Traffic accident scene real scene reconstruction method, device, equipment and medium

Abstract

The invention relates to a traffic accident scene live-action reconstruction method and device, equipment and a medium, and the method comprises the steps: carrying out the thinning optimization of a traffic accident scene image through employing a preset Douglas-Peucker algorithm, so as to generate a thinned and optimized traffic accident scene image; performing patch expansion and filtering on the sparse point cloud by adopting a preset PMVS algorithm to obtain dense point cloud corresponding to each traffic accident element; taking the dense point clouds corresponding to the traffic accident elements as geometric constraints, taking the thinned and optimized traffic accident scene image as a texture data source to construct a training data set, and training a preset neural radiation field model to be converged by adopting the training data set to determine a traffic accident scene three-dimensional model; and inputting a to-be-reconstructed traffic accident scene image into the traffic accident scene three-dimensional model to extract three-dimensional space information corresponding to each traffic accident element. According to the method, the accident scene investigation and modeling period can be greatly shortened, and the road blocking time is greatly shortened.

Description

Technical Field

[0001] This application relates to the field of intelligent transportation, and in particular to a method for real-time reconstruction of traffic accident scenes, a corresponding device, electronic equipment, and a computer-readable storage medium. Background Technology

[0002] The highway network system has developed rapidly and is gradually becoming more complete. As the primary means of transportation, the number of automobiles is also growing rapidly, especially in cities, where they have almost become a necessity for every family. With the dramatic increase in the number of vehicles, road traffic accidents are also gradually increasing. In the past three years, the number of traffic accidents in my country has exceeded 240,000 annually, resulting in over 250,000 injuries and fatalities, and direct property losses reaching 1.384 billion yuan. Traffic congestion after accidents easily leads to secondary accidents, and the time required for law enforcement to investigate accident scenes, coupled with the poor accuracy of manual investigations, are particularly prominent problems.

[0003] After a traffic accident, according to the regulations for handling road traffic accidents, the handling of the accident requires multiple procedures. If it is a minor accident and the parties involved have no dispute over the facts and causes, they can immediately leave the scene on their own and settle the matter through negotiation. If they cannot leave the scene on their own, they must protect the scene and promptly report to the public security authorities. After receiving instructions, law enforcement officers will quickly rush to the scene to conduct an investigation and handle the situation. The on-site investigation includes a series of tasks such as on-site interviews, photography, mapping, measurement, and examination. Law enforcement officers need to carry a large amount of investigation equipment to the accident site and need to cordon off the scene for a considerable period of time for investigation. Moreover, it places higher demands on the investigation skills of law enforcement officers; otherwise, it will prolong the road closure time and cause serious congestion. Therefore, whenever a traffic accident occurs, traffic congestion and safety issues become more prominent.

[0004] Currently, the information recorded and carried by traffic accident photos, videos, and texts is very limited, often leading to inaccurate judgments. Therefore, information scarcity and incompleteness are major problems in accident handling. Furthermore, the creation of traditional scene investigation maps requires various complex manual interventions, with almost every step requiring human involvement, which is time-consuming and labor-intensive. With the continuous development of technologies such as autonomous driving, intelligent vehicles, and vehicle-to-everything (V2X), intelligent technologies are increasingly integrated into the vehicle R&D industry, placing higher demands on the efficiency, accuracy, and intelligence of accident handling.

[0005] In summary, given the existing technologies that require law enforcement officers to conduct on-site investigations of accidents for extended periods, have poor accuracy in manual investigations, and require various complex manual interventions to create traditional on-site investigation maps (almost every step requires human involvement, which is time-consuming and labor-intensive), the applicant has made corresponding explorations to address these issues. Summary of the Invention

[0006] The purpose of this application is to solve the above-mentioned problems by providing a method, device, electronic equipment and computer-readable storage medium for real-time reconstruction of traffic accident scenes.

[0007] To achieve the various objectives of this application, the following technical solution is adopted: A method for real-scene reconstruction of traffic accident scenes, proposed to meet one of the purposes of this application, includes: The system acquires multi-view traffic accident scene images corresponding to each traffic accident scene, and uses a preset Douglas-Puk algorithm to thin and optimize the traffic accident scene images to generate thinned and optimized traffic accident scene images. An improved SIFT algorithm is used to narrow the neighborhood range of the descriptor, and features are extracted and matched on the thinned and optimized traffic accident scene image to determine the matching feature point pairs corresponding to each traffic accident element in the traffic accident scene image. The preset openMVG tool is called to triangulate the matching feature point pairs to generate sparse point clouds corresponding to each traffic accident element. The preset PMVS algorithm is used to expand and filter the sparse point clouds to obtain dense point clouds corresponding to each traffic accident element. The dense point cloud corresponding to each traffic accident element is used as a geometric constraint, and the thinned and optimized traffic accident scene image is used as a texture data source to construct a training dataset. The training dataset is used to train the preset neural radiation field model to convergence, so as to determine the traffic accident scene 3D model that has been trained to convergence. The traffic accident scene image to be reconstructed is input into the traffic accident scene 3D model that has been trained to convergence, so as to extract the 3D spatial information corresponding to each traffic accident element in the traffic accident scene image to complete the real-scene reconstruction of the traffic accident scene.

[0008] Optionally, the elements of the traffic accident include the vehicles involved in the accident, road facilities, debris involved in the accident, accident traces, and persons involved in the accident. The vehicles involved in the accident include motor vehicles and non-motor vehicles; the road facilities include lane markings, zebra crossings, directional arrows, curbs, guardrails, traffic lights, traffic signs, and speed bumps; the debris involved in the accident includes parts detached from the vehicles; the accident traces include tire brake marks, bloodstains on the ground, and scattered goods or personal belongings; and the persons involved in the accident include drivers, passengers, pedestrians, or non-motorized vehicle riders. The three-dimensional spatial information includes the three-dimensional coordinates, relative positional relationships, shape and size, and collision-related details of each traffic accident element.

[0009] Optionally, before the step of inputting the traffic accident scene image to be reconstructed into the traffic accident scene 3D model that has been trained to a convergent state, the following steps are included: The system obtains the coordinates of the traffic accident location at the target traffic accident scene and the flight constraints of the UAV's flight environment. The flight constraints include no-fly zone constraints, battery life constraints, flight parameter constraints, static obstacle constraints, and dynamic obstacle constraints. Based on the flight constraints of the UAV's flight environment, the objective function of the flight control model is constructed to minimize the total length of the flight path, maximize the minimum safe distance between the flight path and obstacles, and maximize the smoothness of the flight path. The objective function of the flight control model is solved based on a preset particle swarm optimization algorithm to determine the optimal evidence collection path parameters of the UAV. The drone is controlled to fly to the traffic accident location coordinates of the target traffic accident scene according to the optimal evidence collection path parameters, so as to obtain the traffic accident scene image to be reconstructed at the target traffic accident scene.

[0010] Optionally, based on the flight constraints of the UAV's flight environment, the objective function of the flight control model is constructed to minimize the total length of the flight path, maximize the minimum safe distance between the flight path and obstacles, and maximize the smoothness of the flight path. The objective function of the flight control model is then solved using a preset particle swarm optimization algorithm to determine the optimal evidence-gathering path parameters for the UAV. This step includes: Based on the flight constraints of the UAV flight environment, an objective function for the flight control model is constructed with the optimization objectives of minimizing the total length of the flight path, maximizing the minimum safe distance between the flight path and obstacles, and maximizing the smoothness of the flight path. The total length of the path represents the sum of the Euclidean distances between each discrete path point in the flight path, and the minimum safe distance represents the minimum distance from each discrete path point in the flight path to the boundary of the obstacle. Initialize each initial particle in the initial population of the particle swarm optimization algorithm, where each particle represents the evidence path parameters of the candidate evidence path corresponding to the UAV. Within the flight constraint conditions, the velocity and position of each initial particle are generated. Based on the objective function of the flight control model, the objective function value of the evidence path parameters of the candidate evidence path corresponding to each particle is calculated to determine the individual optimal solution and the global optimal solution. The particle velocity and particle position are iteratively updated. Repeat the above steps until the number of iterations reaches a preset iteration threshold, then stop the iteration to determine the optimal evidence collection path parameters corresponding to the global optimal solution. The optimal evidence collection path parameters include the three-dimensional coordinates of discrete path points and the flight speed between adjacent path points.

[0011] Optionally, the step of using a preset Douglas-Puk algorithm to thin and optimize the traffic accident scene image to generate a thinned and optimized traffic accident scene image includes: The process involves acquiring traffic accident scene images captured by drones, extracting the contour feature point set of each traffic accident element in the traffic accident scene image using an edge detection algorithm, converting the contour feature point set into a two-dimensional point sequence according to pixel coordinates, and removing duplicate and abnormal coordinate noise points in the two-dimensional point sequence. Based on the preset distance threshold corresponding to each traffic accident element of the Douglas-Puk algorithm, the first and last two points of the two-dimensional point sequence corresponding to each traffic accident element are used as the initial baseline segment. The vertical distance from all intermediate feature points in the two-dimensional point sequence to the baseline segment is calculated, and the maximum vertical distance value and the corresponding feature point are selected. Each traffic accident element corresponds one-to-one with each distance threshold, and the distance threshold represents the maximum allowable offset distance from the feature point to the baseline segment. If the maximum vertical distance value is greater than the distance threshold, the feature point corresponding to the maximum vertical distance value is retained, and the two-dimensional point sequence is split into two sub-sequences using it as the dividing point. The vertical distance calculation operation is repeated for the two sub-sequences respectively. If the maximum vertical distance value is less than or equal to the distance threshold, all intermediate points of the current sub-sequence are discarded, and only the first and last points of the sub-sequence are retained. All retained feature points are obtained and reconnected according to the spatial order of the original two-dimensional point sequence to form the thinned traffic accident feature outline. This traffic accident feature outline is then merged with the background layer of the traffic accident scene image to generate the thinned and optimized traffic accident scene image.

[0012] Optionally, the steps of calling the preset OpenMVG tool to triangulate the matching feature point pairs to generate sparse point clouds corresponding to each traffic accident element, and using the preset PMVS algorithm to perform patch expansion and filtering on the sparse point clouds to obtain dense point clouds corresponding to each traffic accident element include: The matching feature point pairs corresponding to each traffic accident element in the traffic accident scene image are obtained, and the camera intrinsic parameters and exterior orientation elements at the time of the traffic accident scene image acquisition are retrieved simultaneously. The camera intrinsic parameters include focal length and principal point coordinates, and the exterior orientation elements include RTK positioning coordinates and IMU attitude angles. The preset openMVG tool is invoked, and the matching feature point pairs, camera intrinsic parameters, and exterior orientation elements are input into the openMVG tool. After optimizing the camera attitude through bundle adjustment, triangulation calculation is performed to generate sparse point clouds corresponding to each traffic accident element. The sparse point clouds corresponding to each traffic accident element are imported into the preset PMVS algorithm. The points of the sparse point cloud are used as seed points to construct initial patches. The patch range is expanded based on the texture consistency constraint of the traffic accident scene image, while abnormal patches that deviate from the contour of the corresponding traffic accident element are removed. The expanded and filtered patches are processed into point clouds to obtain dense point clouds corresponding to each traffic accident element.

[0013] Optionally, after the step of inputting the traffic accident scene image to be reconstructed into the traffic accident scene 3D model that has been trained to convergence state to extract the 3D spatial information corresponding to each traffic accident element in the traffic accident scene image to be reconstructed, the method includes: Law enforcement officers determine liability at the target traffic accident scene based on the spatial information corresponding to each traffic accident element in the image of the traffic accident scene to be reconstructed, thereby completing the real-scene reconstruction of the traffic accident scene.

[0014] A traffic accident scene reconstruction device provided for another purpose of this application includes: The image processing module is configured to acquire multi-view traffic accident scene images corresponding to each traffic accident scene, and use a preset Douglas-Puk algorithm to thin and optimize the traffic accident scene images to generate thinned and optimized traffic accident scene images. The feature point pair determination module is configured to use an improved SIFT algorithm to narrow the descriptor neighborhood range, and to perform feature extraction and matching on the thinned and optimized traffic accident scene image in order to determine the matching feature point pairs corresponding to each traffic accident element in the traffic accident scene image. The dense point cloud determination module is configured to call the preset openMVG tool to triangulate the matching feature point pairs to generate sparse point clouds corresponding to each traffic accident element, and use the preset PMVS algorithm to expand and filter the sparse point clouds to obtain the dense point clouds corresponding to each traffic accident element. The 3D model construction module is configured to use the dense point cloud corresponding to each traffic accident element as a geometric constraint, and the thinned and optimized traffic accident scene image as a texture data source to construct a training dataset. The training dataset is used to train the preset neural radiation field model to convergence, so as to determine the 3D model of the traffic accident scene that has been trained to convergence. The traffic accident reconstruction module is configured to input the traffic accident scene image to be reconstructed into the traffic accident scene 3D model that has been trained to convergence, so as to extract the 3D spatial information corresponding to each traffic accident element in the traffic accident scene image to complete the real-scene reconstruction of the traffic accident scene.

[0015] An electronic device provided for another purpose of this application includes a central processing unit and a memory, the central processing unit being configured to invoke and run a computer program stored in the memory to perform the steps of the traffic accident scene reconstruction method of this application.

[0016] A computer-readable storage medium is provided for another purpose of this application, which stores, in the form of computer-readable instructions, a computer program implemented according to the traffic accident scene reconstruction method, which, when called by a computer, executes the steps included in the corresponding method.

[0017] Compared to existing technologies, this application addresses the problems of long time required for law enforcement personnel to conduct accident scene surveys, poor accuracy of manual surveys, and the need for complex manual intervention in the creation of traditional scene survey maps, where almost every step requires human involvement, resulting in time-consuming and labor-intensive processes. This application offers the following beneficial effects, including but not limited to: Firstly, this application achieves end-to-end automated processing from UAV path planning and image acquisition to image thinning, feature matching, point cloud generation, and 3D model training and reconstruction, which significantly shortens the accident scene investigation and modeling cycle, reduces road closure time, and lowers the possibility of secondary accidents.

[0018] Secondly, this application improves matching accuracy through an enhanced SIFT algorithm, combines point cloud generation and optimization algorithms to complete the geometric information of elements, and then uses 3D modeling technology to restore details such as the shape of the accident vehicle and the distribution of scattered objects, ensuring the geometric accuracy and texture realism of the model and meeting the needs of precise surveying and accident analysis. This application avoids the subjective bias of traditional manual surveying, and data can be retrieved retrospectively for subsequent disputes over liability determination, providing objective evidence for law enforcement and enhancing public credibility.

[0019] Thirdly, this application reduces on-site safety risks. The drone remotely completes image acquisition, eliminating the need for law enforcement officers to stay at the accident scene for extended periods and avoiding safety hazards associated with investigations in traffic. The rapid completion of on-site processing also indirectly reduces additional safety and economic problems caused by traffic congestion.

[0020] Fourth, the three-dimensional model of the traffic accident scene constructed in this application, the objective spatial information provided by the three-dimensional model, combined with traffic regulations, can quickly clarify the attribution of accident responsibility, reduce disputes caused by insufficient evidence, and improve the efficiency and fairness of accident handling. Attached Figure Description

[0021] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1This is a flowchart illustrating the method for real-scene reconstruction of traffic accident scenes in the embodiments of this application; Figure 2 This is a schematic block diagram of the traffic accident scene reconstruction device in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the computer device in the embodiments of this application. Detailed Implementation

[0022] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0023] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0024] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0025] Those skilled in the art will understand that the terms "client," "terminal," and "terminal device" as used herein include both devices that receive wireless signals, devices that only possess wireless signal receiver capabilities without transmission capabilities, and devices with receiving and transmitting hardware, devices that have receiving and transmitting hardware capable of bidirectional communication over a bidirectional communication link. Such devices may include: cellular or other communication devices such as personal computers or tablets, having single-line displays, multi-line displays, or cellular or other communication devices without multi-line displays; PCS (Personal Communications Service) that can combine voice, data processing, fax, and / or data communication capabilities; PDAs (Personal Digital Assistants) that may include radio frequency receivers, pagers, internet / intranet access, web browsers, notebooks, calendars, and / or GPS (Global Positioning System) receivers; and conventional laptops and / or handheld computers or other devices that have and / or include radio frequency receivers. As used herein, "client," "terminal," and "terminal device" can be portable, transportable, installed in a means of transportation (air, sea, and / or land), or suitable and / or configured to operate locally and / or in a distributed manner, operating in any other location on Earth and / or in space. "Client," "terminal," and "terminal device" as used herein can also be a communication terminal, an internet access terminal, or a music / video playback terminal, such as a PDA, a MID (Mobile Internet Device), and / or a mobile phone with music / video playback capabilities, or a smart TV, set-top box, etc.

[0026] The hardware referred to by the names "server," "client," and "service node" in this application is essentially an electronic device with the equivalent capabilities of a personal computer. It is a hardware device with the necessary components revealed by the von Neumann architecture, such as a central processing unit (including an arithmetic logic unit and a control unit), memory, input devices, and output devices. The computer program is stored in its memory, and the central processing unit loads the program stored in the secondary storage into the main memory to run it, execute the instructions in the program, and interact with the input and output devices to complete specific functions.

[0027] It should be noted that the concept of "server" used in this application can also be extended to the case of server clusters. Based on the network deployment principles understood by those skilled in the art, the servers should be logically divided. Physically, these servers can be independent of each other but accessible through interfaces, or they can be integrated into a single physical computer or a computer cluster. Those skilled in the art should understand this flexibility and should not use it to constrain the implementation of the network deployment method in this application.

[0028] One or more of the technical features of this application, unless explicitly specified herein, can be deployed on a server and accessed by a client remotely calling the online service interface provided by the server, or can be directly deployed and run on a client for access.

[0029] Unless otherwise specified, the neural network models referenced or potentially referenced in this application may be deployed on a remote server and invoked remotely on the client, or deployed on a client with the capability to invoke directly. In some embodiments, when running on the client, the corresponding intelligence may be acquired through transfer learning in order to reduce the requirements on the client's hardware resources and avoid excessive consumption of the client's hardware resources.

[0030] Unless otherwise specified, all data involved in this application may be stored remotely on a server or on a local terminal device, as long as it is suitable for use by the technical solution of this application.

[0031] Those skilled in the art will understand that although the various methods in this application are described based on the same concept and thus present commonality among them, they can be performed independently unless otherwise specified. Similarly, the various embodiments disclosed in this application are all based on the same inventive concept; therefore, concepts expressed in the same way, as well as concepts that are appropriately changed for convenience but are expressed differently, should be understood equivalently.

[0032] Unless otherwise expressly stated, the various embodiments disclosed in this application can be combined in a cross-cutting manner to flexibly construct new embodiments, as long as such combination does not depart from the inventive spirit of this application and can meet the needs of the prior art or solve a certain deficiency in the prior art. Those skilled in the art should be aware of such modifications.

[0033] Please see Figure 1 In one embodiment of the traffic accident scene reconstruction method of this application, the method includes: Step S10: Obtain multi-view traffic accident scene images corresponding to each traffic accident scene, and use the preset Douglas-Puk algorithm to thin and optimize the traffic accident scene images to generate thinned and optimized traffic accident scene images. The traffic management platform in the terminal device can acquire multi-view traffic accident scene images corresponding to each traffic accident scene, and use a preset Douglas-Puk algorithm to thin and optimize the traffic accident scene images to generate thinned and optimized traffic accident scene images. Specifically, the traffic accident scene images are multi-view image data covering the core area of ​​the accident and its surrounding environment, collected by drones (such as the DJI Phantom 4 RTK model, equipped with a DM5-3600 five-lens tilt camera) at the accident scene. This data serves as the core foundational data for subsequent 3D reconstruction, precise mapping, and accident liability determination. The traffic accident scene images clearly record various elements of the accident scene, including static elements such as vehicles involved in the accident, road facilities, debris, accident traces, and personnel involved, as well as environmental backgrounds such as the terrain, surrounding buildings, and traffic facilities. They also implicitly contain the camera's intrinsic parameters (focal length, principal point coordinates) and external orientation elements (RTK positioning coordinates, IMU attitude angles) at the time of capture.

[0034] In some embodiments, the elements of a traffic accident include vehicles involved in the accident, road facilities, debris involved in the accident, accident traces, and persons involved in the accident. The vehicles involved in the accident include motor vehicles and non-motor vehicles; the road facilities include lane markings, crosswalks, directional arrows, curbs, guardrails, traffic lights, traffic signs, and speed bumps; the debris involved in the accident includes parts detached from the vehicles upon impact; the accident traces include tire brake marks, bloodstains on the ground, and scattered goods or personal belongings; and the persons involved in the accident include drivers, passengers, pedestrians, or non-motorized vehicle riders. The three-dimensional spatial information includes the three-dimensional coordinates, relative positional relationships, shape and size, and collision-related details of each traffic accident element.

[0035] In some embodiments, the step of using a preset Douglas-Puk algorithm to thin and optimize the traffic accident scene image to generate a thinned and optimized traffic accident scene image includes: Step S101: Obtain traffic accident scene images collected by drones, extract the contour feature point set of each traffic accident element in the traffic accident scene image through edge detection algorithm, convert the contour feature point set into a two-dimensional point sequence according to pixel coordinates, and remove duplicate and abnormal noise points in the two-dimensional point sequence. Step S102: Based on the preset distance threshold corresponding to each traffic accident element of the Douglas-Puk algorithm, take the first and last two points of the two-dimensional point sequence corresponding to each traffic accident element as the initial baseline segment, calculate the vertical distance from all intermediate feature points in the two-dimensional point sequence to the baseline segment, and filter out the maximum vertical distance value and the corresponding feature point. Each traffic accident element corresponds one-to-one with each distance threshold, and the distance threshold represents the maximum allowable offset distance from the feature point to the baseline segment. Step S103: If the maximum vertical distance value is greater than the distance threshold, retain the feature point corresponding to the maximum vertical distance value, and use it as the dividing point to split the two-dimensional point sequence into two sub-sequences. Repeat the vertical distance calculation operation for the two sub-sequences respectively. If the maximum vertical distance value is less than or equal to the distance threshold, discard all the intermediate points of the current sub-sequence and retain only the first and last points of the sub-sequence. Step S104: Obtain all retained feature points, reconnect them according to the spatial order of the original two-dimensional point sequence to form the thinned traffic accident element outline, and merge the traffic accident element outline with the background layer of the traffic accident scene image to generate the thinned and optimized traffic accident scene image.

[0036] As demonstrated in the above embodiments, edge detection extracts the contour feature point set of traffic accident elements. Combined with pixel coordinate transformation and noise point removal, core elements such as accident vehicles and debris are accurately separated from the background. This avoids duplicate or abnormal points interfering with subsequent thinning, providing high-quality input data for the algorithm and ensuring the accuracy of the element contours after thinning. A specific distance threshold is matched for each traffic accident element. A baseline segment is constructed based on the beginning and end of the element's two-dimensional point sequence, and the vertical distance to the intermediate point is calculated. This not only accommodates the different contour complexity of different elements (such as vehicles and road markings) but also uses the maximum distance to filter and lock key inflection points, providing accurate criteria for subsequent point selection and avoiding the loss of details or redundant residues caused by a uniform threshold. Based on the comparison results of the maximum distance and the threshold, through recursive operations of segmentation retention and discarding intermediate points, redundant points are removed while preserving the core geometric shape of the elements (such as vehicle collision dents and road marking corners), achieving simplification and improving the efficiency of subsequent feature matching and point cloud generation. By connecting the retained points in the original order and merging them with the background, the topological structure of the elements is consistent with the original image. At the same time, the data volume is reduced by 40% to 60% through thinning. This balances scene integrity and processing efficiency, laying an efficient and accurate data foundation for improving subsequent steps such as SIFT feature extraction and openMVG sparse reconstruction.

[0037] Step S20: The improved SIFT algorithm is used to narrow the neighborhood range of the descriptor, and the features are extracted and matched on the thinned and optimized traffic accident scene image to determine the matching feature point pairs corresponding to each traffic accident element in the traffic accident scene image. The system acquires multi-view traffic accident scene images corresponding to each traffic accident scene, and uses a preset Douglas-Puk algorithm to thin and optimize the traffic accident scene images to generate thinned and optimized traffic accident scene images. Then, it uses an improved SIFT algorithm to narrow the descriptor neighborhood range and performs feature extraction and matching on the thinned and optimized traffic accident scene images to determine the matching feature point pairs corresponding to each traffic accident element in the traffic accident scene images. Specifically, the steps of using an improved SIFT algorithm to narrow the descriptor neighborhood and extracting and matching features from the thinned and optimized traffic accident scene image to determine the matching feature point pairs corresponding to each traffic accident element in the traffic accident scene image include: Step S201: Using the thinned and optimized traffic accident scene image as input, this image has been processed using the Douglas-Puk algorithm to remove redundant pixels, retaining only the complete outlines of core elements such as accident vehicles, road markings, and debris. The mapping relationship between the element pixel coordinates and the spatial location of the original scene has been locked. The input image is converted to grayscale to eliminate redundant information interference from the color channels. Simultaneously, a correspondence table of "grayscale pixel coordinates - original image element outline coordinates" is saved, providing a data benchmark for subsequent accurate association of feature points to specific traffic accident elements and avoiding misalignment between feature points and element outlines.

[0038] Step S202: Based on the grayscale image obtained in step S201, a Gaussian pyramid (by sequentially blurring the image using Gaussian kernels of different standard deviations to generate multi-scale image layers) and a Difference-of-Gaussian (DoG) pyramid (by subtracting Gaussian images from adjacent scales) are constructed. For each pixel in the Difference-of-Gaussian (DoG) pyramid, its grayscale value is compared with that of its 26 neighboring points (8 pixels at the same scale + 9 pixels at each of the adjacent scales above and below). Local extrema across scales are selected. These extrema initially correspond to potential feature points with scale recognition in accident scenes, such as vehicle corners, road marking inflection points, and debris edges. This lays the foundation for accurate extraction of multi-scale data and adapts to the image scale differences brought about by multi-view drone shooting.

[0039] Step S203: For the potential feature points selected in step S202, their coordinates in scale space and image space are fitted using a quadratic function. Low-quality points with contrast below a preset threshold (set in conjunction with the accuracy requirements of traffic accident mapping to ensure the stability of feature points) are eliminated. Then, the principal curvature ratio of each potential feature point is calculated using the Hessian matrix, and edge response points with a principal curvature ratio > 10 are eliminated (these points are mostly transition areas between the background and the feature, and have no actual feature significance). The final stable key points are precisely distributed at key positions on the contour of the accident feature (such as the edge of a vehicle collision dent, the corner of scattered objects), and the coordinates of each key point can be associated with specific traffic accident features through the correspondence table in step S201.

[0040] Step S204: Using the stable keypoints obtained in step S203 as the center, a 16×16 pixel neighborhood is selected based on the scale value corresponding to the keypoint. The larger the scale value, the wider the actual scene area covered by the neighborhood, ensuring the consistency of neighborhood information at different scales. The Sobel operator is used to calculate the gradient magnitude and direction of each pixel in this neighborhood, generating a gradient direction histogram of 36 equal intervals (bins) (each equal interval corresponds to 10°). The direction corresponding to the peak of the histogram is taken as the main direction of the keypoint. If the secondary peak exceeds 80% of the main peak, an auxiliary direction is added to the keypoint. By assigning direction values, each keypoint has rotation invariance, adapting to image rotation scenarios caused by drone tilt shooting, and ensuring that the direction features of keypoints for the same element are consistent in images from different viewpoints.

[0041] Step S205: Based on the key points assigned directions in step S204, rotate the neighborhood around the key points according to their main directions (eliminating the effect of rotation), reducing the 16×16 pixel descriptor neighborhood range of the traditional SIFT algorithm to 12×12 pixels. Divide the reduced neighborhood into 4×4 sub-blocks (each sub-block corresponds to 3×3 pixels), calculate the gradient histograms of 8 directions (each direction corresponds to 45°) within each sub-block, and integrate them to generate a 128-dimensional feature descriptor. Reducing the neighborhood can reduce the interference of background noise (such as road stains, distant trees) on the feature features, and more accurately depict local details such as scattered textures and vehicle scratches; at the same time, the 128-dimensional descriptor maintains feature recognition and reduces the computational cost compared to the traditional SIFT. Each descriptor is bound to the corresponding key point, and can be associated with specific traffic accident elements through the key point.

[0042] Step S206: Perform L2 normalization on the 128-dimensional feature descriptors generated in step S205: Calculate the L2 norm of each descriptor, divide each element in the descriptor by this norm, and eliminate the gradient magnitude differences caused by different lighting conditions at the accident scene (such as backlighting and cloudy days). The normalized descriptors ensure that the feature representation of the same traffic accident element is consistent in images under different lighting conditions, providing a unified data standard for subsequent cross-view feature matching. Furthermore, each normalized descriptor remains associated with its corresponding key point and traffic accident element, preparing data for the final determination of feature point pairs for element matching.

[0043] Step S30: Call the preset openMVG tool to triangulate the matching feature point pairs to generate sparse point clouds corresponding to each traffic accident element, and use the preset PMVS algorithm to expand and filter the sparse point clouds to obtain dense point clouds corresponding to each traffic accident element. An improved SIFT algorithm is used to narrow the neighborhood range of the descriptor. Feature extraction and matching are performed on the traffic accident scene image after thinning and optimization. After determining the matching feature point pairs corresponding to each traffic accident element in the traffic accident scene image, the preset openMVG tool is called to triangulate the matching feature point pairs to generate sparse point clouds corresponding to each traffic accident element. The preset PMVS algorithm is used to expand and filter the sparse point clouds to obtain dense point clouds corresponding to each traffic accident element. In some embodiments, after using a preset PMVS algorithm to expand and filter the sparse point cloud to obtain the dense point cloud corresponding to each traffic accident element, a triangular mesh can be constructed based on the Delaunay criterion and growth method; a square regular mesh is used for surface rendering to eliminate cracks and improve the visual effect.

[0044] In some embodiments, the steps of calling a preset openMVG tool to triangulate the matching feature point pairs to generate sparse point clouds corresponding to each traffic accident element, and using a preset PMVS algorithm to perform patch expansion and filtering on the sparse point clouds to obtain dense point clouds corresponding to each traffic accident element include: Step S301: Obtain matching feature point pairs corresponding to each traffic accident element in the traffic accident scene image, and simultaneously retrieve the camera intrinsic parameters and exterior orientation elements when the traffic accident scene image was acquired. The camera intrinsic parameters include focal length and principal point coordinates, and the exterior orientation elements include RTK positioning coordinates and IMU attitude angles. Step S302: Call the preset openMVG tool, input the matching feature point pairs and camera intrinsic and extrinsic orientation elements into the openMVG tool, optimize the camera attitude through bundle adjustment, perform triangulation calculation, and generate sparse point clouds corresponding to each traffic accident element. Step S303: Import the sparse point cloud corresponding to each traffic accident element into the preset PMVS algorithm, use the points of the sparse point cloud as seed points to construct the initial patch, expand the patch range based on the texture consistency constraint of the traffic accident scene image, and remove abnormal patches that deviate from the contour of the corresponding traffic accident element. Step S304: Perform point cloudification on the expanded and filtered surface patches to obtain dense point clouds corresponding to each traffic accident element.

[0045] As can be seen from steps S301 to S304 above, synchronously acquiring matching feature point pairs and camera parameters clarifies the correspondence between feature elements and provides a basis for mapping two-dimensional pixels to spatial coordinates through intrinsic parameters (focal length, principal point) and exterior orientation elements (RTK coordinates, IMU attitude angles). This avoids point cloud position deviations caused by missing parameters during subsequent triangulation, laying a data foundation for accurate modeling. Based on the OpenMVG tool, camera attitude is optimized and triangulation is performed. First, shooting errors are corrected through bundle adjustment, and then three-dimensional coordinates are calculated based on matching point pairs to generate a feature-specific sparse point cloud. This preserves the core spatial structure of the features while eliminating redundancy, providing efficient and accurate seed data for subsequent dense reconstruction, adapting to the multi-feature separation modeling needs of accident scenarios.

[0046] Initial patches are constructed using sparse points as seeds. These patches are then expanded and anomalous patches are removed based on texture consistency constraints. This approach leverages image textures to complete feature details while avoiding background interference, ensuring patches closely match feature contours (such as vehicle shapes and debris boundaries). This improves the geometric accuracy and feature correlation of subsequent dense point clouds. The optimized patches are then converted into dense point clouds, significantly increasing point cloud density (e.g., over 50 points per square meter). This fully restores feature details (such as collision dents and marking widths) while maintaining the independence of each feature's point cloud. This provides high-quality geometric data for subsequent NeRF model training and triangular mesh construction, supporting accurate 3D reconstruction of accident scenes.

[0047] Step S40: Use the dense point cloud corresponding to each traffic accident element as a geometric constraint, use the thinned and optimized traffic accident scene image as a texture data source to construct a training dataset, and use the training dataset to train the preset neural radiation field model to convergence, so as to determine the traffic accident scene 3D model that has been trained to convergence. The pre-defined OpenMVG tool is invoked to triangulate the matching feature point pairs to generate sparse point clouds corresponding to each traffic accident element. After the pre-defined PMVS algorithm is used to expand and filter the sparse point clouds to obtain dense point clouds corresponding to each traffic accident element, the dense point clouds corresponding to each traffic accident element are used as geometric constraints. The thinned and optimized traffic accident scene image is used as a texture data source to construct a training dataset. The pre-defined neural radiation field model is trained to convergence using the training dataset to determine the 3D model of the traffic accident scene that has been trained to convergence.

[0048] In some embodiments, prior to the step of inputting the traffic accident scene image to be reconstructed into the traffic accident scene 3D model that has been trained to a convergent state, the method includes: Step S401: Obtain the traffic accident location coordinates of the target traffic accident scene and the flight constraints of the UAV flight environment, wherein the flight constraints include no-fly zone constraints, battery life constraints, flight parameter constraints, static obstacle constraints, and dynamic obstacle constraints. Step S402: Based on the flight constraints of the UAV flight environment, construct the objective function of the flight control model to minimize the total length of the flight path, maximize the minimum safe distance between the flight path and obstacles, and maximize the smoothness of the flight path. Solve the objective function of the flight control model based on the preset particle swarm optimization algorithm to determine the optimal evidence collection path parameters of the UAV. Step S403: Control the drone to fly to the traffic accident location coordinates of the target traffic accident scene according to the optimal evidence collection path parameters, so as to obtain the traffic accident scene image to be reconstructed at the target traffic accident scene.

[0049] In some embodiments, the step of constructing an objective function for a flight control model based on flight constraints of the UAV's flight environment—minimizing the total length of the flight path, maximizing the minimum safe distance between the flight path and obstacles, and maximizing the smoothness of the flight path—and solving the objective function of the flight control model using a preset particle swarm optimization algorithm to determine the optimal evidence-gathering path parameters for the UAV includes: Step S4001: Based on the flight constraints of the UAV flight environment, construct the objective function of the flight control model with the optimization objectives of minimizing the total length of the flight path, maximizing the minimum safe distance between the flight path and obstacles, and maximizing the smoothness of the flight path. The total length of the path represents the sum of the Euclidean distances between each discrete path point in the flight path, and the minimum safe distance represents the minimum distance from each discrete path point in the flight path to the boundary of the obstacle. Step S4002: Initialize each initial particle in the initial population of the particle swarm optimization algorithm, wherein each particle represents the evidence path parameters of the candidate evidence path corresponding to the UAV. Step S4003: Generate the velocity and position of each initial particle within the flight constraint conditions, calculate the objective function value of the evidence path parameters of the candidate evidence path corresponding to each particle according to the objective function of the flight control model, so as to determine the individual optimal solution and the global optimal solution, and iteratively update the particle velocity and particle position. Step S4004: Repeat step S4003 until the number of iterations reaches a preset iteration threshold and then stop iterating, in order to determine the optimal evidence collection path parameters corresponding to the global optimal solution. The optimal evidence collection path parameters include the three-dimensional coordinates of discrete path points and the flight speed between adjacent path points.

[0050] As illustrated in the above embodiments, obtaining the accident location coordinates and flight constraints clarifies the spatial objectives and limiting boundaries for UAV evidence collection. Constraints such as no-fly zones and battery limitations prevent unauthorized flights, while static and dynamic obstacle constraints proactively mitigate collision risks, ensuring path planning remains relevant to the actual scenario and guaranteeing the safety and compliance of evidence collection. The objective function of the flight control model is constructed by minimizing the total flight path length, maximizing the minimum safe distance between the flight path and obstacles, and maximizing flight path smoothness. The optimal path is solved using a particle swarm optimization algorithm. This minimizes path length to save battery life and improve evidence collection efficiency, maximizes the safe distance from obstacles to reduce collision risks, and ensures a smooth path to reduce UAV flight jitter (avoiding image blurring). The output optimal evidence collection path parameters are precise and controllable, providing technical support for efficient and safe evidence collection. Controlling the UAV flight for evidence collection according to the optimal path ensures the UAV accurately arrives at the accident scene, avoiding missing key areas due to path deviations, while stabilizing the flight state improves the quality of acquired images (reducing blurring and misalignment). The acquired images to be reconstructed fully cover the accident elements, providing high-quality data for subsequent 3D model input, ensuring the accuracy of real-scene reconstruction from the source, while shortening the time spent at the scene and reducing traffic congestion and the risk of secondary accidents.

[0051] Step S50: Input the traffic accident scene image to be reconstructed into the traffic accident scene 3D model that has been trained to convergence state, so as to extract the 3D spatial information corresponding to each traffic accident element in the traffic accident scene image to complete the real-scene reconstruction of the traffic accident scene.

[0052] The dense point cloud corresponding to each traffic accident element is used as a geometric constraint, and the thinned and optimized traffic accident scene image is used as a texture data source to construct a training dataset. The preset neural radiation field model is trained to convergence using the training dataset to determine the traffic accident scene 3D model that has been trained to convergence. The traffic accident scene image to be reconstructed is input into the traffic accident scene 3D model that has been trained to convergence to extract the 3D spatial information corresponding to each traffic accident element in the traffic accident scene image to complete the real-scene reconstruction of the traffic accident scene.

[0053] In some embodiments, after the step of inputting the traffic accident scene image to be reconstructed into the traffic accident scene 3D model that has been trained to a converged state to extract the 3D spatial information corresponding to each traffic accident element in the traffic accident scene image to be reconstructed, the method includes: Law enforcement officers determine liability at the target traffic accident scene based on the spatial information corresponding to each traffic accident element in the image of the traffic accident scene to be reconstructed, thereby completing the real-scene reconstruction of the traffic accident scene.

[0054] Specifically, law enforcement officers can use traffic management platforms to retrieve 3D spatial information of traffic accident elements extracted from 3D models of traffic accident scenes. This includes the 3D coordinates of the vehicles involved (such as their position at the time of collision and their orientation), relative positional relationships (such as the collision angle between the two vehicles and their offset from road markings), and key detail dimensions (such as braking distance and the distribution distance between debris and vehicles). This data comes from the precise calculations of the 3D model of the traffic accident scene, which has been trained to convergence, to reconstruct the image of the traffic accident scene. Law enforcement officers can combine the 3D spatial information of the traffic accident elements with traffic regulations (such as determining whether vehicles crossed the line or failed to maintain a safe distance) to reconstruct the accident process (such as estimating vehicle speed through braking distance and determining the responsible party through the collision angle), avoiding the bias caused by traditional manual investigations relying on subjective experience. For example, if the 3D model of the traffic accident scene shows that vehicle A's braking distance exceeds the prescribed range, and vehicle B did not cross the line, it can help determine the primary responsibility of vehicle A for not slowing down in time. Ultimately, by combining spatial information with legal provisions, law enforcement officers can generate a liability determination document. At the same time, the 3D model of the traffic accident scene can be archived as a result of real-scene reconstruction, and can be retrieved retrospectively if there are any disputes. This not only completes the visual real-scene restoration of the accident scene, but also provides objective and accurate technical support for liability determination, realizing the empowerment of law enforcement by technology and improving the efficiency and credibility of accident handling.

[0055] As can be seen from the above embodiments, compared with the prior art, this application addresses the problems in the prior art, such as the long time required for law enforcement officers to conduct accident scene surveys, the poor accuracy of manual surveys, and the need for various complex manual interventions in the drawing of traditional scene survey maps, with almost every step requiring manual participation, which is time-consuming and labor-intensive. This application has, but is not limited to, the following beneficial effects: Firstly, this application achieves end-to-end automated processing from UAV path planning and image acquisition to image thinning, feature matching, point cloud generation, and 3D model training and reconstruction, which significantly shortens the accident scene investigation and modeling cycle, reduces road closure time, and lowers the possibility of secondary accidents.

[0056] Secondly, this application improves matching accuracy through an enhanced SIFT algorithm, combines point cloud generation and optimization algorithms to complete the geometric information of elements, and then uses 3D modeling technology to restore details such as the shape of the accident vehicle and the distribution of scattered objects, ensuring the geometric accuracy and texture realism of the model and meeting the needs of precise surveying and accident analysis. This application avoids the subjective bias of traditional manual surveying, and data can be retrieved retrospectively for subsequent disputes over liability determination, providing objective evidence for law enforcement and enhancing public credibility.

[0057] Thirdly, this application reduces on-site safety risks. The drone remotely completes image acquisition, eliminating the need for law enforcement officers to stay at the accident scene for extended periods and avoiding safety hazards associated with investigations in traffic. The rapid completion of on-site processing also indirectly reduces additional safety and economic problems caused by traffic congestion.

[0058] Fourth, the three-dimensional model of the traffic accident scene constructed in this application, the objective spatial information provided by the three-dimensional model, combined with traffic regulations, can quickly clarify the attribution of accident responsibility, reduce disputes caused by insufficient evidence, and improve the efficiency and fairness of accident handling.

[0059] Please see Figure 2This application provides a traffic accident scene reconstruction device, comprising an image processing module 1100, a feature point pair determination module 1200, a dense point cloud determination module 1300, a 3D model construction module 1400, and a traffic accident reconstruction module 1500. The image processing module 1100 is configured to acquire multi-view traffic accident scene images corresponding to various traffic accident scenes, and to perform thinning optimization on the traffic accident scene images using a preset Douglas-Puk algorithm to generate thinned and optimized traffic accident scene images. The feature point pair determination module 1200 is configured to use an improved SIFT algorithm to narrow the descriptor neighborhood range, and to perform feature extraction and matching on the thinned and optimized traffic accident scene images to determine the matching feature point pairs corresponding to each traffic accident element in the traffic accident scene images. The dense point cloud determination module 1300 is configured to call a preset OpenMVG tool to triangulate the matching feature point pairs to generate sparse point clouds corresponding to each traffic accident element, and to use a preset PMVS algorithm. The sparse point cloud is expanded and filtered to obtain a dense point cloud corresponding to each traffic accident element. The 3D model construction module 1400 is configured to use the dense point cloud corresponding to each traffic accident element as a geometric constraint and the thinned and optimized traffic accident scene image as a texture data source to construct a training dataset. The training dataset is used to train a preset neural radiation field model to convergence, so as to determine the traffic accident scene 3D model that has been trained to convergence. The traffic accident reconstruction module 1500 is configured to reconstruct the traffic accident scene image to be reconstructed by inputting it into the traffic accident scene 3D model that has been trained to convergence, so as to extract the 3D spatial information corresponding to each traffic accident element in the traffic accident scene image to be reconstructed, so as to complete the real scene reconstruction of the traffic accident scene.

[0060] Based on any embodiment of this application, please refer to Figure 3 Another embodiment of this application also provides an electronic device, which can be implemented by a computer device, such as... Figure 3The diagram shows the internal structure of a computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable storage medium stores an operating system, a database, and computer-readable instructions. The database may store a sequence of control information. When the computer-readable instructions are executed by the processor, the processor can implement a method for real-time reconstruction of a traffic accident scene. The processor of the computer device provides computing and control capabilities to support the operation of the entire computer device. The memory of the computer device may store computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor can execute the traffic accident scene reconstruction method of this application. The network interface of the computer device is used for communication with a terminal. Those skilled in the art will understand that… Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0061] In this embodiment, the processor is used to execute... Figure 2 The specific functions of each module are defined within the device, and the memory stores the program code and various data required to execute these modules. A network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program code and data required to execute all modules in the traffic accident scene reconstruction device of this application, and the server can call the server's program code and data to execute the functions of all modules.

[0062] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the traffic accident scene reconstruction method described in any embodiment of this application.

[0063] This application also provides a computer program product, including a computer program / instructions that, when executed by one or more processors, implement the steps of the traffic accident scene reconstruction method described in any embodiment of this application.

[0064] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0065] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for real-scene reconstruction of a traffic accident scene, characterized in that, include: The system acquires multi-view traffic accident scene images corresponding to each traffic accident scene, and uses a preset Douglas-Puk algorithm to thin and optimize the traffic accident scene images to generate thinned and optimized traffic accident scene images. An improved SIFT algorithm is used to narrow the neighborhood range of the descriptor, and features are extracted and matched on the thinned and optimized traffic accident scene image to determine the matching feature point pairs corresponding to each traffic accident element in the traffic accident scene image. The preset openMVG tool is called to triangulate the matching feature point pairs to generate sparse point clouds corresponding to each traffic accident element. The preset PMVS algorithm is used to expand and filter the sparse point clouds to obtain dense point clouds corresponding to each traffic accident element. The dense point cloud corresponding to each traffic accident element is used as a geometric constraint, and the thinned and optimized traffic accident scene image is used as a texture data source to construct a training dataset. The training dataset is used to train the preset neural radiation field model to convergence, so as to determine the traffic accident scene 3D model that has been trained to convergence. The traffic accident scene image to be reconstructed is input into the traffic accident scene 3D model that has been trained to convergence, so as to extract the 3D spatial information corresponding to each traffic accident element in the traffic accident scene image to complete the real-scene reconstruction of the traffic accident scene.

2. The method for real-scene reconstruction of traffic accident scenes according to claim 1, characterized in that, The elements of the traffic accident include the vehicles involved, road facilities, debris, accident traces, and persons involved. The vehicles involved include motor vehicles and non-motorized vehicles; the road facilities include lane markings, crosswalks, directional arrows, curbs, guardrails, traffic lights, traffic signs, and speed bumps; the debris includes parts detached from the vehicles; the accident traces include tire brake marks, bloodstains, and scattered goods or personal belongings; and the persons involved include drivers, passengers, pedestrians, and non-motorized vehicle riders. The three-dimensional spatial information includes the three-dimensional coordinates, relative positional relationships, shape and size, and collision-related details of each traffic accident element.

3. The method for real-scene reconstruction of traffic accident scenes according to claim 1, characterized in that, Before the step of inputting the image of the traffic accident scene to be reconstructed into the 3D model of the traffic accident scene that has been trained to convergence, the following steps are included: The system obtains the coordinates of the traffic accident location at the target traffic accident scene and the flight constraints of the UAV's flight environment. The flight constraints include no-fly zone constraints, battery life constraints, flight parameter constraints, static obstacle constraints, and dynamic obstacle constraints. Based on the flight constraints of the UAV's flight environment, the objective function of the flight control model is constructed to minimize the total length of the flight path, maximize the minimum safe distance between the flight path and obstacles, and maximize the smoothness of the flight path. The objective function of the flight control model is solved based on a preset particle swarm optimization algorithm to determine the optimal evidence collection path parameters of the UAV. The drone is controlled to fly to the traffic accident location coordinates of the target traffic accident scene according to the optimal evidence collection path parameters, so as to obtain the traffic accident scene image to be reconstructed at the target traffic accident scene.

4. The method for real-scene reconstruction of traffic accident scenes according to claim 3, characterized in that, Based on the flight constraints of the UAV's flight environment, the objective function of the flight control model is constructed to minimize the total flight path length, maximize the minimum safe distance between the flight path and obstacles, and maximize the flight path smoothness. The objective function of the flight control model is solved using a preset particle swarm optimization algorithm to determine the optimal evidence collection path parameters for the UAV. The steps include: Based on the flight constraints of the UAV flight environment, an objective function for the flight control model is constructed with the optimization objectives of minimizing the total length of the flight path, maximizing the minimum safe distance between the flight path and obstacles, and maximizing the smoothness of the flight path. The total length of the path represents the sum of the Euclidean distances between each discrete path point in the flight path, and the minimum safe distance represents the minimum distance from each discrete path point in the flight path to the boundary of the obstacle. Initialize each initial particle in the initial population of the particle swarm optimization algorithm, where each particle represents the evidence path parameters of the candidate evidence path corresponding to the UAV. Within the flight constraints, the velocity and position of each initial particle are generated. Based on the objective function of the flight control model, the objective function value of the evidence path parameters of the candidate evidence path corresponding to each particle is calculated to determine the individual optimal solution and the global optimal solution. The particle velocity and particle position are iteratively updated. Repeat the above steps until the number of iterations reaches a preset iteration threshold, then stop the iteration to determine the optimal evidence collection path parameters corresponding to the global optimal solution. The optimal evidence collection path parameters include the three-dimensional coordinates of discrete path points and the flight speed between adjacent path points.

5. The method for real-scene reconstruction of traffic accident scenes according to claim 1, characterized in that, The step of using a preset Douglas-Puk algorithm to thin and optimize the traffic accident scene image to generate a thinned and optimized traffic accident scene image includes: The process involves acquiring traffic accident scene images captured by drones, extracting the contour feature point set of each traffic accident element in the traffic accident scene image using an edge detection algorithm, converting the contour feature point set into a two-dimensional point sequence according to pixel coordinates, and removing duplicate and abnormal coordinate noise points in the two-dimensional point sequence. Based on the preset distance threshold corresponding to each traffic accident element of the Douglas-Puk algorithm, the first and last two points of the two-dimensional point sequence corresponding to each traffic accident element are used as the initial baseline segment. The vertical distance from all intermediate feature points in the two-dimensional point sequence to the baseline segment is calculated, and the maximum vertical distance value and the corresponding feature point are selected. Each traffic accident element corresponds one-to-one with each distance threshold, and the distance threshold represents the maximum allowable offset distance from the feature point to the baseline segment. If the maximum vertical distance value is greater than the distance threshold, the feature point corresponding to the maximum vertical distance value is retained, and the two-dimensional point sequence is split into two sub-sequences using it as the dividing point. The vertical distance calculation operation is repeated for the two sub-sequences respectively. If the maximum vertical distance value is less than or equal to the distance threshold, all intermediate points of the current sub-sequence are discarded, and only the first and last points of the sub-sequence are retained. All retained feature points are obtained and reconnected according to the spatial order of the original two-dimensional point sequence to form the thinned traffic accident feature outline. This traffic accident feature outline is then merged with the background layer of the traffic accident scene image to generate the thinned and optimized traffic accident scene image.

6. The method for real-scene reconstruction of traffic accident scenes according to claim 1, characterized in that, The steps of calling the preset OpenMVG tool to triangulate the matched feature point pairs to generate sparse point clouds corresponding to each traffic accident element, and using the preset PMVS algorithm to perform patch expansion and filtering on the sparse point clouds to obtain dense point clouds corresponding to each traffic accident element include: The matching feature point pairs corresponding to each traffic accident element in the traffic accident scene image are obtained, and the camera intrinsic parameters and exterior orientation elements at the time of the traffic accident scene image acquisition are retrieved simultaneously. The camera intrinsic parameters include focal length and principal point coordinates, and the exterior orientation elements include RTK positioning coordinates and IMU attitude angles. The preset openMVG tool is invoked, and the matching feature point pairs, camera intrinsic parameters, and exterior orientation elements are input into the openMVG tool. After optimizing the camera attitude through bundle adjustment, triangulation calculation is performed to generate sparse point clouds corresponding to each traffic accident element. The sparse point clouds corresponding to each traffic accident element are imported into the preset PMVS algorithm. The points of the sparse point cloud are used as seed points to construct initial patches. The patch range is expanded based on the texture consistency constraint of the traffic accident scene image, while abnormal patches that deviate from the contour of the corresponding traffic accident element are removed. The expanded and filtered patches are processed into point clouds to obtain dense point clouds corresponding to each traffic accident element.

7. The method for real-scene reconstruction of traffic accident scenes according to any one of claims 1 to 6, characterized in that, After the step of inputting the traffic accident scene image to be reconstructed into the pre-trained converged 3D model of the traffic accident scene to extract the corresponding 3D spatial information of each traffic accident element in the traffic accident scene image, the following steps are included: Law enforcement officers determine liability at the target traffic accident scene based on the spatial information corresponding to each traffic accident element in the image of the traffic accident scene to be reconstructed, thereby completing the real-scene reconstruction of the traffic accident scene.

8. A device for real-scene reconstruction of traffic accident scenes, characterized in that, include: The image processing module is configured to acquire multi-view traffic accident scene images corresponding to each traffic accident scene, and use a preset Douglas-Puk algorithm to thin and optimize the traffic accident scene images to generate thinned and optimized traffic accident scene images. The feature point pair determination module is configured to use an improved SIFT algorithm to narrow the descriptor neighborhood range, and to perform feature extraction and matching on the thinned and optimized traffic accident scene image in order to determine the matching feature point pairs corresponding to each traffic accident element in the traffic accident scene image. The dense point cloud determination module is configured to call the preset openMVG tool to triangulate the matching feature point pairs to generate sparse point clouds corresponding to each traffic accident element, and use the preset PMVS algorithm to expand and filter the sparse point clouds to obtain the dense point clouds corresponding to each traffic accident element. The 3D model construction module is configured to use the dense point cloud corresponding to each traffic accident element as a geometric constraint, and the thinned and optimized traffic accident scene image as a texture data source to construct a training dataset. The training dataset is used to train the preset neural radiation field model to convergence, so as to determine the 3D model of the traffic accident scene that has been trained to convergence. The traffic accident reconstruction module is configured to input the traffic accident scene image to be reconstructed into the traffic accident scene 3D model that has been trained to convergence, so as to extract the 3D spatial information corresponding to each traffic accident element in the traffic accident scene image to complete the real-scene reconstruction of the traffic accident scene.

9. An electronic device comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 7, which, when invoked by a computer, executes the steps included in the corresponding method.

Images