type, road, weather, and traffic signal information. It simulates the process of a traffic accident and continuously adjusts the speed and heading angle of the traditional vehicle in the model to optimize the simulation until the absolute error between the simulated distance after the accident vehicle collision and the relative spatial distance obtained from S4 is found. If the score is 5%, the road traffic accident reconstruction process is considered to meet the requirements and the road traffic accident reconstruction is considered complete.
[0132] Professionals can determine the cause of the accident and assign responsibility based on the operating status information of the intelligent vehicle and the speed and heading angle of the traditional vehicle obtained through simulation optimization.
Claims
1. A method for digital reconstruction of road traffic accidents, characterized in that, Includes the following steps: S1, to take aerial photos and point cloud scans of the accident site to obtain aerial images and point cloud data; S2, preprocessing the aerial images obtained in S1; S3 uses a Transformer-based panoramic segmentation model for traffic accident scenes to perform panoramic segmentation on the aerial images obtained in S2, extracting the masks, types, and number of traffic participants at the accident scene. S4 projects the mask of traffic participants obtained in S3 onto the three-dimensional coordinate system of the point cloud data obtained in S1 to help determine the point set contained in each traffic participant, thereby calculating the relative spatial position of each traffic participant at the accident scene. S5, obtains comprehensive vehicle-road information: Based on the communication status of the vehicle-road cooperative system, information on road traffic flow, traffic signal control, and road channelization design is determined through roadside units; If all the vehicles involved in the accident are intelligent vehicles, the operating status information of the vehicles involved in the accident will be obtained directly from the on-board unit installed on the vehicle to determine the cause of the accident and to assign responsibility. If all the vehicles involved in the accident are traditional vehicles and are not equipped with an on-board unit, proceed to step S61; if one of the vehicles involved in the accident is an intelligent vehicle and the other is a traditional vehicle, obtain the operating status information of the intelligent vehicle through the on-board unit and proceed to step S62. S6, Traffic Accident Simulation Analysis: S61, in the traffic accident collision simulation software, establish an accident collision mathematical model, and continuously adjust the speed and heading angle of the accident vehicles in the model to perform simulation optimization until the absolute error between the simulated distance after the accident vehicle collision and the relative spatial distance obtained in S4 is found. 5%, completing the reconstruction of road traffic accidents; S62, in traffic accident collision simulation software, establishes a mathematical model of accident collision to simulate the process of a traffic accident. By continuously adjusting the speed and heading angle of the traditional vehicles in the model, simulation optimization is performed until the absolute error between the simulated distance after the collision and the relative spatial distance obtained in S4 is found. 5%, completing the reconstruction of road traffic accidents.
2. The method for digital reconstruction of road traffic accidents according to claim 1, characterized in that, The specific operation of step S2 is as follows: Gamma correction is used to improve the quality of aerial images obtained from S1. The correction formula is as follows: in, Aerial images; These are corrected aerial images; To correct parameters, if the aerial image has a high exposure, If the aerial image has low exposure, .
3. The method for digital reconstruction of road traffic accidents according to claim 2, characterized in that, The specific operation of step S3 is as follows: S31, Build a panoramic segmentation model of a traffic accident scene based on Transformer: The Transformer-based panoramic segmentation model for traffic accident scenes includes a feature enhancement backbone, a pixel decoder, and a Transformer decoder, which are used to process the three-channel images of the input model. The feature enhancement backbone is obtained by introducing a balanced pooling attention module to improve the original backbone of the Mask2Former model. S32, training the panoramic segmentation model of the traffic accident scene built in S31: A weighted loss function is constructed. Based on the weighted loss function, virtual accident scene data containing extreme weather is generated using an open-source autonomous driving simulation platform. Noise simulating rain, snow, obstruction, and fog is injected into the input. The panoramic segmentation performance of the traffic accident scene panoramic segmentation model in complex environments is optimized by using adversarial training to complete the training. S33, use the trained panoramic segmentation model of the traffic accident scene to perform panoramic segmentation on the preprocessed aerial image obtained in S2, and then extract the mask, type and number of traffic participants at the accident scene, including accident vehicles, non-motorized vehicles and pedestrians.
4. The method for digital reconstruction of road traffic accidents according to claim 3, characterized in that, In step S31: The process of processing the three-channel image of the input model is as follows: (1) The feature enhancement backbone processes the input three-channel image to obtain four multi-scale feature maps. ; The feature enhancement backbone includes a Win-Tiny network and four balanced pooling attention modules, wherein: The Swin-Tiny network includes a Patch Partition module and four hierarchical stages (Stage 1 to Stage 4). First, the Patch Partition module processes the input image, converting the number of channels to 48 and reducing the width and height to match the input image. Then, feature maps are output through the concatenated Stages 1 through Stage 4. ; The balanced pooling attention module is used to process the feature maps respectively. Feature extraction is performed to obtain feature maps. ; Further analysis of feature maps Multi-scale fusion is performed to obtain new multi-scale features. Specifically: for ; for The upsampling results and Weighted summation; for The upsampling results and Weighted summation; for The upsampling results and Weighted summation; (2) Feature map output by the feature enhancement backbone Feature maps obtained after Conv4d processing The length and width remain the same, but the number of channels becomes 256; (3) The pixel decoder processes the input feature map Multi-scale fusion is performed to obtain feature maps. With feature map : For feature maps A 6-layer multi-scale deformable attention Transformer is applied for processing; the width, height, and number of channels of the output feature remain unchanged after processing. Then, the feature map is obtained through upsampling, weighted summation, and convolution. Feature map Compared with the upsampled feature map Perform a weighted sum, then... Convolution and Convolution processing yields feature maps ; (4) The Transformer decoder has 9 stacked Transformer decoder layers, initialized with a set of learnable query vectors, including Each query receives feature maps from every three decoder layers. As a pixel embedding, it receives the query vector output from the previous decoder layer, processes it through Masked Cross-Attention, Self-Attention, and FFN, and outputs the updated query vector; each query has a corresponding mask image and corresponding category; (5) The final query vector output by the Transformer decoder is fed into two fully connected layers. One layer is used for classification to obtain the corresponding category and category probability. The other layer is mapped to the mask embedding vector and the feature map. Perform a dot product to obtain the mask image.
5. The method for digital reconstruction of road traffic accidents according to claim 4, characterized in that, In step S31: Stages 1 through 4 contain stacked 2, 2, 6, and 2 Swing Transformer blocks, respectively. In Stage 1, the Linear Embedding layer processes the input feature map, increasing the number of channels. The output feature map is obtained, and then the SwingTransformer block is used for feature extraction to form the output of Stage 1. Stages 2 through 4 use Patch Merging layers to downsample the input feature maps, reducing their width and height to their original values. The number of channels is expanded to twice the input.
6. The method for digital reconstruction of road traffic accidents according to claim 4, characterized in that, In step S31: The balanced pooling attention module processes the input features. Average pooling and max pooling are performed using horizontal and vertical pooling kernels respectively, generating four one-dimensional feature maps. Then, a fused one-dimensional horizontal feature is obtained by performing a one-dimensional convolution with a kernel size of 3 and weighted summation. and one-dimensional vertical features Then the two are connected along the spatial dimension to form ,pass Convolution compresses the number of channels, then splits them along the spatial dimension into and ,pass Convolution and spatial broadcasting operations respectively connect the two to the initial input features. Having the same number of channels and spatial dimensions, then applying them... Activation function generates attention weights and Finally, the original input features With weight , Element-wise multiplication yields the enhanced output features. .
7. The method for digital reconstruction of road traffic accidents according to claim 6, characterized in that, In step S32: The formula for the weighted loss function is as follows: ; in: For classification loss, the corresponding weights ; For mask loss; For boundary-aware loss, the corresponding weights The value range is 0.2-0.5; The formula for the boundary-aware loss function is: ; in: ; ; ; ; expansion With corrosion All operations use structural elements; It is the continuous probability mask of the model's output prediction. Refers to the actual mask; It should be a small positive number to prevent the denominator from being 0.
8. The method for digital reconstruction of road traffic accidents according to claim 1, characterized in that, In step S5: The vehicle operation status information refers to the vehicle's operation status information 10 to 20 seconds before the traffic accident. The operational status information includes position, speed, acceleration, heading angle, and braking.
9. The method for digital reconstruction of road traffic accidents according to claim 8, characterized in that, The specific operation of step S61 is as follows: Based on the type of the accident vehicle and the road traffic flow, traffic signal control, and road channelization design information obtained from S5, a mathematical model of the accident collision is established by setting vehicle type, road, weather, and traffic signal information in the traffic accident collision simulation software.
10. The method for digital reconstruction of road traffic accidents according to claim 8, characterized in that, The specific operation of step S62 is as follows: Based on the type of vehicle involved in the accident, the road traffic flow obtained from S5, traffic signal control, road channelization design information, and the operating status information of the intelligent vehicle, a mathematical model of the accident collision is established in the traffic accident collision simulation software by setting vehicle type, road, weather, and traffic signal information.
Citation Information
Patent Citations
Traffic accident scene three-dimensional reconstruction system and method based on unmanned aerial vehicle aircraft aerial photography
CN107194989A
Laser point cloud-based road traffic accident auxiliary processing method and system
CN112446958A
Road traffic accident intelligent reproduction method based on automatic Pc-Crash simulation and MLP algorithm
CN120124451A
Road traffic accident three-dimensional reconstruction system and method
CN120495503A
