A traffic accident scene real scene three-dimensional modeling drawing method and system

Abstract

The application provides a traffic accident scene real scene three-dimensional modeling and drawing method and system. Based on the characteristics of wide coverage and high collection efficiency of panoramic video, the accident scene data is collected comprehensively and quickly. Combined with an optimized panoramic video segmentation method, motion structure recovery and 3D technology, a three-dimensional model of the accident scene with better visual effect and higher precision is constructed. Meanwhile, a standardized calibration, measurement and drawing whole process adapting to the standard system of traffic accident scene drawing is built, and a compliant scene real scene drawing and scene record drawing are generated.

Description

Technical Field

[0001] This invention relates to the field of traffic accident scene investigation technology, specifically to a method and system for real-world 3D modeling and drawing of traffic accident scenes. Background Technology

[0002] According to the "Regulations on Procedures for Handling Road Traffic Accidents" and related work specifications and technical standards, on-site investigation of traffic accidents requires the preservation of on-site evidence through methods such as drawing on-site diagrams, taking on-site photographs, and preparing on-site investigation records. Among these, on-site diagrams can visually reflect the on-site environment, the form of the accident, and the location and interrelationships of relevant vehicles, personnel, objects, and traces, and are an important type of evidentiary material for analyzing the causes of the accident and determining liability.

[0003] Traffic accident scene diagrams possess unique technical characteristics that distinguish them from other fields such as engineering drawing, exhibiting independent and strict industry standards: First, they are specific in their drawing types. According to "Drawing Road Traffic Accident Scene Diagrams" (GA / T 49-2019), traffic accident scene diagrams include six categories: record diagrams, real-scene diagrams, scale diagrams, cross-sectional diagrams, elevation diagrams, and analysis diagrams. Among these, real-scene diagrams are a unique drawing type for traffic accident investigation, relying on real-scene images or 3D models for annotation and drawing. Second, they are specific in their graphic symbols. According to "Graphic Symbols for Road Traffic Accident Scenes" (GB / T11797-2005), standardized graphic symbols for traffic accident scenes are specified, including those for motor vehicles, non-motor vehicles, pedestrians, brake marks, road markings, scattered objects, roadside facilities, bloodstains, and scratches. Third, they are specific in their positioning methods. Traffic accident scene diagrams require the use of specific investigation and positioning methods such as right-angle positioning, comprehensive positioning, and triangulation positioning, relying on benchmark points and baselines to complete the precise positioning and measurement of scene elements, ensuring the legality and accuracy of the drawing data.

[0004] Traditional road traffic accident scene mapping relies on measuring tapes and manual drawing, resulting in low efficiency, potential traffic congestion and secondary accidents, and problems such as large measurement errors and non-standard drawing. Drone aerial mapping technology improves the efficiency of large-scale scene mapping and is suitable for open scenes such as large urban intersections and highways. However, it cannot be used in areas with tree cover, severe weather conditions such as wind, rain, and snow, or in restricted flight areas such as around airports, limiting its application scenarios. In recent years, real-scene 3D modeling technology has developed rapidly, becoming a new direction for the digital and intelligent investigation of traffic accident scenes. Existing lightweight laser scanning equipment integrates multiple sensors such as LiDAR, visual acquisition, and high-precision positioning, which can quickly collect point cloud and image data from the scene to generate high-precision 3D models. However, the purchase and maintenance costs of such equipment are high, and the professional skills required for operation are also high, making it difficult to promote its widespread application at the grassroots level in the short term.

[0005] Multi-view geometry-based real-scene 3D modeling technology has become a research hotspot in the industry due to its low cost and ease of operation. This technology uses ordinary mobile phones and cameras to capture on-site videos or image sequences, combines motion structure recovery algorithms to calculate camera pose and on-site spatial structure, and then uses 3D Gaussian splash rendering technology to quickly generate 3D models of accident scenes, significantly lowering the equipment threshold for 3D modeling. However, existing technologies still have two major shortcomings in standardized traffic accident mapping scenarios: First, traditional single-view surround shooting only focuses on the vehicle at the center of the accident, easily resulting in blurred images and missing details of surrounding roads, minor road surface marks, and edge markings. This leads to poor integrity of the generated 3D model, failing to meet the detail requirements of real-scene mapping. To increase model integrity and detail clarity, it is necessary to increase the shooting angle and expand the shooting path range, inevitably increasing shooting time and affecting the efficiency of on-site investigation and handling. Second, existing technologies only focus on generating, rendering, and simple ranging of 3D models, lacking a mapping process adapted to the standard system for traffic accident scene mapping. This prevents the direct generation of compliant on-site maps, hindering the full realization of the application value of 3D models and severely limiting the practical application of real-scene 3D technology in accident scene investigation. Summary of the Invention

[0006] To address the issues of insufficient model detail and lack of standardized drawing processes when applying existing real-scene 3D modeling techniques to traffic accident scenes, this invention provides a real-scene 3D modeling and drawing method for traffic accident scenes. This method can construct 3D accident scene models with better visual appeal and higher accuracy, while significantly shortening the time required for on-site investigation, reducing equipment and operational barriers, and improving the accuracy and standardization of on-site drawing.

[0007] The technical solution of this invention is as follows: a method for three-dimensional modeling and drawing of a traffic accident scene, characterized by comprising the following steps: S1: Fast panoramic video acquisition; Determine the closed-loop acquisition path based on the scope of the accident; A panoramic camera is used to collect full-area video of the traffic accident scene along the closed-loop acquisition path. Two circles of panoramic video are collected around the accident scene at different heights. The first circle is collected at a height higher than the top of the accident vehicle to capture the overall environment of the scene, and the second circle is collected at a height lower than the height of the operator to capture the fine elements of the road surface, thus obtaining a complete video data source covering the overall layout of the scene and the detailed elements of the road surface. S2: Refined 3D modeling, including the following steps: Panoramic video perspective segmentation: Differentiated perspective segmentation is performed on the acquired panoramic video, the spherical image is converted into a standard perspective image, and keyframes are extracted in batches at fixed time intervals to construct a continuous and complete sequence of on-site image frames; specifically, the core accident area is segmented with a small angle of 45°~60° for fine segmentation, and the road background area is segmented with a large angle of 90°~120° for coarse segmentation, generating an image frame sequence in which adjacent perspectives retain 15%~20% of the feature overlap area; Based on the image frame sequence, a motion structure recovery algorithm is used to calculate the camera pose and reconstruct the three-dimensional spatial structure to build a real-world three-dimensional model of the accident scene. S3: Standard site diagram drawing; Based on the real-scene 3D model, the computer performs model scale calibration, benchmark point and baseline setting, positioning and distance measurement and on-site element annotation. After generating the annotated 3D model scene view, the national standard graphic symbol library is called to draw graphic symbols and nest standard templates to generate a real-scene picture and on-site record picture that conform to the national standard specification for traffic accident scene drawing.

[0008] Its further features are: In step S1, the method for confirming the closed-loop collection path is as follows: taking the area composed of core accident elements as the center, and combining the on-site road direction, lane distribution, and roadside facility layout, a closed-loop collection path is planned; so that the collection path fully covers the core accident area, lanes, intersection markings, roadside facilities, and vegetation-covered areas; the core accident elements include: accident vehicles, collision marks, human bodies, scattered objects, and bloodstains and scratches. Based on the scale of the scene covered by the closed-loop acquisition path, the acquisition time is adaptively matched, with shorter acquisition time for smaller scales; the scene scale includes: large intersections or highway accident scenes as well as small accident scenes. In step S2, a three-dimensional model of the accident scene is constructed based on the image frame sequence, specifically including the following steps: a1: Camera pose and spatial structure reconstruction: The motion structure recovery algorithm is used to extract and accurately match feature points in key frames, solve the camera intrinsic parameters, extrinsic parameters and 3D spatial pose of each frame, restore the overall 3D spatial topology of the accident scene, and generate a highly matched sparse point cloud model of the scene. a2: Dense point cloud generation and optimization: Global dense reconstruction is carried out based on sparse point cloud model to complete the subtle spatial and texture information of the scene, filter the floating noise of the scene, and fully restore the small evidence; the small evidence includes: road scratches, scattered objects and road markings. a3: 3D model rendering: Convert high-precision dense point clouds into a visual 3D model using a 3D Gaussian algorithm to generate a real-world 3D model of the accident scene; Step S3 specifically includes the following steps: S31: Model display and trimming / editing; Load and render the pre-processed 3D model of the traffic accident scene, and perform custom trimming and editing. The custom trimming and editing includes two modes: planar trimming and 3D trimming. Planar trimming involves customizing the cutting plane to remove invalid model data from one side of the plane. 3D trimming involves customizing the effective area of ​​a cube, retaining the core accident model data inside the cube, removing irrelevant external environmental data, and simplifying the model scene. S32: Model calibration measurement; Establish a standardized process, and complete model calibration and precise measurement based on the standardized process; The standard drawing process includes the following steps executed in sequence: setting dimension reference, setting ground reference, setting benchmark point, setting benchmark line, selecting the point to be measured, selecting positioning method, and annotating on-site elements. The setting of the size reference is as follows: Select a size reference object in the 3D model, set measurement points at both ends of the reference object, and use the known standard length of the reference object as a benchmark to complete the scale calibration between the model space and the actual physical space, ensuring that the model size restores the actual size on site in a 1:1 ratio; if the model already has absolute geographic coordinates and precise scale, this step can be skipped. The ground reference setting is as follows: Select a flat, undulating standard road surface area at the accident site, set three non-collinear measurement points, and use the three points to generate a standard ground reference plane to unify the model's ground benchmark and eliminate measurement errors caused by road surface undulations; if the model has accurate ground coordinate information, this step can be skipped. The established reference points are: the reference carriers of the three-dimensional model include: permanent fixed objects on site and temporary references; If there are permanent fixed objects on site, select the permanent fixed objects as the reference carrier, and set measuring points at the grounding point, end point and center point of the permanent fixed objects as reference points; otherwise, if there are no permanent fixed objects on site, use temporary reference points, and record the relative distance between the temporary reference points and the surrounding fixed objects and reference lines at the same time to ensure the traceability of the positioning. The baseline is defined as follows: a baseline is constructed by selecting orthogonal elements in the model; The selection of test points involves: selecting measurement points for various accident elements based on the investigation standards; the accident elements include: vehicle elements, human elements, road surface traces, and scattered object evidence. The selected positioning method is as follows: based on the characteristics of the accident scenario, a suitable method for accurate positioning and measurement of on-site elements is selected; for regular and orderly road scenarios, the right-angle positioning method or the comprehensive positioning method is preferred; for scenarios without regular baselines, complex intersections, or non-standard roads, the triangulation positioning method is used. The element labeling refers to: three-dimensional spatial line labeling and text annotation of accident vehicles, road traces, and on-site evidence to fully record accident element information; the on-site elements corresponding to the accident vehicles include: vehicle number, vehicle type, and license plate information; the on-site elements corresponding to the road traces include: brake marks, scratches, and skid marks; the on-site elements corresponding to the on-site evidence include: glass fragments, paint fragments, bloodstains, and scattered debris. S33: Graphic symbol creation; Based on the 3D model completed by calibration measurement, complete the standardized graphic symbol drawing to meet the standard requirements for traffic accident scene drawing; S34: On-site diagram generation; Based on the annotated 2D screenshot, a standardized on-site diagram template is nested to generate an accident scene diagram; The accident scene diagrams include: actual scene photos and scene record diagrams; The method for creating the on-site real-scene image includes the following operations: The screenshots of the actual scene with superimposed dimensions and standard graphic symbols are nested into the site map template, and basic information is filled in to generate a standardized site scene record map; the basic information includes: accident time, location, weather, survey personnel, and site description; The method for creating the on-site record diagram includes the following operations: Hide the background image of the real-scene screenshot, and only retain the standard graphic symbols, dimension lines, text annotations and standard drawing frames to generate a pure vector standardized on-site record image; In step S32, the model scale calibration operation includes the following: In the operation of setting size references, the national standard vehicle license plate is preferred as the size reference object. If the license plate is abnormal, a known size reference object on site is selected. The known size reference objects include: north markers, vehicle wheelbase, or standard road markings. The setting of the reference point includes the following operations: The reference point should be selected from permanent fixed objects on site, including: roadside lampposts, traffic sign posts, guardrail posts or fixed trees; if there are no permanent fixed objects on the core site, the endpoints of road markings should be selected as temporary reference points, or temporary reference points marked by road surface spraying or chalk should be used. Setting the baseline includes the following operations: The baseline is preferably selected from standard road markings, which include: road edge lines, lane dividing lines and stop lines. When there are no clear markings, the shoulder edge and the grounding edge of the guardrail are selected as the baseline carrier. For scenarios where the road markings are irregular and have bending or offset, it is supported to set multiple baselines in segments or to generate curved baselines through multi-point fitting. The selection of the test point includes the following operations: The test points for vehicle elements include: the contact point of the tire on the same side or the projection point of the wheel axle on the road surface; the test points for human elements include: the top of the head and the heel; the test points for road traces and scattered material evidence include: the starting point, the ending point, the turning point, the point of change in width, and the center point. Step S33 includes the following operations: S331: View adjustment and screenshot; Freely adjust the 3D model viewpoint to set the viewpoint where the overall layout of the accident and all labeled elements are clearly visible as the optimal viewpoint; S332: After locking the optimal viewpoint, take a screenshot of the scene and save it. When saving the scene screenshot, the annotation layer and the real scene base map are stored in separate layers. S333: Opens the system's built-in standard graphic symbol library; S334: Start the pre-trained target detection model; the target detection model performs target detection on the scene screenshot; The target detection model outputs the target type and target pixel bounding box for all identified targets; S335: Load the corresponding national standard vector symbol from the standard graphic symbol library according to the target type, and then scale and adjust the national standard vector symbol according to the size and angle of the target pixel bounding box to complete the initial automatic plotting; The target detection model is built based on the YOLO model; Step S332 includes the following operations: b1: Read the intrinsic parameter matrix K of the 3D camera under the current viewpoint, including the rotation matrix R and the translation vector t of the extrinsic parameter matrix; b2: The three-dimensional coordinates P of all labeled information in three-dimensional space. world (X,Y,Z) performs a coordinate transformation to generate two-dimensional pixel coordinates; the annotation information includes: endpoints of the annotation line segments, text, and measurement points; the calculation process includes the following steps: First, transform to the camera coordinate system P using rotation and translation matrices. cam =R·P world +t; Then, the projection formula through the pinhole camera is u=f x ·X / Z+C x and v=f y ·Y / Z+c y The calculated cutoff Figure 2 3D pixel coordinates (u,v); In the formula, f x and f y c is the lens focal length; x and c y Principal point coordinates; b3: Independently project and map all annotation information to generate a two-dimensional transparent layer, and bind the pixel coordinates of the projection layer to the projection calculation result (u,v) one by one; b4: Store the two-dimensional projection layer of the annotation information and the base image of the real scene screenshot in layers.

[0009] A 3D modeling and drawing system for real-time traffic accident scenes, characterized in that it includes: a panoramic video processing module, a 3D modeling algorithm module, and a scene drawing module; The panoramic video processing module is used to perform panoramic video frame extraction and differentiated perspective segmentation, eliminate image distortion, and generate a standardized image frame sequence. The panoramic video processing module includes a panoramic video frame extraction unit and a panoramic video perspective segmentation unit. The panoramic video frame extraction unit is used to extract key frames from the acquired panoramic video in batches at fixed time intervals, filter redundant frames, and generate a continuous and complete image frame sequence. The panoramic video perspective segmentation unit is used to execute differentiated perspective segmentation strategies of fine segmentation at small angles in the core accident area and coarse segmentation at large angles in the road background area, complete the conversion from spherical image to standard perspective image, and ensure a feature overlap of 15% to 20% between adjacent perspectives. The 3D modeling algorithm module is used to reconstruct the 3D structure of the scene based on image frame sequences. The reconstruction includes camera pose calculation, dense point cloud optimization, and outputs a high-precision real-scene 3D model. The 3D modeling algorithm module includes: a sparse point construction unit, a 3D Gaussian training unit, and a model format conversion unit. The sparse point construction unit is used to run the motion structure recovery algorithm, complete the camera pose calculation, and generate a sparse point cloud model of the scene. The 3D Gaussian training unit is used for dense point cloud reconstruction, detail enhancement, and noise filtering, and outputs a high-detail real-scene 3D model. The model format conversion unit is used to support the interoperability of multiple 3D model formats and adapt to multi-terminal rendering and drawing requirements. The on-site drawing module is used to automatically perform interactive editing of 3D models, national standard calibration measurement, graphic symbol plotting, and standardized on-site drawing generation and output; the on-site drawing module includes: a model rendering and calibration measurement unit, a graphic symbol production unit, and an on-site drawing generation unit; The model rendering and calibration measurement unit includes a model rendering component, a size calibration component, a ground reference component, a benchmark setting component, a positioning and ranging component, and a field element annotation component, which are used for model interactive operation and national standard standardized measurement and calibration. The graphic symbol creation unit includes a screenshot saving component, a layer display component, and a graphic symbol component, used for lossless screenshotting, annotation layer overlay, and national standard symbol drawing; the site map generation unit includes a site map nesting component and a site map output component, used for nesting, exporting, storing, and printing dual-type site map templates.

[0010] This application provides a method and system for 3D modeling and drawing of traffic accident scenes. Based on the wide coverage and high acquisition efficiency of panoramic video, it achieves comprehensive and rapid data collection from the accident scene. Combined with an optimized panoramic video segmentation method, motion structure restoration, and 3D technology, it constructs a more visually appealing and accurate 3D model of the accident scene. Simultaneously, it establishes a standardized calibration, measurement, and drawing process adapted to the standard system for traffic accident scene diagrams, generating compliant scene images and scene records. In terms of data acquisition, this method abandons the traditional single-view local shooting method and innovatively adopts a scheme combining panoramic video with full-area, blind-spot-free acquisition and differentiated acquisition using both high and low-level dual-circles. This ensures the complete recording of traditionally easily lost elements such as minor road surface traces, edge markings, and obstructed areas from the source. In terms of data processing, it constructs a standardized calibration, measurement, and drawing process adapted to national standards, achieving a technological leap from "visual display" to "standardized measurement and standardized drawing" of the 3D model. This method designs a standardized process for model calibration and precise measurement, which can shorten the time for on-site investigation, reduce equipment and operation thresholds, improve the accuracy and standardization of on-site map drawing, and promote the large-scale application of real-scene 3D technology in the field of road traffic accident investigation. Attached Figure Description

[0011] Figure 1 A schematic diagram illustrating the process of creating a 3D model of a traffic accident scene. Figure 2 A schematic diagram of a 3D modeling and drawing system for real-world traffic accident scenes; Figure 3 This is a real-scene image of the site drawn in the embodiment; Figure 4 This is a scene recording diagram drawn in the embodiment. Detailed Implementation

[0012] To address the shortcomings of existing multi-view geometric 3D modeling technologies, such as large blind spots in on-site data acquisition, poor detail reproduction of subtle traces and evidence, lack of standardized drawing processes adapted to the standard system for on-site diagram drawing, and inability to directly generate compliant on-site diagrams, this invention adopts a panoramic video full-domain, no-blind-spot acquisition scheme. It uses a high-low dual-circle differentiated acquisition method to fully cover the core area of ​​the accident scene, subtle road traces, and occluded areas; combined with a differentiated perspective segmentation strategy, it constructs a high-precision, high-detail 3D model of the accident scene. This method fully conforms to the standardized operating procedures of the traffic accident scene investigation and drawing standards and specifications stipulated in "Road Traffic Accident Scene Diagram Drawing" and "Road Traffic Accident Scene Graphic Symbols," and designs the 3D modeling and drawing process for traffic accident scenes. For example... Figure 1 As shown, this application proposes a method for three-dimensional modeling and drawing of a traffic accident scene, which includes the following steps.

[0013] S1: Fast panoramic video acquisition; A panoramic camera is used to collect full-area video of the traffic accident scene along a closed-loop acquisition path. The camera captures images around the accident scene at different heights to fully cover the core area of ​​the accident, the surrounding road environment, subtle traces and physical evidence, and obstructed areas, completely eliminating blind spots and providing a complete data source for high-precision modeling.

[0014] Specifically, it includes two sub-steps: S11 On-site data acquisition planning: Define a closed-loop data acquisition path based on the scope of the accident.

[0015] The method for confirming the closed-loop collection path designed in this application is as follows: taking the area composed of core accident elements as the center, and combining the on-site road direction, lane distribution, and roadside facility layout, a closed-loop collection path is planned; so that the collection path fully covers the core accident area, lanes, intersection markings, roadside facilities, and vegetation-covered areas; the core accident elements include: accident vehicles, collision marks, scattered objects, and bloodstains and scratches. S12 panoramic video capture.

[0016] The operator uses panoramic equipment to move at a constant speed along a preset circular path to collect panoramic video; the entire process is completed by circling the accident scene twice; the collection heights of the two circles are set differently: the first circle is above the top of the accident vehicle to capture the overall environment of the scene, fully covering the overall layout and surrounding environment. The second circle is below the operator's height to capture minor road surface details, obtaining a complete video data source covering the overall layout and detailed road surface elements; in this embodiment, the camera height of the second circle is 0.5 meters to 1 meter, focusing on low-altitude details such as road markings, minor scratches, and small debris.

[0017] Based on the scale of the accident scene covered by the closed-loop acquisition path, the acquisition time is adaptively matched, with shorter acquisition time for smaller scenes. The accident scene scale includes: large intersections or highway accident scenes as well as small accident scenes. Small accident scenes include: small single-vehicle and two-vehicle accident scenes. In this embodiment, the acquisition time for small single-vehicle and two-vehicle accident scenes is 1-2 minutes, and the acquisition time for large intersections and highway multi-vehicle accident scenes is 2-3 minutes, so as to ensure data integrity and avoid data redundancy.

[0018] S2: Refined 3D Modeling: Based on the acquired panoramic video, the viewpoint is segmented, and the SFM motion structure recovery algorithm and 3D Gaussian splash rendering technology are integrated to specifically solve the problems of missing edge details, low restoration of small physical evidence, and large scale deviation in traditional models, generating a high-precision realistic 3D model adapted to the needs of accident drawing. This includes the following steps.

[0019] S21: Panoramic video perspective segmentation: Differentiated perspective segmentation is performed on the acquired panoramic video, key frames are extracted in batches at fixed time intervals from the acquired original panoramic video, a continuous and complete sequence of on-site image frames is constructed, and the spherical image is converted into a standard perspective image. Based on the core area of ​​the accident, horizontal view blocks are dynamically divided in a differentiated manner, employing a segmentation strategy that combines fine and extensive approaches. Specifically, this includes: (1) The core area of ​​the accident is divided into small angles and the single-view field of view covers 45°~60° to maximize the preservation of local traces, scattered objects and marking details; the core area of ​​the accident includes: the accident-intensive area and the area where traces and physical evidence are concentrated; (2) The road background area adopts a large-angle coarse segmentation, with a single-view field of view covering 90°~120°, effectively reducing invalid data redundancy; the road background area includes: open road background and featureless area; All segmented sub-view images are uniformly converted into standard perspective projection images to eliminate panoramic spherical distortion, while retaining 15%~20% feature overlap between adjacent viewpoints, generating an image frame sequence. This provides sufficient matching points for subsequent SFM feature matching, completely avoiding the problems of missing or broken key accident features.

[0020] Based on image frame sequences, a motion structure recovery algorithm is used to calculate camera pose and reconstruct the three-dimensional spatial structure, thereby constructing a real-world three-dimensional model of the accident scene.

[0021] S22: Camera pose and spatial structure reconstruction: A motion structure recovery algorithm is used to extract and accurately match feature points in key frames, solve the camera intrinsic and extrinsic parameters and 3D spatial pose of each frame, restore the overall 3D spatial topology of the accident scene, and generate a high-density, high-matching sparse point cloud model of the scene. S23: Dense Point Cloud Generation and Optimization: Based on the sparse point cloud model, global dense reconstruction is carried out to complete the subtle spatial and texture information of the scene, filter the floating noise of the scene, and fully restore the small physical evidence; the small physical evidence includes: road scratches, scattered objects and light-colored markings.

[0022] S24: 3D Model Rendering: Convert high-precision dense point clouds into a visualized 3D model using a 3D Gaussian algorithm, generating a real-world 3D model of the accident scene with clear textures, accurate scale, and complete details.

[0023] This method employs panoramic video acquisition across the entire area without blind spots, combined with differentiated viewpoint segmentation, weak feature enhancement matching, and adaptive Gaussian rendering techniques. It effectively solves the problems of missing or blurred road surface traces, small pieces of evidence, and edge markings in traditional modeling. The model's detail fidelity is significantly improved, and the overall measurement accuracy reaches the centimeter level, fully meeting the national standard accuracy requirements for traffic accident investigation mapping. It can greatly enhance the completeness and measurement accuracy of 3D accident scene models.

[0024] S3: Standard site diagram drawing; Based on a realistic 3D model, and strictly following the national standards "Road Traffic Accident Scene Drawing" (GA / T 49-2019) and "Road Traffic Accident Scene Graphic Symbols" (GB / T 11797-2005), the system automatically performs model scale calibration, benchmark point and baseline setting, positioning and distance measurement, and on-site element annotation. After generating an annotated 3D model scene view, it calls the national standard graphic symbol library to draw graphic symbols and nests standard templates to generate a real-world scene map and a scene record map that conform to the national standards for traffic accident scene drawing. Step S3 specifically includes the following steps.

[0025] S31: Model display and trimming / editing; Load and render the pre-processed 3D model of the traffic accident scene, and perform custom trimming and editing.

[0026] This method addresses the issues of environmental occlusion and excessive invalid and redundant data in models by supporting custom trimming and editing. Custom trimming and editing includes two modes: planar trimming and 3D trimming. Planar trimming involves customizing the cutting plane to remove invalid model data from one side of the plane. 3D trimming involves customizing the effective area of ​​a cube, retaining the core accident model data inside the cube, removing irrelevant external environmental data, simplifying the model scene, and improving drawing accuracy.

[0027] This method supports model formats including but not limited to general 3D mesh models such as OBJ, FBX, PLY, Splat, and PCD, as well as Gaussian models and industry-specific model formats. The model data source supports both local file import and network transmission. The model rendering and display supports dual-platform adaptation for web browsers and client software.

[0028] S32: Model calibration measurement; This method incorporates a standardized process, which is used to complete model calibration and precise measurement. The standard drawing process includes the following steps performed in sequence: setting dimension reference, setting ground reference, setting benchmark point, setting baseline, selecting the point to be measured, selecting positioning method, and annotating on-site elements.

[0029] S321 sets the size reference as follows: Select a size reference object in the 3D model, set measurement points at both ends of the reference object, and use the known standard length of the reference object as a benchmark to complete the global scale mapping between the model space and the actual physical space, ensuring that the model size reproduces the actual size on site in a 1:1 ratio. The size reference object should preferably be a national standard vehicle license plate that conforms to the national standard size. If the license plate is abnormal, the on-site standard north marker should be used. If there is no north marker, the standard wheelbase of the vehicle and the standard road marking length can be used as the calibration basis. If the model already has absolute geographic coordinates and precise scale, this step can be skipped. License plate abnormalities include: license plate obstruction, deformation, or no license plate scene.

[0030] S322 sets the ground reference as follows: Select a flat, undulating standard road surface area at the accident site, set three non-collinear measurement points, and use the three points to generate a standard ground reference plane, unify the model's ground benchmark, and eliminate measurement errors caused by road surface undulations; if the model has accurate ground coordinate information, this step can be skipped.

[0031] S323 sets the reference points as follows: In the 3D model, select permanent fixed objects on site as reference carriers, and set measuring points at their grounding points, endpoints, and center points as reference points. Permanent fixed objects on site are preferred as reference points, including buildings, roadside lampposts, traffic sign posts, guardrail posts, or fixed trees. If there are no permanent fixed objects on site, select the endpoints of road markings as temporary references, or use temporary reference points marked by road surface spraying or chalk. Simultaneously record the relative distances between the temporary references and surrounding fixed objects and reference lines to ensure location traceability. Specifically, the right-angle positioning method and the comprehensive positioning method set one reference point per drawing, while the triangulation positioning method sets at least two reference points.

[0032] S324 sets the baseline as follows: Select regular linear elements in the model to construct the baseline; prioritize the use of standard road markings such as road edge lines, lane dividing lines, and stop lines; when there are no clear markings, use the shoulder edge and guardrail grounding edge as the reference carrier; for scenarios where the road markings are irregular and have bending offsets, it supports setting multiple baselines in segments or generating curved baselines through multi-point fitting to adapt to complex road scenarios.

[0033] S325 selects measurement points as follows: Measurement points for various accident elements are selected specifically according to the investigation standards. Accident elements include: vehicle elements, human elements, road surface traces, and debris evidence. Specifically, the measurement points for vehicle elements include: the contact point between the tires on the same side or the road surface projection point of the wheel axle; the measurement points for human elements include: the top of the head and the heel; the measurement points for road surface traces and debris evidence include: the starting point, the ending point, the turning point, the point of width change, and the center point. This comprehensively covers key measurement points for accident investigation.

[0034] S326 selects the following positioning methods: Based on the characteristics of the accident scenario, select a suitable method for accurate positioning and measurement of on-site elements; for regular and orderly road scenarios, prioritize the use of right-angle positioning or comprehensive positioning methods; for scenarios without regular baselines, complex intersections, or non-standard roads, use triangulation positioning to meet the positioning needs of all scenarios.

[0035] The S327 scene elements are marked as follows: three-dimensional spatial lead lines and text annotations are used to mark the accident vehicles, road traces and scene evidence, and the accident element information is recorded in a complete manner; the scene elements corresponding to the accident vehicles include: vehicle number, vehicle type and license plate information; the scene elements corresponding to the road traces include: brake marks, scratches and skid marks; the scene elements corresponding to the scene evidence include: glass fragments, paint fragments, bloodstains and scattered debris.

[0036] This application establishes a standardized 3D model exploration and mapping system, filling a technological gap in the industry. Existing real-scene 3D technology only achieves model visualization and simple distance measurement, lacking a standardized mapping process adapted to the traffic management industry. This application innovatively creates a standardized operation system covering the entire process of "dimension calibration → ground reference → benchmark setting → measurement point selection → positioning measurement → element annotation," achieving a technological leap from "visual display" to "standardized measurement and standardized mapping" of 3D models. It unifies the mapping standards for different operators, ensuring the consistency, legality, and standardization of exploration results.

[0037] S33: Graphic Symbol Creation; Based on the 3D model completed through calibration and measurement, complete the standardized graphic symbol drawing to meet the standard requirements for traffic accident scene drawing. Step S33 includes the following operations.

[0038] S331: View Adjustment and Screenshotting; Freely adjust the 3D model's perspective to set the optimal perspective where the overall layout of the accident and all labeled elements are clearly visible. In practice, the optimal perspective is preferably a top-down or oblique-top view to ensure that the overall layout of the accident and all labeled elements are clearly visible.

[0039] S332: After locking the optimal viewpoint, take a screenshot of the scene and save it. When saving the scene screenshot, the annotation layer and the real scene base map are stored in separate layers.

[0040] Step S332 includes the following operations: b1: Read the intrinsic parameter matrix K of the 3D camera under the current viewpoint, including the rotation matrix R and the translation vector t of the extrinsic parameter matrix; b2: The three-dimensional coordinates P of all labeled line segment endpoints, measuring points, and other labeled information in three-dimensional space. world (X,Y,Z) is transformed into two-dimensional pixel coordinates; First, transform to the camera coordinate system P using rotation and translation matrices.cam =R·P world +t; Then, the projection formula through the pinhole camera is u=f x ·X / Z+C x and v=f y ·Y / Z+c y The two-dimensional pixel coordinates (u,v) of the screenshot are calculated. In the formula, f x and f y c is the lens focal length; x and c y Principal point coordinates; b3: Independently project all annotation information such as lines, text and measurement points to generate a two-dimensional transparent layer, and bind the pixel coordinates of the two-dimensional projection layer to the projection calculation result (u,v) one by one; b4: Finally, the two-dimensional projection layer of the annotation information and the base image of the real scene are stored in layers to achieve lossless migration of the annotation information, which can effectively avoid the problems of misalignment, deformation and loss.

[0041] S333: Open the system's built-in standard graphic symbol library.

[0042] In this embodiment, the system has a built-in graphic symbol library for traffic accidents that conforms to the GB / T 11797-2005 standard, covering all categories of symbols for vehicles, pedestrians, road facilities, traces, and physical evidence.

[0043] Regarding the operation of graphic symbol drawing, this method supports two modes: manual drag-and-drop drawing and AI intelligent recognition automatic drawing. The process of graphic symbol drawing based on AI intelligent recognition automatic drawing is described in detail below.

[0044] AI-powered automatic object recognition is achieved through offline training of a YOLO object detection model: 3D screenshots of various accidents and real-world road images are collected to construct a dedicated training dataset. The dataset is labeled with categories including cars, SUVs, trucks, non-motorized vehicles, lane dividers, solid / dashed lines on road edges, brake marks, and debris. A lightweight YOLO network is used for supervised training. During the training phase, the CIoU loss function is used to optimize bounding box regression accuracy, and the Adam optimizer dynamically adjusts the learning rate. During the inference phase, the screenshot layer is input into the trained YOLO model, which outputs the target category, bounding box pixel coordinates, and confidence score. Invalid detection boxes with a confidence score below 0.7 are filtered out. The corresponding national standard symbol is automatically matched based on the center pixel coordinates of the detection box. Simultaneously, the symbol size is adaptively scaled according to the aspect ratio of the detection box, and the symbol angle is rotated to match the road direction. The symbol size, angle, and position can be adjusted according to the actual situation on site to accurately match the on-site elements in the screenshot and complete the standardized symbol drawing.

[0045] S334: Start the pre-trained target detection model for recognizing graphic symbols; the target detection model performs target detection on the scene screenshot; The object detection model outputs the target type and target pixel bounding box for all identified targets; the object detection model is built based on the YOLO model.

[0046] S335: Load the corresponding national standard vector symbol from the standard graphic symbol library according to the target type, and then scale and adjust the national standard vector symbol according to the size and angle of the target pixel bounding box to complete the initial automatic drawing of the graphic symbol.

[0047] S34: On-site diagram generation; Based on the annotated 2D screenshot, a standardized on-site diagram template is nested to generate an accident scene diagram; The accident scene diagrams include: actual scene photos and scene record diagrams; The method for creating on-site photos includes the following steps: The screenshots with superimposed dimensions and standard graphic symbols are nested into the site map template. Basic information is then filled in to generate a standardized site scene record map. The basic information includes: accident time, location, weather, survey personnel, and site description. The method for creating a field record diagram includes the following steps: The background image of the real-world screenshot is hidden, and only standard graphic symbols, dimension lines, text annotations, and standard drawing frames are retained to generate a pure vector standardized on-site record diagram. This meets the standard requirements for accident archiving and liability determination.

[0048] To implement the aforementioned method for 3D modeling and drawing of real-world traffic accident scenes, this application also designs a 3D modeling and drawing system for real-world traffic accident scenes, such as... Figure 2 As shown, it includes: panoramic video processing module M1, 3D modeling algorithm module M2, and on-site map drawing module M3.

[0049] The panoramic video processing module M1 is used to complete panoramic video frame extraction and differential viewpoint segmentation, eliminate image distortion, and generate a standardized image frame sequence. The panoramic video processing module M1 includes a panoramic video frame extraction unit M11 and a panoramic video viewpoint segmentation unit M12.

[0050] The panoramic video frame extraction unit M11 is used to extract key frames from the acquired panoramic video in batches at fixed time intervals, filter redundant frames, and generate a continuous and complete sequence of image frames to ensure data continuity.

[0051] The panoramic video view segmentation unit M12 is used to execute differentiated view segmentation strategies: fine segmentation at small angles for the core accident area and detailed areas, and coarse segmentation at large angles for the road background area. This differentiated view segmentation strategy completes the conversion from spherical image to standard perspective image, ensures 15%~20% feature overlap between adjacent viewpoints, and eliminates image distortion.

[0052] The 3D modeling algorithm module M2 is used to reconstruct the 3D structure of the scene based on the image frame sequence. The reconstruction includes camera pose calculation and dense point cloud optimization, and outputs a high-precision real-scene 3D model. The 3D modeling algorithm module M2 includes: sparse point construction unit M21, 3D Gaussian training unit M22 and model format conversion unit M23.

[0053] The sparse point construction unit M2 is used to run the motion structure recovery algorithm, complete the camera pose calculation, and generate a sparse point cloud model of the site. The 3D Gaussian training unit M22 completes the reconstruction of dense point clouds, detail enhancement, and noise filtering, adaptively adjusts the radius of the Gaussian sphere, achieves refined rendering of key areas and lightweight rendering of background areas, and outputs a high-detail realistic 3D model. The model format conversion unit M23 supports mutual conversion between multiple 3D model formats, adapts to the rendering and drawing needs of web and client applications, and is compatible with mainstream exploration equipment and software platforms.

[0054] The site drawing module M3 is used to automatically perform interactive editing of 3D models, national standard calibration measurement, graphic symbol plotting, and standardized site drawing generation and output. The site drawing module includes M3: model rendering and calibration measurement unit M31, graphic symbol production unit M32, and site drawing generation unit M33.

[0055] The model rendering and calibration measurement unit M31 includes a model rendering component M311, a dimension calibration component M312, a ground reference component M313, a benchmark setting component M314, a positioning and ranging component M315, and a field element annotation component M316, which are used for model interactive operation and national standard standardized measurement calibration.

[0056] Model rendering component M311: Supports full-dimensional interactive operations such as loading, displaying, rotating, scaling, translating, and trimming planar / stereoscopic models in multiple formats, adapting to multi-terminal display needs.

[0057] Dimension calibration component M312: Receives the reference object and standard dimension parameters selected by the user, establishes a scale mapping between model space and physical space, and completes global dimension calibration.

[0058] Ground reference component M313: Receives three non-co-located ground measurement points selected by the user, fits them to generate a standard ground plane, and unifies the measurement benchmark.

[0059] The M314 benchmark setting component supports user-defined selection and setting of benchmark points and benchmark lines, and is compatible with three types of national standard positioning methods.

[0060] The M315 positioning and ranging component integrates three national standard positioning algorithms: right-angle positioning, triangulation positioning, and comprehensive positioning. It automatically calculates the relative distance between each measuring point and generates accurate dimension annotations.

[0061] The M316 on-site element annotation component supports 3D text and leader line annotations for vehicles, personnel, traces, and physical evidence, and completely preserves the exploration information.

[0062] The graphic symbol creation unit M32 includes a screenshot saving component M321, a layer display component M322, and a graphic symbol component M323, which are used for lossless screenshots, annotation layer overlays, and national standard symbol drawing.

[0063] Screenshot saving component M321: Captures the optimal 3D scene view and retains all 3D annotation information.

[0064] Layer display component M322: accurately converts 3D spatial annotation information into 2D layers and overlays them onto the screenshot surface without loss of quality.

[0065] Graphic symbol component M323: Built-in national standard exclusive graphic symbol library, supports two modes: manual drawing and AI intelligent recognition drawing, to quickly complete the restoration of on-site element symbols.

[0066] The site diagram generation unit M33 includes a site diagram nesting component M331 and a site diagram output component M332, which are used for nesting, exporting, storing and printing dual-type site diagram templates.

[0067] The M331 on-site image nesting component supports nesting of both real-scene images and record images, and can generate real-scene record images with background images and standard record images without background images with one click.

[0068] The M332 on-site image output component supports local export, cloud storage, and printing of on-site images. It is compatible with common formats such as PDF and images, meeting the needs of archiving and case handling.

[0069] This proposed solution can effectively solve the problems of lack of detail in traditional modeling, non-standard drawing output, and low work efficiency. It achieves centimeter-level modeling and drawing accuracy, significantly reduces equipment and operation thresholds, is not restricted by weather, obstructions, or restricted flight areas, and can quickly complete digital survey and drawing of accident scenes. It is suitable for various road accident scenarios and is conducive to large-scale promotion and application by grassroots traffic management units.

[0070] The present invention will be further described in detail below with reference to specific embodiments.

[0071] The example is: implementation of scene modeling and drawing of a two-vehicle rear-end collision on an ordinary urban road.

[0072] Combination Figure 1 The method and process shown Figure 2 The system architecture shown in this embodiment fully demonstrates the entire operational process of the method and system of this invention in a typical small two-vehicle rear-end collision in an urban area. The accident scenario is a two-lane road in both directions in an urban area, with ordinary asphalt road surface, white lane dividing lines on both sides of the road, yellow lane dividing lines in the center, and directional arrow markings at intersections. There are no tall roadside trees, billboards, or other obstructions. The weather on that day was sunny, windless, and with sufficient light, without rain, snow, backlight, or other interfering conditions. The accident involved a small red family sedan (Vehicle A) being rear-ended by a gray small sedan (Vehicle B) in the straight lane. There were no bloodstains or brake marks on the road surface at the scene, which is a common accident. The complete operation steps are detailed as follows.

[0073] S1, Fast Panoramic Video Acquisition Upon arrival at the scene, the investigators first placed warning triangles and cones in the direction of oncoming traffic to ensure on-site safety and prevent secondary accidents. Then, the two investigators divided the work: one was responsible for on-site information registration, recording the time, location, vehicle license plate, and information of the parties involved in the accident; the other investigator performed panoramic video acquisition, using a common consumer-grade handheld panoramic camera. The equipment only supports panoramic video recording and does not have LiDAR or high-precision positioning modules. The hardware cost is low and it can be mass-produced at the grassroots level.

[0074] The first step is to implement on-site data collection planning: Investigators visually delineate the accident data collection boundaries on-site. The core area includes the entire bodies of Vehicle A and Vehicle B, the point of contact between the two vehicles, and the area where scattered paint and glass fragments are concentrated. Based on the east-west straight road surface at the scene, a closed-loop walking path is planned. The loop extends outwards to cover a 10-meter lane radius in front of and behind the accident vehicles, fully encompassing the white lane markings, the solid yellow curb line on the right, the wave-shaped guardrail posts, and the straight-ahead directional arrows on the ground. The path avoids the vehicle chassis and blind spots inside the guardrails, ensuring that all traces and evidence on the road surface are within the area enclosed by the loop collection path, with no blind spots. The second step involves performing differentiated panoramic video acquisition using both high and low elevations: 1) First high-altitude data collection: The operator fixes the panoramic camera to the top of a 1.8m handheld extension pole, which is higher than the roofs of the two accident vehicles (the car roof is about 1.45m). The operator walks slowly and at a constant speed along the planned circular path, with the walking speed controlled at 12 meters per minute. The camera lens is facing the inside of the circle to continuously record panoramic video, fully capturing the overall direction of the entire road, the relative positions of the two vehicles, the roadside guardrails, the distant intersection signs, the overall layout of the lane markings, and recording the macroscopic environmental information of the scene; 2) Second loop of low-altitude data collection: Remove the extension pole, hold the panoramic camera at a fixed height of 1.2m, tilt the camera lens slightly downwards to aim at the road surface, and walk around the same circular path again to collect low-altitude details of the road surface at close range, and fully capture details such as white and yellow lane markings and the sides of vehicles.

[0075] The total recording time for this minor two-vehicle rear-end collision was 1 minute and 30 seconds, with 48 seconds of high-circle recording and 42 seconds of low-circle recording. The video file was automatically stored to the camera's local memory card. After the recording was completed, the video could be directly transferred to the modeling and drawing system of this invention via a data cable.

[0076] S2, Refined 3D Modeling Surveyors import panoramic video into the modeling and drawing system. The system automatically starts the panoramic video processing module and the 3D modeling algorithm module, performing fully automated calculations without manual intervention in parameter adjustment. The step-by-step processing flow is as follows: 1) Panoramic video frame extraction and differentiated perspective segmentation: The system extracts key frames from the video in batches at a fixed time interval of 0.5 seconds, automatically filters out jittery and repetitive frames without change, and generates a continuous and uniform image frame sequence; the system automatically identifies the core accident area (the bodies of the two vehicles and the area of ​​scattered objects) and the open road background area in the picture, and uses a 50° single field of view small angle fine segmentation for the core area, and a 100° large angle coarse segmentation for the open road surface without traces in front and behind. All spherical panoramic images are uniformly converted into standard perspective images to eliminate edge stretching distortion of the panoramic picture. Adjacent segmented images automatically retain 18% feature overlap area to ensure that trace and marking features are continuous and uninterrupted.

[0077] 2) Camera pose and 3D spatial structure reconstruction: The system runs a motion structure recovery algorithm to extract and match feature points for all segmented images, automatically solves for internal and external parameters such as camera shooting position and shooting angle for each image, fully restores the 3D spatial topology of the scene, and generates a sparse point cloud model of the scene. The point cloud can identify large feature objects such as vehicle outlines, road markings, and guardrail posts.

[0078] 3) Dense point cloud generation and noise optimization: Global dense reconstruction is carried out based on sparse point cloud to supplement weak texture details such as small scratches, debris, and light-colored markings on the road surface; the system automatically identifies floating and unsupported noise points in the image, filters dust in the air, irrelevant pedestrians and vehicle shadows in the distance, and retains only the real spatial point cloud data of the ground, accident vehicles, and roadside fixed facilities.

[0079] 4) Unified rendering output of 3D models: The system converts the optimized dense point cloud into a general 3D mesh model and outputs a standard ply format file. The model fully restores the centimeter-level details of the scene and can clearly distinguish the length of brake marks, the distribution range of scattered objects, and the contact position of the tires of the two vehicles, without any issues of detail blurring or loss.

[0080] S3, Standard Site Diagram Drawing The system loads the generated 3D mesh model and renders it on the web-based interactive interface. Since the site has no large number of irrelevant trees or redundant buildings, no planar / stereoscopic pruning is required. Surveyors completed calibration measurements, feature annotation, symbol plotting, and site map generation step-by-step according to standardized national procedures. The complete operational details are as follows: 1) Model scale calibration and ground benchmark setting: ① Dimensional reference calibration: The system interface provides a dimension calibration tool. Surveyors can click on both ends of the national standard motor vehicle license plate at the rear of vehicle B. The system has a built-in standard license plate length parameter of 44cm. With one click, the global scale mapping between the model's three-dimensional space and the real physical space is completed. All dimensions of the model are uniformly restored to a 1:1 true scale, and the measurement error is controlled within ±1%. ② Ground reference plane fitting: On a smooth, pothole-free asphalt road surface between vehicles A and B, click to select three non-collinear ground measurement points in sequence. The system automatically fits the three points to generate a unified standard ground reference plane, eliminating the vertical height measurement deviation caused by minor road surface undulations.

[0081] 2) Establishment of benchmark points and baselines, and selection of points to be measured: ① Reference point setting: Select the tip of the straight guide arrow on the ground as the reference point, and click on the point in the model interface to complete the reference point marking; ② Baseline setting: Select the solid white edge line on the left side of the road as the baseline, and the system will automatically generate a continuous linear reference line along the road markings; ③ Differentiated test point selection: Mark the measurement points of the two vehicles according to the national standard requirements. Mark the ground contact points of the left front and left rear tires of vehicle A, and mark the ground contact points of the left front and left rear tires of vehicle B.

[0082] 3) Positioning and ranging, and 3D feature annotation: The road markings on site are neat and orderly. The system defaults to using the right-angle positioning algorithm, automatically calculating the vertical distance of all test points from the baseline and the longitudinal distance from the baseline, and automatically generating white dimension markers and values. Surveyors use the annotation tool to complete the three-dimensional text markers: vehicle A is marked "Vehicle A", vehicle B is marked "Vehicle B", and all markers float above the corresponding elements of the three-dimensional model without obscuring traces or vehicle bodies.

[0083] 4) Perspective screenshot and national standard graphic symbol plotting: ① Viewpoint Locked Screenshot: The surveyor drags the 3D model to a 45° overhead view, fully displaying both vehicles, all traces, and baseline markings, without obstructions or missing elements. After locking the viewpoint, clicking the screenshot button automatically reads the camera's internal and external parameters for the current viewpoint. Through mathematical conversion using a 3D pinhole projection matrix, it accurately maps all 3D spatial dimensions, text annotations, and measurement point coordinates to the screenshot. Figure 2 The annotation layer is a pixel-level layer, with the annotation layer and the actual scene base map being isolated in layers. The coordinate mapping error is less than 1 pixel, enabling lossless transfer of annotations. ② National Standard Symbol Plotting: Open the system's built-in GB / T 11797-2005 graphic symbol library. You can choose to manually drag and drop for plotting or use YOLO for automatic recognition. After starting the AI ​​recognition function, the system calls the local lightweight YOLO inference model to perform target detection on the panoramic screenshot. The model is specially trained based on the traffic management accident image dataset and accurately identifies vehicle A, vehicle B, yellow dashed lines, and white curb lines in the screenshot. It automatically outputs the bounding boxes of each target pixel. The system automatically loads the corresponding national standard vector symbols according to the detection results, and adaptively scales and rotates the symbols according to the length and width of the detection box and the road tilt angle to complete the initial automatic plotting. Surveyors only need to make minor adjustments to the symbol positions to match the actual scale of the scene, greatly reducing the workload of manual plotting.

[0084] 5) Two types of standardized sites Figure 1 Key generation and export: ① Generation of on-site real-scene images: A real-scene screenshot with overlaid annotations and national standard symbols is nested within the GA / T 49-2019 standard real-scene image template. The time, location, weather, road surface type, surveyor, draftsman, and brief on-site description are manually filled in. The template includes a signature field and an inspection record field, generating a standardized on-site real-scene image with a real-scene base map. See attached details. Figure 3 ; ② Generation of On-site Recording Diagrams: With one click, the background image of the actual scene photo is hidden, retaining only vector national standard symbols, dimension annotations, text descriptions, and standard drawing frames to generate a standardized on-site recording diagram with pure lines, as detailed in the attached document. Figure 4 It fully meets the requirements for archiving traffic accident files and providing written evidence for liability determination.

[0085] The entire process, from on-site filming to exporting two compliant on-site maps, takes only 1.5 minutes after modeling is completed, saving at least 60% of the time compared to traditional manual measurement and drawing. The model's measurement accuracy, drawing format, symbol usage, and positioning methods all comply with current traffic accident scene drawing standards, requiring no secondary modifications. The maps can be directly printed, signed, and archived. The equipment is lightweight and easy to operate, allowing a single person to complete all survey and drawing work, perfectly meeting the needs of rapid handling of daily accidents at the grassroots level.

[0086] Using the technical solution of this invention, data acquisition can be completed using ordinary panoramic cameras, eliminating the need for expensive laser scanning equipment and significantly reducing hardware costs. Simultaneously, through automated algorithm processing and AI-assisted plotting, it replaces traditional manual measurement and drawing, reducing on-site drawing time from over 30 minutes to less than 10 minutes. The operation is simple, highly efficient, and suitable for widespread application by grassroots traffic management units. This significantly lowers the barrier to entry for the implementation of real-scene 3D survey technology. This method is not limited by weather, obstructions, or restricted flight areas, and can be used normally in scenarios where drones and traditional shooting methods cannot operate, such as when trees obstruct the view, in bad weather, or in urban restricted flight zones. It has wider scene adaptability and greater practicality. This method is suitable for accident investigation needs across all scenarios and has strong anti-interference capabilities.

Claims

1. A method for three-dimensional modeling and drawing of a traffic accident scene, characterized in that, It includes the following steps: S1: Fast panoramic video acquisition; Determine the closed-loop acquisition path based on the scope of the accident; A panoramic camera is used to collect full-area video of the traffic accident scene along the closed-loop acquisition path. Two circles of panoramic video are collected around the accident scene at different heights. The first circle is collected at a height higher than the top of the accident vehicle to capture the overall environment of the scene, and the second circle is collected at a height lower than the height of the operator to capture the fine elements of the road surface, thus obtaining a complete video data source covering the overall layout of the scene and the detailed elements of the road surface. S2: Refined 3D modeling, including the following steps: Panoramic video perspective segmentation: Differentiated perspective segmentation is performed on the acquired panoramic video, the spherical image is converted into a standard perspective image, and keyframes are extracted in batches at fixed time intervals to construct a continuous and complete sequence of on-site image frames; specifically, the core accident area is segmented with a small angle of 45°~60° for fine segmentation, and the road background area is segmented with a large angle of 90°~120° for coarse segmentation, generating an image frame sequence in which adjacent perspectives retain 15%~20% of the feature overlap area; Based on the image frame sequence, a motion structure recovery algorithm is used to calculate the camera pose and reconstruct the three-dimensional spatial structure to build a real-world three-dimensional model of the accident scene. S3: Standard site diagram drawing; Based on the real-scene 3D model, the computer performs model scale calibration, benchmark point and baseline setting, positioning and distance measurement and on-site element annotation. After generating the annotated 3D model scene view, the national standard graphic symbol library is called to draw graphic symbols and nest standard templates to generate a real-scene picture and on-site record picture that conform to the national standard specification for traffic accident scene drawing.

2. The method for three-dimensional modeling and drawing of a traffic accident scene according to claim 1, characterized in that: In step S1, the method for confirming the closed-loop collection path is as follows: taking the area composed of core accident elements as the center, and combining the on-site road direction, lane distribution, and roadside facility layout, a closed-loop collection path is planned. The data collection path fully covers the accident core area, lanes, intersection markings, roadside facilities, and areas obscured by vegetation. The core accident elements include: the accident vehicle, collision marks, human body, scattered objects, and bloodstains and scratches; Based on the scale of the scene covered by the closed-loop acquisition path, the acquisition time is adaptively matched, with shorter acquisition time for smaller scenes; the scene scale includes: large intersections or highway accident scenes as well as small accident scenes.

3. The method for three-dimensional modeling and drawing of a traffic accident scene according to claim 1, characterized in that: In step S2, a three-dimensional model of the accident scene is constructed based on the image frame sequence, specifically including the following steps: a1: Camera pose and spatial structure reconstruction: The motion structure recovery algorithm is used to extract and accurately match feature points in key frames, solve the camera intrinsic parameters, extrinsic parameters and 3D spatial pose of each frame, restore the overall 3D spatial topology of the accident scene, and generate a highly matched sparse point cloud model of the scene. a2: Dense point cloud generation and optimization: Global dense reconstruction is carried out based on sparse point cloud model to complete the subtle spatial and texture information of the scene, filter the floating noise of the scene, and fully restore the small evidence; the small evidence includes: road scratches, scattered objects and road markings. a3: 3D model rendering: Converts high-precision dense point clouds into a visualized 3D model using a 3D Gaussian algorithm, generating a realistic 3D model of the accident scene.

4. The method for three-dimensional modeling and drawing of a traffic accident scene according to claim 1, characterized in that: Step S3 specifically includes the following steps: S31: Model display and trimming / editing; Load and render the pre-processed 3D model of the traffic accident scene, and perform custom trimming and editing. The custom trimming and editing includes two modes: planar trimming and 3D trimming. Planar trimming involves customizing the cutting plane to remove invalid model data from one side of the plane. 3D trimming involves customizing the effective area of ​​a cube, retaining the core accident model data inside the cube, removing irrelevant external environmental data, and simplifying the model scene. S32: Model calibration measurement; Establish a standardized process, and complete model calibration and precise measurement based on the standardized process; The standard drawing process includes the following steps executed in sequence: setting dimension reference, setting ground reference, setting benchmark point, setting benchmark line, selecting the point to be measured, selecting positioning method, and annotating on-site elements. The setting of the size reference is as follows: Select a size reference object in the 3D model, set measurement points at both ends of the reference object, and use the known standard length of the reference object as a benchmark to complete the scale calibration between the model space and the actual physical space, ensuring that the model size restores the actual size on site in a 1:1 ratio; if the model already has absolute geographic coordinates and precise scale, this step can be skipped. The ground reference setting is as follows: Select a flat, undulating standard road surface area at the accident site, set three non-collinear measurement points, and fit the three points to generate a standard ground reference plane to unify the model's ground benchmark and eliminate measurement errors caused by road surface undulations; if the model has accurate ground coordinate information, this step can be skipped. The reference points are defined as follows: the reference carriers of the three-dimensional model include: permanent fixed objects on site and temporary reference points; If there are permanent fixed objects on site, select the permanent fixed objects as the reference carrier, and set measuring points at the grounding point, end point and center point of the permanent fixed objects as reference points; otherwise, if there are no permanent fixed objects on site, use temporary reference points, and record the relative distance between the temporary reference points and the surrounding fixed objects and reference lines at the same time to ensure the traceability of the positioning. The baseline is defined as follows: a baseline is constructed by selecting orthogonal elements in the model; The selection of test points involves: selecting measurement points for various accident elements based on the investigation standards; the accident elements include: vehicle elements, human elements, road surface traces, and scattered object evidence. The selected positioning method is as follows: based on the characteristics of the accident scenario, a suitable method for accurate positioning and measurement of on-site elements is selected; for regular and orderly road scenarios, the right-angle positioning method or the comprehensive positioning method is preferred; for scenarios without regular baselines, complex intersections, or non-standard roads, the triangulation positioning method is used. The element labeling refers to: three-dimensional spatial line labeling and text annotation of accident vehicles, road traces, and on-site evidence to fully record accident element information; the on-site elements corresponding to the accident vehicles include: vehicle number, vehicle type, and license plate information; the on-site elements corresponding to the road traces include: brake marks, scratches, and skid marks; the on-site elements corresponding to the on-site evidence include: glass fragments, paint fragments, bloodstains, and scattered debris. S33: Graphic symbol creation; Based on the 3D model completed by calibration measurement, complete the standardized graphic symbol drawing to meet the standard requirements for traffic accident scene drawing; S34: On-site diagram generation; Based on the annotated 2D screenshot, a standardized on-site diagram template is nested to generate an accident scene diagram; The accident scene diagrams include: actual scene photos and scene record diagrams; The method for creating the on-site real-scene image includes the following operations: The screenshots of the actual scene with superimposed dimensions and standard graphic symbols are nested into the site map template, and basic information is filled in to generate a standardized site scene record map; the basic information includes: accident time, location, weather, survey personnel, and site description; The method for creating the on-site record diagram includes the following operations: Hides the background image of the real-scene screenshot, retaining only standard graphic symbols, dimension lines, text annotations, and standard drawing frames, and generates a pure vector standardized on-site record image.

5. The method for three-dimensional modeling and drawing of a traffic accident scene according to claim 4, characterized in that: In step S32, the model scale calibration operation includes the following: In the operation of setting size references, the national standard vehicle license plate is preferred as the size reference object. If the license plate is abnormal, a known size reference object on site is selected. The known size reference objects include: north markers, vehicle wheelbase, or standard road markings. The setting of the reference point includes the following operations: The reference point should be selected from permanent fixed objects on site, including: roadside lampposts, traffic sign posts, guardrail posts or fixed trees; if there are no permanent fixed objects on the core site, the endpoints of road markings should be selected as temporary reference points, or temporary reference points marked by road surface spraying or chalk should be used. Setting the baseline includes the following operations: The baseline is preferably selected from standard road markings, which include: road edge lines, lane dividing lines and stop lines. When there are no clear markings, the shoulder edge and the grounding edge of the guardrail are selected as the baseline carrier. For scenarios where the road markings are irregular and have bending or offset, it is supported to set multiple baselines in segments or to generate curved baselines through multi-point fitting. The selection of the test point includes the following operations: The test points for vehicle elements include: the contact point of the tire on the same side or the projection point of the wheel axle on the road surface; the test points for human elements include: the top of the head and the heel; the test points for road traces and scattered evidence include: the starting point, the ending point, the turning point, the point of change in width, and the center point.

6. The method for three-dimensional modeling and drawing of a traffic accident scene according to claim 4, characterized in that: Step S33 includes the following operations: S331: View Adjustment and Screenshotting; Freely adjust the 3D model viewpoint to set the viewpoint where the overall layout of the accident and all labeled elements are clearly visible as the optimal viewpoint; S332: After locking the optimal viewpoint, take a screenshot of the scene and save it. When saving the scene screenshot, the annotation layer and the real scene base map are stored in separate layers. S333: Opens the system's built-in standard graphic symbol library; S334: Start the pre-trained target detection model; the target detection model performs target detection on the scene screenshot; The target detection model outputs the target type and target pixel bounding box for all identified targets; S335: Load the corresponding national standard vector symbol from the standard graphic symbol library according to the target type, and then scale and adjust the national standard vector symbol according to the size and angle of the target pixel bounding box to complete the initial automatic plotting.

7. The method for three-dimensional modeling and drawing of a traffic accident scene according to claim 6, characterized in that: The target detection model is built based on the YOLO model.

8. The method for three-dimensional modeling and drawing of a traffic accident scene according to claim 6, characterized in that: Step S332 includes the following operations: b1: Read the intrinsic parameter matrix K of the 3D camera under the current viewpoint, including the rotation matrix R and the translation vector t of the extrinsic parameter matrix; b2: The three-dimensional coordinates P of all labeled information in three-dimensional space. world (X,Y,Z) performs a coordinate transformation to generate two-dimensional pixel coordinates; the annotation information includes: endpoints of the annotation line segments, text, and measurement points; the calculation process includes the following steps: First, transform to the camera coordinate system P using rotation and translation matrices. cam =R·P world +t; Then, the projection formula through the pinhole camera is u=f x ·X / Z+C x and v=f y ·Y / Z+c y The two-dimensional pixel coordinates (u,v) of the screenshot are calculated. In the formula, f x and f y c is the lens focal length; x and c y Principal point coordinates; b3: Independently project and map all annotation information to generate a two-dimensional transparent layer, and bind the pixel coordinates of the projection layer to the projection calculation result (u,v) one by one; b4: Store the two-dimensional projection layer of the annotation information and the base image of the real scene screenshot in layers.

9. A three-dimensional modeling and drawing system for real-scene traffic accident scenes, characterized in that, It includes: Panoramic video processing module, 3D modeling algorithm module, and on-site map drawing module; The panoramic video processing module is used to perform panoramic video frame extraction and differential viewpoint segmentation, eliminate image distortion, and generate a standardized image frame sequence. The panoramic video processing module includes a panoramic video frame extraction unit and a panoramic video view segmentation unit. The panoramic video frame extraction unit is used to extract key frames from the acquired panoramic video in batches at fixed time intervals, filter redundant frames, and generate a continuous and complete image frame sequence. The panoramic video view segmentation unit is used to execute a differentiated view segmentation strategy of fine segmentation at small angles in the core accident area and coarse segmentation at large angles in the road background area, complete the conversion from spherical image to standard perspective image, and ensure a feature overlap of 15% to 20% between adjacent viewpoints. The 3D modeling algorithm module is used to reconstruct the 3D structure of the scene based on image frame sequences. The reconstruction includes camera pose calculation, dense point cloud optimization, and outputs a high-precision real-scene 3D model. The 3D modeling algorithm module includes: a sparse point construction unit, a 3D Gaussian training unit, and a model format conversion unit. The sparse point construction unit is used to run the motion structure recovery algorithm, complete the camera pose calculation, and generate a sparse point cloud model of the scene. The 3D Gaussian training unit is used for dense point cloud reconstruction, detail enhancement, and noise filtering, and outputs a high-detail real-scene 3D model. The model format conversion unit is used to support the interoperability of multiple 3D model formats and adapt to multi-terminal rendering and drawing requirements. The on-site drawing module is used to automatically perform interactive editing of 3D models, national standard calibration measurement, graphic symbol plotting, and standardized on-site drawing generation and output; the on-site drawing module includes: a model rendering and calibration measurement unit, a graphic symbol production unit, and an on-site drawing generation unit; The model rendering and calibration measurement unit includes a model rendering component, a size calibration component, a ground reference component, a benchmark setting component, a positioning and ranging component, and a field element annotation component, which are used for model interactive operation and national standard standardized measurement and calibration. The graphic symbol creation unit includes a screenshot saving component, a layer display component, and a graphic symbol component, used for lossless screenshotting, annotation layer overlay, and national standard symbol drawing; the site map generation unit includes a site map nesting component and a site map output component, used for nesting, exporting, storing, and printing dual-type site map templates.

Images