Traffic accident responsibility determination method and device, equipment and medium

Abstract

The application relates to the technical field of image processing, and discloses a traffic accident responsibility determination method, device, equipment and medium. The method comprises the following steps: acquiring an effective accident scene image and an accident auxiliary feature; performing accident feature extraction on the effective accident scene image to obtain an accident scene feature; performing simulation reconstruction on a traffic accident scene based on the accident scene feature and the accident auxiliary feature to obtain an accident scene simulation model; and generating a traffic accident responsibility determination result for the traffic accident scene based on the accident scene simulation model and a preset traffic accident responsibility determination strategy. The embodiment of the application can improve the efficiency and accuracy of traffic accident responsibility determination.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method, apparatus, equipment and medium for determining liability in traffic accidents. Background Technology

[0002] Vehicle-to-VRU (pedestrian / non-motorized vehicle) collisions are frequent in road traffic operations. On-site handling and liability determination for such accidents are crucial aspects of traffic management and insurance claims. Current methods for determining liability in vehicle-to-VRU collision accidents suffer from fragmented and disorganized on-site data, lacking a systematic integration solution. Furthermore, traditional 3D reconstruction methods for accident scenes have low processing accuracy for unstructured images, lack standardized reference point selection, and struggle to accurately reconstruct the actual accident scene. Simultaneously, traditional liability determination methods rely on on-site investigations and subjective judgments by staff, resulting in low efficiency and susceptibility to human error, leading to a lack of objectivity and scientific rigor in the determinations. Moreover, existing accident handling methods are mostly simple combinations of conventional technologies, lacking targeted innovative designs in key areas such as image preprocessing, viewpoint coverage assessment, and high-precision map fusion, thus failing to meet the current demands for efficient and intelligent traffic management. Summary of the Invention

[0003] The purpose of this application is to provide a method, device, equipment, and medium for determining liability in traffic accidents, aiming to improve the efficiency and accuracy of determining liability in traffic accidents.

[0004] This application provides a method for determining liability in traffic accidents, including: Acquire valid accident scene images and auxiliary accident features; the valid accident scene images include valid image features of the traffic accident scene, and the auxiliary accident features include the location features of the traffic accident scene, the weather features, the road condition features, and / or the preliminary accident liability determination features; Accident features are extracted from the valid accident scene images to obtain accident scene features; Based on the accident scene features and the accident auxiliary features, the traffic accident scene is simulated and reconstructed to obtain an accident scene simulation model; Based on the accident scene simulation model and the preset traffic accident liability determination strategy, a traffic accident liability determination result is generated for the traffic accident scene.

[0005] In some embodiments, prior to acquiring valid accident scene images and accident auxiliary features, the method further includes: Acquire multiple sets of accident scene images taken from multiple angles at the traffic accident scene; The accident scene images are preprocessed to obtain preprocessed scene images; The preprocessed on-site images are subjected to coordinate transformation to map each preprocessed on-site image to the same coordinate system, thereby obtaining coordinate transformed on-site images; The validity of the coordinate-transformed on-site image is evaluated, and information is supplemented based on the validity evaluation results to obtain the valid accident scene image.

[0006] In some embodiments, the step of evaluating the validity of the coordinate transformation scene image and supplementing the information of the coordinate transformation scene image based on the validity evaluation result includes: Based on a preset evaluation strategy, the sharpness, viewpoint coverage, and information integrity of the coordinate transformation scene image are evaluated to obtain the effectiveness evaluation result. When the effectiveness evaluation result indicates that the evaluation fails, the coordinate transformation scene image is replaced and / or information is supplemented until the effectiveness evaluation result indicates that the evaluation passes, thus obtaining the effective accident scene image.

[0007] In some embodiments, the step of extracting accident features from the valid accident scene images includes: The key areas of the accident are obtained by performing key area detection on the effective accident scene images; Extract the regional contour features and key point features of the critical area of ​​the accident; Based on the region contour features and the key point features, semantic label features matching the key accident region are determined; the semantic label features are used to describe the type information, damage degree information and / or spatial location information of the target in the key accident region. The accident scene features are obtained by integrating the region contour features, the key point features, and the semantic label features.

[0008] In some embodiments, prior to simulating and reconstructing the traffic accident scene, the method further includes: Reference images of the traffic accident scene are captured from at least three shooting angles, and the positional deviation of the accident scene features is corrected to obtain the corrected accident scene features. The accident auxiliary features are matched with the map features corresponding to the traffic accident scene, and the accident auxiliary features are updated based on the feature matching results until the similarity between the updated accident auxiliary features and the map features reaches a preset similarity threshold. The corrected accident scene features and the updated accident auxiliary features are added to the map features corresponding to the traffic accident scene.

[0009] In some embodiments, the step of simulating and reconstructing the traffic accident scene based on the accident scene features and the accident auxiliary features includes: The accident scene features and the accident auxiliary features are converted into corresponding three-dimensional point cloud models; The three-dimensional point cloud model is divided into triangular meshes, and key regions in the three-dimensional point cloud model are identified. The key regions are then subjected to mesh refinement processing to generate an adaptive mesh model. The three-dimensional model of the vehicle or vulnerable road user is fused with the adaptive mesh model in a unified coordinate system to generate a static accident scene model. Physical material properties are assigned to each object in the static accident scene model, and the contact and collision relationships between objects are defined to obtain the accident scene simulation model.

[0010] In some embodiments, generating a traffic accident liability determination result for the traffic accident scene based on the accident scene simulation model and a preset traffic accident liability determination strategy includes: Based on the accident scene simulation model, the traffic accident scene is reverse simulated to deduce the initial motion simulation parameters of the traffic accident scene from the final state. Based on the initial motion simulation parameters, a forward iterative simulation of the traffic accident scene is performed. During the forward iterative simulation, the input parameters are dynamically adjusted so that the error between the forward simulation result model and the accident scene simulation model is less than a preset error threshold. Based on the positive simulation result model and the traffic accident liability determination strategy, a traffic accident liability determination result is generated for the traffic accident scene.

[0011] This application embodiment also provides a traffic accident liability determination device, including: The first module is used to acquire valid accident scene images and accident auxiliary features; the valid accident scene images include valid image features of the traffic accident scene, and the accident auxiliary features include the on-site location features, on-site weather features, on-site road condition features and / or preliminary accident liability determination features of the traffic accident scene; The second module is used to extract accident features from the valid accident scene images to obtain accident scene features; The third module is used to simulate and reconstruct the traffic accident scene based on the accident scene features and the accident auxiliary features to obtain an accident scene simulation model. The fourth module is used to generate a traffic accident liability determination result for the traffic accident scene based on the accident scene simulation model and the preset traffic accident liability determination strategy.

[0012] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method for determining liability in traffic accidents.

[0013] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for determining liability in traffic accidents.

[0014] The beneficial effects of this application are as follows: By extracting accident features from valid accident scene images and performing simulation reconstruction based on the extracted accident scene features and auxiliary features, a more accurate and objective accident scene simulation model can be constructed, and the simulated traffic accident scene is more realistic and credible. Generating liability determination results based on the accident scene simulation model and a preset traffic accident liability determination strategy allows for objective analysis of the accident process based on the simulation model and provides liability judgments according to the preset strategy, thereby improving the efficiency and fairness of liability determination and reducing the uncertainty caused by human intervention. Attached Figure Description

[0015] Figure 1 This is a flowchart of the traffic accident liability determination method provided in the embodiments of this application.

[0016] Figure 2 This is a flowchart of a method for extracting accident features from valid accident scene images provided in an embodiment of this application.

[0017] Figure 3 This is a flowchart of a method for simulating and reconstructing traffic accident scenes provided in an embodiment of this application.

[0018] Figure 4 This is a flowchart of a method for generating traffic accident liability determination results at a traffic accident scene, provided in an embodiment of this application.

[0019] Figure 5 This is a schematic diagram of the traffic accident liability determination device provided in the embodiments of this application.

[0020] Figure 6 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and drawings are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application. Furthermore, the information, data, and signals involved in the embodiments of this application are all authorized by relevant parties or have been fully authorized by all parties, and the collection, use, and processing of related data comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0024] In traditional road traffic accident handling processes, accident scene data is fragmented and lacks a systematic integration mechanism, resulting in significantly insufficient processing accuracy of unstructured image data in the 3D reconstruction stage. Furthermore, the lack of unified technical standards in the benchmark selection process makes it difficult to accurately reconstruct the spatial relationships of the accident scene. Moreover, the liability determination process relies excessively on subjective experience and judgment from manual on-site investigations, systematically restricting the objectivity and scientific rigor of the determination conclusions. Consequently, the technical solutions for key aspects such as image preprocessing, viewpoint coverage assessment, and high-precision map fusion are merely simple combinations of conventional methods, failing to develop adaptive and innovative designs for accident liability determination scenarios. This, in turn, has a continuous negative impact on core performance indicators such as liability determination efficiency, scene reconstruction accuracy, and decision-making objectivity.

[0025] For example, in a vehicle-pedestrian collision at a major urban intersection, multiple accident images taken by different mobile devices remain at the scene. Due to random shooting angles and complex ambient lighting conditions, these images exhibit significant unstructured characteristics. Location features include worn road markings, meteorological features include road surface reflections caused by precipitation, and road conditions are characterized by a slippery surface. Furthermore, while traditional image processing tools are used for benchmark point calibration, the lack of unified coordinate mapping rules leads to systematic biases in the spatial coordinate calculations of the vehicle collision points. Consequently, the contour features and key point features of critical accident areas cannot be fully extracted, the correlation between semantic label features and the degree of damage to the target is broken, ultimately causing the construction of the static accident scene model to be interrupted due to missing information, and forcing a delay in the generation of liability determination.

[0026] If the above problems are not resolved, the spatial topology of the accident scene simulation model will continue to be distorted, making it impossible to accurately reconstruct the initial motion parameters during the reverse engineering process. In particular, the error between the forward iterative simulation and the actual state of the accident scene will exceed the acceptable range, causing a logical conflict between the liability determination results and physical laws. Furthermore, the delay in the liability determination process will exacerbate the traffic congestion effect at the accident scene, and the fragmented state of the on-site data will lead to a continuous deterioration in the integrity of the basis for liability determination. As a result, the intelligent evolution of the traffic management system will be fundamentally hindered, the credibility of the accident handling conclusions will be systematically weakened due to technical defects, and ultimately affect the overall effectiveness of the road traffic safety management system.

[0027] Based on this, embodiments of this application provide a method, apparatus, equipment, and medium for determining liability in traffic accidents. By integrating effective accident scene images and auxiliary accident features for feature extraction and simulation reconstruction, objective liability determination can be achieved, thereby improving the efficiency and accuracy of determining liability in traffic accidents.

[0028] Figure 1 This is a flowchart of the traffic accident liability determination method provided in the embodiments of this application. (See attached document.) Figure 1 In one embodiment, the method includes, but is not limited to, steps S101 to S104.

[0029] Step S101: Obtain valid accident scene images and accident auxiliary features.

[0030] Valid accident scene images contain the essential visual features of a traffic accident scene. In essence, valid accident scene images refer to image data that, after screening and processing, clearly and completely reflects the conditions of a traffic accident scene. This image data typically includes the relative positions of the vehicles involved, the extent of damage, road marks, traffic signs, and other visual information related to the accident, forming the basis for accident analysis and reconstruction.

[0031] Accident auxiliary features include the location characteristics of the traffic accident scene, meteorological characteristics, road condition characteristics, and / or preliminary accident liability determination features. In essence, accident auxiliary features refer to data, in addition to image information, used to supplement and describe the environment and background of the traffic accident scene. These features may include the specific geographic coordinates of the accident, the time, weather conditions (e.g., sunny, rainy, snowy, foggy), road conditions (e.g., dry, slippery, icy), traffic flow, and preliminary investigation records or statements from the parties involved. These auxiliary features provide crucial environmental contextual information for subsequent simulation reconstruction and liability determination.

[0032] There are several methods for acquiring effective accident scene images and auxiliary accident features. For example, on-site investigators can use standard cameras or smartphones to take multi-angle photos of the accident scene to obtain raw images. Simultaneously, investigators can manually record auxiliary information such as the location, time, weather conditions, and road conditions of the accident, and make a preliminary determination of responsibility. These images and auxiliary information are then manually screened and deemed valid data. Another method is to use multiple fixed surveillance cameras deployed along the roadside to automatically capture static images of the accident scene after it occurs, combining this with environmental sensors integrated into the monitoring system to obtain meteorological and road condition data. While these single-source images and auxiliary features can provide basic information, they may be incomplete or have limited perspectives in certain complex scenarios.

[0033] Step S102: Extract accident features from valid accident scene images to obtain accident scene features.

[0034] Accident scene features refer to structured information extracted from valid accident scene images to describe the core elements of the accident. These features typically include the type, posture, damaged parts and extent of the vehicles involved, the location and posture of pedestrians or non-motorized vehicles, and the distribution of debris (such as fragments and scattered objects) left at the scene. These features are key inputs for constructing accident scene simulation models.

[0035] In extracting accident features from valid accident scene images, manual identification or basic algorithm processing can be employed. Specifically, experienced accident analysts can directly observe valid accident scene images, manually delineate key objects such as vehicles, pedestrians, and obstacles, and mark their approximate locations and damage levels, thereby forming accident scene features. Alternatively, basic image processing algorithms, such as edge detection and color segmentation, can be used to identify salient regions in the image to extract visual features such as vehicle outlines and collision points. While these extracted features can initially outline the accident scene, they may have limitations in terms of detail accuracy and semantic understanding.

[0036] Step S103: Based on the accident scene features and accident auxiliary features, the traffic accident scene is simulated and reconstructed to obtain an accident scene simulation model.

[0037] An accident scene simulation model refers to a virtual three-dimensional or multi-dimensional model that uses computer simulation technology to transform the various characteristic data of a traffic accident scene into a virtual model. This model can present the static scene of the accident in a highly realistic manner and can be further used to simulate the dynamic changes during the accident process. This model provides a visual and quantifiable platform for objective accident analysis.

[0038] In simulating and reconstructing traffic accident scenes based on accident scene features and auxiliary accident features, a simplified model can be constructed. For example, extracted accident scene features (such as vehicle positions and damage points) can be manually input into a 2D drawing software, and combined with auxiliary accident features (such as road type and lane lines) to draw a planar layout of the accident scene. Based on this, basic 3D modeling tools can be used to place simplified 3D models of vehicles and pedestrians onto this planar layout, and their approximate height and posture can be adjusted according to the auxiliary features, thus forming a static accident scene simulation model. This reconstruction method can provide an intuitive overview of the scene, but its physical realism and dynamic simulation capabilities may be limited.

[0039] Step S104: Based on the accident scene simulation model and the preset traffic accident liability determination strategy, generate the traffic accident liability determination result for the traffic accident scene.

[0040] A traffic accident liability determination strategy refers to a set of pre-defined rules, standards, and algorithms used to determine the attribution of responsibility in traffic accidents. This strategy is typically built upon traffic regulations, industry practices, historical case data, and expert experience, aiming to provide a fair and objective basis for determining liability in different types of traffic accidents.

[0041] The determination of liability in a traffic accident refers to the conclusion drawn regarding the proportion or type of responsibility of each party involved, based on accident scene simulation models and traffic accident liability determination strategies. This result serves as a crucial basis for traffic management departments to impose penalties, insurance companies to process claims, and judicial institutions to make rulings.

[0042] In generating traffic accident liability determination results based on accident scene simulation models and pre-set traffic accident liability determination strategies, rule matching can be used. Specifically, the generated accident scene simulation model can be used as input, allowing human experts to intuitively judge the key elements of the accident based on the model and match them against a pre-set traffic accident liability determination strategy (e.g., a liability division table based on traffic regulations) to arrive at a preliminary liability determination result. Alternatively, a simple expert system can be developed with a built-in series of rule-based liability determination logic. When key parameters (such as relative vehicle positions, collision points, and traffic signal status) are input from the accident scene simulation model, rule matching is automatically executed, and the corresponding liability determination result is output. This method can provide a preliminary liability determination conclusion, but its ability to analyze the dynamic process of accidents may be insufficient, and its adaptability to complex and changing scenarios needs improvement.

[0043] The following example will provide a more detailed explanation of the above technical solution: Suppose that on a rainy day, user A's vehicle collides with a pedestrian at location A. Traditional methods of determining liability may face problems such as blurred on-site data due to rain and subjective judgment by investigators, leading to low efficiency and unobjective results.

[0044] The method provided in this embodiment can effectively solve these problems. First, after the accident, on-site investigators use portable equipment to take multiple photos of the accident scene and record the specific time of the accident, the GPS coordinates of location A, the weather conditions (light rain), the road surface conditions (slippery), and the initial observed positions of vehicles and pedestrians. These images and auxiliary information are initially screened to ensure their clarity and completeness, thereby obtaining effective accident scene images and accident auxiliary features. For example, the screened images clearly show the collision points of the vehicles and the positions of the pedestrians, and the auxiliary features accurately record the key environmental factor of the slippery road surface.

[0045] Next, accident features are extracted from these valid accident scene images. Image processing technology automatically identifies key elements such as vehicles, pedestrians, collision points, and brake marks. For example, it identifies a significant dent in the front of the vehicle, a pedestrian approximately 2 meters in front of the vehicle, and about 5 meters of brake marks on the road. This identified information is integrated into accident scene features, which describe in detail the static layout at the time of the accident.

[0046] Subsequently, based on the accident scene characteristics and auxiliary accident features, a simulation reconstruction of the traffic accident scene was performed. The extracted scene features, including vehicles, pedestrians, and roads, were combined with the geographical information of location A and auxiliary features such as slippery road surfaces to construct a three-dimensional accident scene simulation model. This model not only accurately reproduces the relative positions, postures, and damage of vehicles and pedestrians, but also considers the potential impact of slippery road surfaces on vehicle braking distance. This simulation model provides a highly visualized and near-realistic accident scene environment.

[0047] Finally, based on the accident scene simulation model and the preset traffic accident liability determination strategy, a traffic accident liability determination result is generated for this specific accident scene. Using the simulation model as input, and combining it with the built-in liability determination strategy (which may include rules such as "slow down on wet roads" and "pedestrians should use crosswalks when crossing the road"), the accident liability is analyzed. For example, by analyzing the relationship between the vehicle's braking distance and the wet road surface in the simulation model, as well as the pedestrian's position when crossing the road, the respective liability proportions of the vehicle and the pedestrian are finally determined. This liability determination result is based on objective data and simulation analysis, avoiding the subjective bias that may exist in traditional methods.

[0048] Based on the above examples, the technical concept of this embodiment demonstrates a significant technical contribution. Compared to the problems of scattered and disorganized on-site data and the lack of a systematic integration scheme in existing technologies, this embodiment achieves comprehensive and structured collection of accident scene information by systematically acquiring effective accident scene images and accident auxiliary features. For example, in rainy accidents, traditional methods may result in incomplete data due to rain blurring, while this embodiment ensures a solid data foundation for subsequent analysis by evaluating the effectiveness of images and integrating auxiliary features.

[0049] Furthermore, addressing the issues of low accuracy in processing unstructured images and the lack of a unified standard for selecting reference points in existing 3D accident scene reconstruction methods, this embodiment extracts accident features from valid accident scene images and performs simulation reconstruction based on these features and auxiliary features, thereby constructing a more accurate and objective accident scene simulation model. In the example above, by extracting scene features such as vehicle damage, pedestrian positions, and brake marks, and combining them with auxiliary features such as slippery road surfaces for simulation, the resulting model is more realistic and reliable than traditional reconstruction results that rely on manual visual inspection or simple measurement.

[0050] Furthermore, this embodiment generates liability determination results based on an accident scene simulation model and a preset traffic accident liability determination strategy. This effectively solves the drawbacks of traditional liability determination methods, which rely on on-site investigations and subjective judgments by staff, resulting in low efficiency and susceptibility to human factors that lead to a lack of objectivity and scientific rigor in the determination results. In the aforementioned example of a rainy day accident, the system can objectively analyze the accident process based on the simulation model and provide a liability judgment according to the preset strategy, thereby improving the efficiency and fairness of liability determination and reducing the uncertainty caused by human intervention.

[0051] In summary, this embodiment provides a more systematic, intelligent, and objective method for determining liability in traffic accidents, overcoming the limitations of existing technologies in data integration, scene reconstruction, and liability determination, and providing strong technical support for traffic management and insurance claims.

[0052] In some embodiments, before acquiring valid accident scene images and accident auxiliary features, the method further includes: acquiring multiple sets of accident scene images taken from multiple angles at the traffic accident scene; preprocessing the accident scene images to obtain preprocessed scene images; performing coordinate transformation processing on the preprocessed scene images to map each preprocessed scene image to the same coordinate system to obtain coordinate transformed scene images; evaluating the validity of the coordinate transformed scene images, and supplementing the coordinate transformed scene images with information based on the validity evaluation results to obtain valid accident scene images.

[0053] Preprocessing accident scene images can involve using image filtering algorithms (such as Gaussian filtering and median filtering) for noise reduction, using camera calibration parameters to correct image distortion, or adjusting image brightness and contrast through methods such as histogram equalization and gamma correction.

[0054] Coordinate transformation processing of preprocessed scene images can be performed by calculating homography matrices or fundamental matrices using common-view feature points in the images (such as road markings, building corners, and specific parts of vehicles), projecting images from different perspectives onto a unified two-dimensional plane coordinate system or three-dimensional world coordinate system; or by using Simultaneous Localization and Mapping (SLAM) technology based on feature matching to construct the three-dimensional structure of the scene and determine the relative poses of each image.

[0055] The effectiveness of coordinate transformation field images is evaluated, and information completion is performed on the coordinate transformation field images based on the effectiveness evaluation results. This can be based on a preset evaluation strategy, assessing image sharpness (e.g., by calculating image gradients and Laplacian operator responses), viewpoint coverage (e.g., by analyzing overlapping areas after image stitching), and information completeness (e.g., by determining whether key objects are missing through object detection or semantic segmentation). When the evaluation results indicate that the image has quality problems or missing information, measures can be taken such as replacing low-quality images, using images from other angles for information interpolation or fusion, or even using image generation techniques (e.g., image inpainting based on deep learning) to complete the missing areas until the preset effectiveness criteria are met.

[0056] This application's solution systematically acquires multi-angle accident scene images and performs preprocessing, coordinate transformation, validity assessment, and information completion, constructing a complete processing flow from raw images to high-quality, valid accident scene images. First, multi-angle shooting ensures the comprehensiveness and redundancy of accident scene information, providing a rich data source for subsequent processing. Second, the preprocessing step effectively improves image quality and reduces interference. Next, coordinate transformation resolves the spatial inconsistency problem of images from different perspectives, enabling all images to be fused and analyzed within a unified framework. Finally, the validity assessment and information completion mechanisms serve as key quality control steps, ensuring that the final output valid accident scene images are high-quality, highly complete, and spatially consistent, thus providing reliable input data for subsequent accident feature extraction and simulation reconstruction. This systematic image preparation process significantly improves the accuracy of subsequent accident scene feature extraction and lays a solid foundation for building high-precision accident scene simulation models, thereby enhancing the reliability of traffic accident liability determination results.

[0057] The following is a concrete example to illustrate this. After a traffic accident, on-site investigators use smartphones to take multiple sets of images of the accident scene from the four sides, front, rear, and from high vantage points (such as standing at a height or using a selfie stick) of the vehicle involved. These images are first automatically preprocessed by image processing software, including removing blur caused by hand shake or insufficient lighting, correcting wide-angle distortion of the phone camera, and adjusting image brightness to maintain consistency under different lighting conditions. Subsequently, the system uses lane lines, lampposts, building edges, etc., in the images as feature points to calculate the homography matrix between each image, mapping all images uniformly to a two-dimensional coordinate system of the accident scene. During this process, the system performs sharpness detection and coverage analysis of key areas (such as the vehicle collision point and debris) on each image after coordinate transformation. If an image is found to be blurry or missing information in a key area, the system will prompt the investigators to retake the image, or automatically extract the corresponding information from clear images from other angles for fusion and completion, until all key information is clearly and completely covered, ultimately generating a valid accident scene image for subsequent analysis.

[0058] The above technical solution effectively solves the problems of poor quality, incomplete information, or spatial inconsistencies that may exist in the original accident scene images, ensuring that the effective accident scene images used for subsequent accident feature extraction and simulation reconstruction have high definition, high integrity, and spatial consistency. This significantly improves the accuracy of accident scene feature extraction, provides a reliable data foundation for building accurate accident scene simulation models, and thus makes the final traffic accident liability determination results more objective, accurate, and credible.

[0059] In some embodiments, the effectiveness of the coordinate transformation scene image is evaluated, and information is supplemented based on the effectiveness evaluation result. This includes: evaluating the clarity, view coverage, and information integrity of the coordinate transformation scene image based on a preset evaluation strategy to obtain an effectiveness evaluation result; when the effectiveness evaluation result characterizes the evaluation as unsuccessful, replacing and / or supplementing information in the coordinate transformation scene image until the effectiveness evaluation result characterizes the evaluation as successful, thereby obtaining an effective accident scene image.

[0060] An evaluation strategy refers to a set of rules, standards, and algorithms pre-defined based on actual needs and experience before image evaluation. These strategies may include image quality evaluation metrics, such as sharpness thresholds and noise tolerance; methods for calculating viewpoint coverage; and key information identification standards. The purpose of setting these strategies is to provide objective and quantitative basis for subsequent automated evaluation.

[0061] Sharpness assessment aims to measure the visual quality of an image, ensuring that details are sufficiently sharp and discernible to eliminate blurry images caused by factors such as camera shake, inaccurate focus, or insufficient lighting. Implementation methods may include, but are not limited to: calculating the energy of the high-frequency components of the Fourier transform of the image, calculating the image gradient using the Laplacian or Sobel operators, and employing blind image quality assessment (BIQA) algorithms such as BRISQUE or NIQE.

[0062] Viewpoint coverage assessment aims to check whether multiple sets of accident scene images have captured the accident scene completely from a sufficient number of angles, avoiding occlusion or omission of key areas to ensure the comprehensiveness of subsequent 3D reconstruction and feature extraction. Implementation methods may include, but are not limited to: estimating camera pose through image feature point matching and multi-view geometry methods, and calculating the overlap rate and coverage between images and between images and the entire scene; alternatively, a list of key areas can be pre-defined, and object detection algorithms can be used to determine whether these key areas are clearly captured in images from at least N different viewpoints.

[0063] Information integrity assessment aims to determine whether an image contains all necessary accident-related information, such as vehicle damage, road signs, and the locations of traffic participants, to ensure the comprehensiveness and accuracy of subsequent accident scene feature extraction. Implementation methods may include, but are not limited to: using deep learning models to perform object detection and semantic segmentation on the image, identifying and statistically analyzing preset key objects and their attributes, and comparing them with a preset integrity list; or, analyzing whether there are large blank areas, occluded areas, or missing data areas in the image.

[0064] When any one or more of the sharpness assessment, viewpoint coverage assessment, and / or information integrity assessment fails to meet the preset assessment strategy standards, the validity assessment result is considered unsuccessful. In this case, replacement refers to requiring a re-capture or acquiring new, higher-quality images from other data sources to replace the unqualified images when the existing image quality is too low or information is severely missing. Information completion refers to supplementing missing information in existing images using technical means without re-acquiring images. Implementation methods may include, but are not limited to: using image inpainting techniques to intelligently fill in occluded or missing areas in the image; or combining other auxiliary data to supplement information on difficult-to-identify or missing objects in the image; or using multimodal data fusion technology to integrate data acquired from different sensors to compensate for the deficiencies of a single image. The validity assessment continues until the result passes, representing an iterative or cyclical process. After performing replacement and / or information completion operations on the coordinate transformation scene image, the system will re-evaluate the validity of the processed image. This process will continue until all assessment indicators meet the preset assessment strategy.

[0065] The proposed solution for evaluating the validity of coordinate-transformed on-site images goes beyond a simple one-time judgment. Instead, it introduces a refined, multi-dimensional evaluation mechanism based on a pre-defined evaluation strategy, combined with an iterative replacement and / or information completion process. Specifically, the system first evaluates the sharpness, viewpoint coverage, and information integrity of the coordinate-transformed on-site images according to the pre-defined evaluation strategy. These evaluations quantify the image quality and usability from different dimensions. If any evaluation result indicates that the image does not meet the requirements, i.e., the validity evaluation result indicates a failure, the system does not directly discard these images but triggers a replacement and / or information completion mechanism. This means that, depending on the specific defects, a higher-quality image can be acquired for replacement, or advanced image processing technology can be used to intelligently complete the missing information in the existing image. Crucially, after the replacement or completion is completed, the system performs another validity evaluation on the updated image, forming a closed-loop feedback and correction process. This iterative process continues until all evaluation metrics meet preset standards, ensuring that the final output of valid accident scene images possesses high definition, wide-view coverage, and complete information. This provides a solid and reliable data foundation for subsequent accident feature extraction, accident scene simulation reconstruction, and final traffic accident liability determination. This mechanism effectively compensates for the data quality deficiencies that may result from single evaluations or one-time processing, significantly improving the accuracy and reliability of the entire liability determination process.

[0066] The following is a concrete example to illustrate this. When evaluating the effectiveness of coordinate transformation scene images, the following process can be adopted: First, for sharpness evaluation, the local variance of the image can be calculated or MTF curve analysis can be used. A threshold is set where the image's average local variance is below a certain empirical value, indicating that the sharpness is unacceptable. For viewpoint coverage evaluation, a deep learning-based scene understanding model can be used to identify key objects in the image. Combined with multi-view geometry algorithms, the projection coverage of these key objects in different images can be calculated. If the coverage of key objects in the scene is less than 90% or there are large blind spots, the viewpoint coverage is considered insufficient. For information integrity evaluation, a list of key information about the accident scene can be pre-defined, such as vehicle type, license plate number, damaged parts, road markings, and traffic light status. Then, OCR technology is used to identify the license plate number, and object detection and semantic segmentation models are used to identify vehicle damage and road markings. If more than 20% of the key information in the list cannot be extracted or identified from the image, the information is considered incomplete. If any evaluation result fails, the user can be prompted to retake an image from a specific angle to replace it, or the image restoration algorithm can be automatically invoked to intelligently fill in occluded areas in the image, such as repairing road markings obscured by leaves. After the replacement or completion is completed, the above evaluation process will be repeated until all evaluation items pass, ultimately resulting in a high-quality and valid accident scene image.

[0067] Through the above technical solution, this application effectively addresses the problems of poor image quality, missing information, or incomplete perspective that may occur during the acquisition of valid accident scene images. By introducing a pre-defined evaluation strategy to conduct multi-dimensional and refined evaluation of clarity, perspective coverage, and information completeness, and combining iterative replacement and / or information completion mechanisms, the final obtained valid accident scene images are ensured to have high reliability and high usability. This significantly improves the accuracy of subsequent accident feature extraction, enabling accident scene features to more realistically and comprehensively reflect the situation at the scene. Furthermore, when performing simulation reconstruction based on these high-quality features, a more accurate accident scene simulation model can be generated, providing a more solid and reliable data foundation for the final determination of liability in traffic accidents, thereby improving the objectivity and fairness of the traffic accident liability determination results.

[0068] See Figure 2 In one embodiment, the method for extracting accident features from valid accident scene images includes, but is not limited to, steps S201 to S204.

[0069] Step S201: Perform key area detection on the valid accident scene images to obtain the key accident areas.

[0070] Step S202: Extract the regional contour features and key point features of the critical area of ​​the accident.

[0071] Step S203: Based on the region contour features and key point features, determine the semantic label features that match the key areas of the accident.

[0072] Step S204: Integrate region contour features, key point features, and semantic label features to obtain accident scene features.

[0073] Key region detection is performed on the valid accident scene images to identify and locate core areas directly related to the accident from complex accident scene images, such as vehicles involved, pedestrians, obstacles, and road traces. This can be achieved through various techniques. For example, deep learning-based object detection models (such as YOLO and Faster R-CNN) can be used to automatically identify and box key targets by training on a large number of accident scene images; alternatively, traditional image processing methods, such as segmentation algorithms based on color, texture, or shape features, can be used to distinguish key objects in the foreground from the background. Key accident regions refer to the image regions directly related to the traffic accident identified after key region detection. These regions are the focus of subsequent feature extraction and analysis, and their accuracy directly affects the quality of the final accident scene features.

[0074] This study extracts regional contour features and key point features from the critical areas of an accident, aiming to accurately describe these areas at both geometric and local detail levels. Regional contour features can be precise depictions of the critical area boundaries, such as edge information obtained through edge detection algorithms (e.g., Canny or Sobel operators) or pixel-level contours obtained through image segmentation techniques (e.g., semantic segmentation or instance segmentation). Key point features refer to salient and repeatable points in the image, such as corner points (e.g., Harris or Shi-Tomasi corners), Scale Invariant Feature Transform (SIFT) feature points, or Speed-Up Robust Feature Transform (SURF) feature points. These points are crucial for image registration, pose estimation, and 3D reconstruction.

[0075] Based on regional contour features and key point features, semantic label features matching the critical areas of the accident are determined, aiming to imbue geometric features with high-level semantic information. Semantic label features are used to describe the type, damage level, and / or spatial location information of targets within the critical accident areas. For example, type information can indicate whether the critical area is a "car," "truck," "pedestrian," or "motorcycle"; damage level information can describe the vehicle as having "minor scratches," "moderate dents," or "severe deformation"; spatial location information can indicate the target's relative or absolute position and posture within the scene. Determining these semantic labels can be achieved through image classification models, attribute recognition models, or posture estimation models. These models are typically based on deep neural networks, learning from large amounts of labeled data to identify high-level semantic information in images.

[0076] This application's solution extracts refined accident features from valid accident scene images. First, it utilizes key region detection technology to focus the analysis on critical accident areas, effectively avoiding interference from irrelevant background information. Then, by extracting the regional contour features and key point features of the critical accident areas, it accurately captures the shape, boundaries, and local details of the targets at a geometric level, providing a high-precision geometric foundation for subsequent simulation reconstruction. Building upon this, it further combines regional contour features and key point features to determine semantic label features matching the critical accident areas. These semantic label features not only provide target type information but also include crucial damage degree and spatial location information, greatly enriching the understanding of the accident scene. Finally, these multi-level, multi-modal features are integrated to form a comprehensive and detailed accident scene feature set. This detailed feature extraction process enables subsequent traffic accident scene simulation reconstruction to obtain more accurate input data, thereby constructing an accident scene simulation model that more closely resembles the real situation. This provides a solid foundation for inverse and forward iterative simulations based on this model, significantly improving the accuracy and reliability of traffic accident liability determination results.

[0077] The following is a concrete example to illustrate this. When extracting accident features from valid accident scene images, a pre-trained YOLOv8 model can first be used to detect key regions in the image, identifying all vehicles and pedestrians, and labeling these detected target regions as key accident regions. Next, for each key accident region, a Mask R-CNN model can be used for instance segmentation, thereby accurately extracting pixel-level region contour features of vehicles or pedestrians. Simultaneously, the SIFT algorithm can be used to detect and extract key point features within these key regions, generating descriptors. Subsequently, based on these region contour features and key point features, a multi-task deep learning network can be trained. One branch of this network is used for vehicle classification (e.g., sedans, SUVs, trucks), another branch is used to assess the degree of vehicle damage (e.g., minor scratches, dented doors, severe front-end deformation), and a third branch is used to estimate the vehicle's pose and relative spatial position in the image. These classification results, damage assessment results, and pose information constitute the semantic label features. Finally, the pixel masks of these region contours, key point descriptors, and semantic labels (such as vehicle type ID, damage level code, and attitude quaternion) are integrated into a structured data object as the final accident scene feature.

[0078] Through the aforementioned technical solution, this application can extract more refined and comprehensive accident scene features from valid accident scene images. This detailed feature description, including precise geometric contours, rich local key points, and high-level semantic information (such as target type, damage level, and spatial location), greatly enhances the depth of understanding of the accident scene. Therefore, in subsequent traffic accident scene simulation and reconstruction processes, a more realistic and accurate accident scene simulation model can be constructed based on these high-quality input features, thereby providing more reliable and refined data support for determining liability in traffic accidents and effectively avoiding liability deviations caused by insufficient feature extraction.

[0079] In some embodiments, before simulating and reconstructing the traffic accident scene, the method further includes: capturing reference images of the traffic accident scene from at least three shooting angles, correcting the positional deviation of the accident scene features to obtain corrected accident scene features; performing feature matching between the accident auxiliary features and the map features corresponding to the traffic accident scene, and updating the accident auxiliary features based on the feature matching results until the similarity between the updated accident auxiliary features and the map features reaches a preset similarity threshold; and supplementing the corrected accident scene features and the updated accident auxiliary features to the map features corresponding to the traffic accident scene.

[0080] The purpose of capturing reference images of the traffic accident scene from at least three shooting angles is to obtain more comprehensive on-site data through multi-view information, providing a foundation for subsequent accurate 3D reconstruction and location correction. These reference images can be obtained by taking pictures of the accident scene from different positions and angles using specialized equipment (e.g., a high-resolution camera mounted on a drone) or general-purpose equipment (e.g., a smartphone).

[0081] Positional deviation correction of accident scene features aims to rectify spatial inaccuracies in the features extracted from valid accident scene images. This correction can be achieved through 3D reconstruction using multi-view reference images. Techniques such as triangulation and bundle adjustment can accurately map feature points from 2D images into 3D space, thereby eliminating or reducing positional deviations. Feature matching between the accident auxiliary features and the map features corresponding to the traffic accident scene aligns the auxiliary information of the accident scene (such as GPS coordinates and road segment information) with authoritative Geographic Information System (GIS) data or high-precision map data.

[0082] Feature matching can be achieved by identifying and comparing common geographic entities (such as road intersections, buildings, traffic signs, etc.), for example, using image-based geolocation technology or text-based semantic matching technology. Updating the accident auxiliary features based on the feature matching results means adjusting and optimizing the original accident auxiliary features according to the accuracy of the matching. For example, if there is a deviation between the initial GPS coordinates and the road location on the map, the GPS coordinates can be updated based on map features. This update can be performed using algorithms such as weighted averaging, Kalman filtering, or iterative optimization. The update process continues until the similarity between the updated accident auxiliary features and the map features reaches a preset similarity threshold, indicating that the update process is an iterative optimization process, the goal of which is to ensure a high degree of consistency between the accident auxiliary features and the map features.

[0083] The corrected accident scene features and updated accident auxiliary features are added to the map features corresponding to the traffic accident scene. The aim is to supplement the corrected accident details and geographical environment information to the map features corresponding to the traffic accident scene, so as to create a spatial data model containing all relevant elements of the accident scene, and provide a comprehensive and accurate data foundation for subsequent simulation reconstruction.

[0084] This application's solution significantly improves the spatial and geographical accuracy of traffic accident scene data by introducing multi-angle reference images, positional deviation correction, map feature matching, and an iterative update mechanism. Specifically, firstly, reference images taken from at least three different angles provide a rich data source for the precise 3D localization of accident scene features. These multi-view images can be finely calibrated using advanced computer vision techniques, such as multi-view geometry or photogrammetry, to obtain more accurately corrected accident scene features. Simultaneously, to ensure high consistency between the geographical location information of the accident scene and the real world, this solution performs feature matching between accident auxiliary features and authoritative map features. This matching process is not completed in one step but is iteratively updated to continuously optimize the accident auxiliary features until their similarity to the map features reaches a preset high-precision threshold. This iterative matching and update mechanism effectively solves the geographical deviation problem that may exist in the initial auxiliary features. Finally, the precisely position-corrected accident scene features and the geographically calibrated accident auxiliary features are added to the map features corresponding to the traffic accident scene. In this way, this solution constructs a highly accurate digital model of the accident scene with complete geographic information and precise spatial location. This model, as input for subsequent traffic accident scene simulation and reconstruction, can significantly improve the realism and reliability of the simulation model, thus laying a solid foundation for generating more accurate and convincing traffic accident liability determination results.

[0085] The following is a concrete example. Before simulating and reconstructing a traffic accident scene, a drone equipped with a high-precision positioning module can be used to take panoramic photos of the accident scene from at least three different altitudes and horizontal positions, acquiring a series of high-resolution reference images. These images not only contain key accident scene features such as the vehicles involved in the accident and debris, but also detailed information about the surrounding environment. Subsequently, using these reference images, combined with accident scene features extracted from valid accident scene images, such as the outlines of the vehicles involved, the point of impact, and the location of debris, the system processes them using 3D reconstruction software (e.g., software based on SfM and MVS algorithms) to generate a precise 3D point cloud model of the accident scene. During this process, the positions of the initially extracted accident scene features in 3D space are precisely corrected, resulting in corrected accident scene features. Simultaneously, the system can acquire auxiliary accident features such as GPS coordinates, road type, and speed limit information. These auxiliary features are matched with pre-loaded high-precision map data (containing geographical features such as road networks, lane lines, traffic signs, and building outlines). For example, the system can identify the road name and GPS coordinates in the auxiliary features and compare them with the corresponding road segments in the map data. If a slight deviation is detected between the GPS coordinates and the actual road location on the map, the system will adjust the GPS coordinates based on the map data and update auxiliary features such as road type and speed limit. This matching and updating process will continue until the similarity between the updated auxiliary features (e.g., the distance error between the corrected GPS coordinates and the map road) and the map features (e.g., a distance error of less than 1 meter) reaches a preset threshold. Finally, this precisely corrected 3D accident scene feature model and the geographically calibrated accident auxiliary features are integrated into the high-precision map data to form a unified digital model of the accident scene with real-world geographic coordinates.

[0086] By employing the aforementioned technical solution, this application effectively addresses the potential spatial location deviations and geographic information inconsistencies between initial accident scene features and auxiliary accident features. Positional deviation correction using multi-angle reference images ensures the spatial accuracy of accident scene features; iterative matching and updating with map features guarantees the geographic precision of auxiliary accident features. This enables subsequent traffic accident scene simulation reconstruction to be based on highly accurate and geographically complete data, significantly improving the realism and reliability of the simulation model. Therefore, this application can generate more accurate and reliable traffic accident liability determination results, effectively reducing disputes over liability determination and improving the efficiency and fairness of accident handling.

[0087] See Figure 3 In one embodiment, the method for simulating and reconstructing a traffic accident scene includes, but is not limited to, steps S301 to S304.

[0088] Step S301: Convert the accident scene features and accident auxiliary features into corresponding three-dimensional point cloud models.

[0089] Step S302: Divide the 3D point cloud model into triangular meshes, identify key regions in the 3D point cloud model, refine the meshes in the key regions, and generate an adaptive mesh model.

[0090] Step S303: The 3D model of the vehicle or vulnerable road user is fused with the adaptive mesh model in a unified coordinate system to generate a static accident scene model.

[0091] Step S304: Assign physical material properties to each object in the static accident scene model, define the contact and collision relationships between objects, and obtain the accident scene simulation model.

[0092] This application's solution first transforms abstract accident scene features and auxiliary accident features into a 3D point cloud model, laying the foundation for subsequent 3D reconstruction. Given the potential inhomogeneity or lack of detail in the original point cloud, this solution further triangulates the 3D point cloud model and intelligently identifies key areas in the accident, such as collision points, vehicle deformation sites, or pedestrian positions. These key areas undergo mesh refinement, generating an adaptive mesh model that covers the entire scene while precisely representing key details. Based on this, to ensure the realism and integrity of the simulation model, pre-set high-precision 3D models of vehicles or vulnerable road users are accurately integrated into this adaptive mesh model. All models are aligned in a unified coordinate system, thus constructing a static accident scene model containing all relevant objects. To enable this static model for physical simulation, this solution further assigns realistic physical material properties to each object in the static accident scene model and explicitly defines the possible contact and collision relationships between objects. Through this series of steps, a high-precision, physically realistic accident scene simulation model is gradually constructed from the original two-dimensional features and auxiliary information. This model not only accurately reproduces the accident scene geometrically, but also provides reliable input for subsequent reverse and forward simulations in terms of physical properties. This effectively solves the problem of insufficient precision and missing physical properties in traditional methods when constructing simulation models, leading to inaccurate simulation results, and provides more solid technical support for determining liability in traffic accidents.

[0093] The following is a concrete example to illustrate this. When converting accident scene features and auxiliary accident features into a 3D point cloud model, a structured light scanner or LiDAR can be used to scan the accident scene and directly acquire high-density 3D point cloud data. Alternatively, multiple accident scene images from different perspectives can be used to reconstruct the 3D point cloud of the scene from the images using the Structure from Motion (SfM) algorithm and the Multi-View Stereo (MVS) algorithm. When triangulating the 3D point cloud model and generating an adaptive mesh model, methods such as Poisson Reconstruction or Screen-Space Ambient Occlusion can be used to convert the point cloud into a surface mesh. Key regions can then be identified by analyzing the curvature of the mesh or by combining semantic segmentation results. For example, the mesh density of the vehicle collision area can be doubled. When fusing 3D models of vehicles or vulnerable road users, a 3D CAD model matching the accident vehicle model or person's body size can be selected from a pre-defined model library. The Iterative Closest Point (ICP) algorithm is then used to precisely align and fuse these models with the adaptive mesh model. Finally, when assigning physical material properties and defining contact and collision relationships, different coefficients of friction and elasticity can be assigned to vehicle components (such as the body and tires), and corresponding roughness can be assigned to the road surface. The API interfaces provided by physics engines (such as NVIDIA PhysX or BulletPhysics) can be used to set the shape of the collider (such as bounding box or convex hull) and collision response rules, ensuring that the simulation model can accurately simulate physical processes such as vehicle collisions and sliding.

[0094] Through the aforementioned technical solution, this application can efficiently and accurately transform the two-dimensional features and auxiliary information obtained from the accident scene into a high-fidelity, physically realistic accident scene simulation model. This model not only meticulously recreates the complex structure of the accident scene geometrically, but also ensures simulation accuracy in critical areas through mesh refinement processing of key regions. Simultaneously, by integrating high-precision three-dimensional models of vehicles or vulnerable road users, and assigning accurate physical material properties to each object and defining contact and collision relationships, the constructed simulation model possesses the capability for accurate physical simulation. This significantly improves the realism and reliability of the accident scene simulation reconstruction, providing a solid foundation for subsequent reverse and forward deductions based on the simulation model. This effectively avoids liability determination biases caused by insufficient model accuracy or missing physical properties, greatly improving the accuracy and persuasiveness of traffic accident liability determination results.

[0095] See Figure 4In one embodiment, the method for generating traffic accident liability determination results for a traffic accident scene includes, but is not limited to, steps S401 to S403.

[0096] Step S401: Based on the accident scene simulation model, perform reverse simulation of the traffic accident scene to deduce the initial motion simulation parameters of the traffic accident scene from the final state.

[0097] Step S402: Based on the initial motion simulation parameters, perform forward iterative simulation of the traffic accident scene. During the forward iterative simulation, dynamically adjust the input parameters so that the error between the forward simulation result model and the accident scene simulation model is less than a preset error threshold.

[0098] Step S403: Based on the positive simulation result model and the traffic accident liability determination strategy, generate the traffic accident liability determination result for the traffic accident scene.

[0099] This application's solution addresses the problem of accurately determining liability based solely on static accident scene simulation models by introducing dynamic simulation analysis. First, a reverse simulation is performed using a pre-constructed accident scene simulation model (representing the final static state of the accident). This step aims to deduce the initial motion simulation parameters before the accident, such as the initial speed and direction of the vehicles involved, from the final outcome of the accident. This process continuously attempts and adjusts the initial conditions through a physics engine or optimization algorithm to ensure that the forward simulation can reproduce the final state of the accident scene. Subsequently, based on these derived initial motion simulation parameters, a forward iterative simulation is performed. During the forward simulation, the system dynamically adjusts input parameters, such as vehicle braking intensity, steering angle, and road friction coefficient, to ensure that the error between the forward simulation result model and the actual accident scene simulation model is less than a preset error threshold. This iterative adjustment mechanism guarantees a high degree of realism and accuracy in the simulation process, enabling the forward simulation result model to accurately reproduce the entire process of the accident. Ultimately, based on this highly accurate and dynamic forward simulation model of the accident, combined with a pre-defined traffic accident liability determination strategy, the system can comprehensively and objectively analyze various dynamic factors and the behaviors of all parties involved in an accident, thereby generating more accurate and convincing traffic accident liability determination results. Through this dynamic simulation method combining reverse inference and forward verification, this application overcomes the limitations of static models in liability determination analysis, significantly improving the scientific rigor and reliability of liability determination.

[0100] The following is a concrete example. After obtaining the accident scene simulation model, a physics engine-based inverse solver can be used to perform inverse simulation on the model. For example, the solver can receive the final position, attitude, and damage information of the vehicles in the accident scene simulation model, and then, through optimization methods such as Monte Carlo search or genetic algorithms, iteratively calculate the initial motion simulation parameters that lead to these final states, such as the velocity vectors and directions of each vehicle before the collision, under preset physical constraints (such as vehicle type and road friction coefficient range). Next, these derived initial motion simulation parameters are used as input to start a forward physics simulation program. During the forward simulation, the differences between the simulation result model and the accident scene simulation model (such as the collision point, the final stopping position of the vehicles, and the degree of deformation) can be monitored in real time. If the error exceeds a preset error threshold (for example, the final position deviation of the vehicles exceeds 50 cm), the simulation input parameters can be dynamically adjusted, such as fine-tuning the initial speed, collision stiffness, or braking delay time, and the simulation can be rerun until the forward simulation result model and the accident scene simulation model are highly consistent. Once a forward simulation model meeting the error requirements is obtained, it will contain dynamic data of the entire accident process, including the trajectories of each vehicle, speed changes, and the forces acting on them at the moment of collision. At this point, a pre-defined traffic accident liability determination strategy can be implemented. This strategy can be an expert system that includes traffic regulations and detailed rules for assigning responsibility. For example, the system can analyze whether vehicles in the simulation model exhibit behaviors such as speeding, running red lights, failing to maintain a safe distance, or failing to yield, and match these behaviors with the rules in the traffic accident liability determination strategy to ultimately generate a detailed traffic accident liability determination result, including the responsibility ratios and basis for each party.

[0101] Through the above technical solution, this application overcomes the limitations of traditional liability determination methods that rely solely on static accident scene information. By introducing a dynamic analysis mechanism combining reverse simulation and forward iterative simulation, this application can accurately deduce the initial motion parameters before the accident from its final static state. Based on this, forward iterative simulation dynamically verifies and refines the entire accident process. This method enables the generated forward simulation model to highly realistically reproduce the dynamic evolution of the accident, including vehicle trajectories, speed changes, forces at the moment of collision, and the behavioral patterns of all parties. Therefore, when determining liability for traffic accidents based on this highly accurate dynamic simulation model, various dynamic factors and the behaviors of all parties during the accident process can be considered more comprehensively and objectively, significantly improving the scientific rigor, accuracy, and persuasiveness of the liability determination results, and effectively avoiding disputes caused by insufficient information or judgment bias.

[0102] See Figure 5This application also provides a traffic accident liability determination device, which can implement the above-mentioned traffic accident liability determination method. The device includes: The first module 501 is used to acquire valid accident scene images and accident auxiliary features; the valid accident scene images include valid image features of the traffic accident scene, and the accident auxiliary features include the location features of the traffic accident scene, the weather features, the road condition features, and / or the preliminary accident liability determination features; The second module 502 is used to extract accident features from valid accident scene images to obtain accident scene features; The third module 503 is used to simulate and reconstruct the traffic accident scene based on the accident scene features and accident auxiliary features, and obtain the accident scene simulation model. Module 4, 504, is used to generate traffic accident liability determination results based on the accident scene simulation model and the preset traffic accident liability determination strategy.

[0103] The specific implementation method of this traffic accident liability determination device is basically the same as the specific implementation method of the above-mentioned traffic accident liability determination method, and will not be described again here.

[0104] Figure 6 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application.

[0105] The following reference Figure 6 To describe an electronic device 600 according to such an embodiment of the present disclosure. Figure 6 The electronic device 600 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0106] like Figure 6 As shown, the electronic device 600 is presented in the form of a general-purpose computing device. The components of the electronic device 600 may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different system components (including storage unit 620 and processing unit 610), a display unit 640, etc.

[0107] The storage unit stores program code, which can be executed by the processing unit 610, causing the processing unit 610 to perform the steps described in the above-described section on the method for determining liability in traffic accidents according to various exemplary embodiments of this disclosure.

[0108] Storage unit 620 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 6201 and / or cache memory 6202, and may further include a read-only memory (ROM) 6203.

[0109] Storage unit 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0110] Bus 630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0111] Electronic device 600 can also communicate with one or more external devices 600' (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 650. Furthermore, electronic device 600 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 660. Network adapter 660 can communicate with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0112] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0113] The traffic accident liability determination method, apparatus, equipment, and medium provided in this application extract accident features from valid accident scene images and perform simulation reconstruction based on the extracted accident scene features and auxiliary features. This enables the construction of a more accurate and objective accident scene simulation model, resulting in a more realistic and credible simulated traffic accident scene. By generating liability determination results based on the accident scene simulation model and a preset traffic accident liability determination strategy, the method allows for objective analysis of the accident process based on the simulation model and provides a liability judgment according to the preset strategy. This improves the efficiency and fairness of liability determination and reduces the uncertainty caused by human intervention.

[0114] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, or network device, etc.) to execute the methods described above according to the embodiments of this disclosure.

[0115] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0116] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0117] Those skilled in the art will understand that the above modules can be distributed in the device as described in the embodiments, or they can be modified accordingly and placed in one or more devices that are unique to this embodiment. The modules in the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0118] Exemplary embodiments of this disclosure have been specifically shown and described above. It should be understood that this disclosure is not limited to the detailed structures, arrangements, or implementations described herein; rather, this disclosure is intended to cover various modifications and equivalent arrangements contained within the spirit and scope of the appended claims.

Claims

1. A method for determining liability in traffic accidents, characterized in that, include: Acquire effective accident scene images and auxiliary accident features; The effective accident scene images contain effective visual features of the traffic accident scene, and the auxiliary accident features include the on-site location features, on-site weather features, on-site road condition features, and / or preliminary accident liability determination features. Accident features are extracted from the valid accident scene images to obtain accident scene features; Based on the accident scene features and the accident auxiliary features, the traffic accident scene is simulated and reconstructed to obtain an accident scene simulation model; Based on the accident scene simulation model and the preset traffic accident liability determination strategy, a traffic accident liability determination result is generated for the traffic accident scene.

2. The method for determining liability in traffic accidents according to claim 1, characterized in that, Before acquiring valid accident scene images and auxiliary accident features, the method further includes: Acquire multiple sets of accident scene images taken from multiple angles at the traffic accident scene; The accident scene images are preprocessed to obtain preprocessed scene images; The preprocessed on-site images are subjected to coordinate transformation to map each preprocessed on-site image to the same coordinate system, thereby obtaining coordinate transformed on-site images; The validity of the coordinate-transformed on-site image is evaluated, and information is supplemented based on the validity evaluation results to obtain the valid accident scene image.

3. The method for determining liability in traffic accidents according to claim 2, characterized in that, The process of evaluating the effectiveness of the coordinate transformation site image and supplementing the information in the coordinate transformation site image based on the effectiveness evaluation results includes: Based on a preset evaluation strategy, the sharpness, viewpoint coverage, and information integrity of the coordinate transformation scene image are evaluated to obtain the effectiveness evaluation result. When the effectiveness evaluation result indicates that the evaluation fails, the coordinate transformation scene image is replaced and / or information is supplemented until the effectiveness evaluation result indicates that the evaluation passes, thus obtaining the effective accident scene image.

4. The method for determining liability in traffic accidents according to claim 1, characterized in that, The step of extracting accident features from the valid accident scene images includes: The key areas of the accident are obtained by performing key area detection on the effective accident scene images; Extract the regional contour features and key point features of the critical area of ​​the accident; Based on the region contour features and the key point features, semantic label features matching the key accident region are determined; the semantic label features are used to describe the type information, damage degree information and / or spatial location information of the target in the key accident region. The accident scene features are obtained by integrating the region contour features, the key point features, and the semantic label features.

5. The method for determining liability in traffic accidents according to claim 1, characterized in that, Before performing a simulation reconstruction of the traffic accident scene, the following is also included: Reference images of the traffic accident scene are captured from at least three shooting angles, and the positional deviation of the accident scene features is corrected to obtain the corrected accident scene features. The accident auxiliary features are matched with the map features corresponding to the traffic accident scene, and the accident auxiliary features are updated based on the feature matching results until the similarity between the updated accident auxiliary features and the map features reaches a preset similarity threshold. The corrected accident scene features and the updated accident auxiliary features are added to the map features corresponding to the traffic accident scene.

6. The method for determining liability in traffic accidents according to claim 1, characterized in that, The process of simulating and reconstructing the traffic accident scene based on the accident scene features and the accident auxiliary features includes: The accident scene features and the accident auxiliary features are converted into corresponding three-dimensional point cloud models; The three-dimensional point cloud model is divided into triangular meshes, and key regions in the three-dimensional point cloud model are identified. The key regions are then subjected to mesh refinement processing to generate an adaptive mesh model. The three-dimensional model of the vehicle or vulnerable road user is fused with the adaptive mesh model in a unified coordinate system to generate a static accident scene model. Physical material properties are assigned to each object in the static accident scene model, and the contact and collision relationships between objects are defined to obtain the accident scene simulation model.

7. The method for determining liability in traffic accidents according to claim 1, characterized in that, The process of generating a traffic accident liability determination result based on the accident scene simulation model and a preset traffic accident liability determination strategy includes: Based on the accident scene simulation model, the traffic accident scene is reverse simulated to deduce the initial motion simulation parameters of the traffic accident scene from the final state. Based on the initial motion simulation parameters, a forward iterative simulation of the traffic accident scene is performed. During the forward iterative simulation, the input parameters are dynamically adjusted so that the error between the forward simulation result model and the accident scene simulation model is less than a preset error threshold. Based on the positive simulation result model and the traffic accident liability determination strategy, a traffic accident liability determination result is generated for the traffic accident scene.

8. A device for determining liability in traffic accidents, characterized in that, include: The first module is used to acquire valid accident scene images and accident auxiliary features; The effective accident scene images contain effective visual features of the traffic accident scene, and the auxiliary accident features include the on-site location features, on-site weather features, on-site road condition features, and / or preliminary accident liability determination features. The second module is used to extract accident features from the valid accident scene images to obtain accident scene features; The third module is used to simulate and reconstruct the traffic accident scene based on the accident scene features and the accident auxiliary features to obtain an accident scene simulation model. The fourth module is used to generate a traffic accident liability determination result for the traffic accident scene based on the accident scene simulation model and the preset traffic accident liability determination strategy.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the traffic accident liability determination method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the traffic accident liability determination method as described in any one of claims 1 to 7.

Images