Traffic scene simulation method based on accident report

Abstract

The invention discloses a traffic scene simulation method based on an accident report, and aims to solve the problem of insufficient complex dynamic scene data in traffic simulation. According to the method, physical clues (such as vehicle speed, collision angle, quality and the like) and context information (such as road types and environmental conditions) are extracted from a real traffic accident report, and a high-fidelity traffic scene data set is generated. In foreground processing, the extracted physical clues are used for accurately constructing a vehicle motion trajectory, diversified dynamic traffic scenes are generated through a pre-collision trajectory planning algorithm, and trajectory prediction is optimized in combination with a large language model to ensure that the real physical law is met. In background processing, high-quality three-dimensional scene reconstruction is carried out according to environment information in an accident report, a vivid background image is generated, and the consistency of a visual angle and illumination is kept. And finally, through fusion of the foreground vehicle track and the background image, a visual and physical double real traffic simulation video is generated. The method significantly improves the diversity and authenticity of traffic scenes, reduces the dependence on real data collection, is widely suitable for high-fidelity traffic simulation scenes such as automatic driving test and traffic flow analysis, and effectively promotes the progress of the traffic simulation technology.

Description

Technical Field

[0001] This invention relates to the field of traffic simulation and autonomous driving testing, and in particular to a simulation method for generating high-fidelity traffic scenarios based on real traffic accident reports. By extracting physical cues and contextual information, and utilizing a physics simulation engine and neural rendering technology, it generates visually and physically realistic traffic simulation videos, which are applicable to scenarios such as autonomous driving testing, traffic flow analysis, and intelligent transportation system development. Background Technology

[0002] Traffic simulation technology is crucial in autonomous driving and intelligent transportation systems, providing diverse driving scenario data for model training and system testing. However, existing methods face the challenge of insufficient data when generating complex dynamic scenarios, especially those involving diverse road types and vehicle interactions. Traditional data acquisition relies on on-site recording, which is costly and struggles to cover rare scenarios, resulting in datasets lacking diversity and physical realism. While existing video generation methods can produce visually realistic driving scenarios, they often fail to accurately simulate vehicle dynamics due to a lack of physical constraints, limiting their application in safety-critical scenarios. To address this, this invention proposes a traffic scenario simulation method based on accident reports. This method utilizes physical and contextual information from accident reports to generate high-fidelity traffic scenarios through physical simulation and neural rendering techniques. Summary of the Invention

[0003] This invention proposes a traffic scenario simulation method based on accident reports, comprising the following steps:

[0004] Accident report information extraction: Extract physical clues (such as vehicle speed, collision angle, vehicle mass, etc.) and contextual information (such as road type, environmental conditions, etc.) from real traffic accident reports to form structured data, providing a foundation for subsequent simulations.

[0005] Vehicle trajectory acquisition: The extracted physical clues are input into a physics simulation engine (such as CARLA) to generate a vehicle motion trajectory dataset that conforms to real physical laws, and dynamic traffic scenarios are generated through pre-collision trajectory planning.

[0006] Vehicle trajectory prediction: Optimize trajectory prediction using large language models (such as LLaMa) to ensure that the simulated scene conforms to physical constraints.

[0007] Background Reconstruction: Based on contextual information, techniques such as Neural Radiation Field (NeRF) are used to reconstruct 3D scenes, generating highly realistic background images while maintaining consistency between lighting and viewpoint.

[0008] Foreground and background fusion: By using channel synthesis technology, the foreground vehicle trajectory is fused with the background image to generate a traffic simulation video that is both visually and physically realistic.

[0009] Video output: Outputs the final traffic simulation video, which has a high degree of physical realism and visual realism, and is widely applicable to autonomous driving testing and traffic flow analysis applications. Attached Figure Description

[0010] Figure 1 This is a schematic diagram illustrating the architecture and process of the traffic scenario simulation method based on accident reports proposed in this invention. Detailed Implementation

[0011] Reference Figure 1 The traffic scenario simulation method based on accident reports proposed in this invention includes the following steps:

[0012] Step S1: Accident Report Information Extraction. Key information is extracted from real traffic accident reports, including physical clues (such as vehicle speed, collision angle, vehicle mass, etc.) and contextual information (such as road type, slope, environmental conditions, etc.). A large language model combined with a few-shot learning strategy is used to automatically parse the pre-collision environmental conditions, collision dynamics, and post-collision results from the report, generating structured data. The extracted information covers road features, vehicle dynamics, and environmental factors, providing comprehensive support for subsequent simulations.

[0013] Step S2: Vehicle Trajectory Acquisition. The extracted physical cues are input into a physics simulation engine (such as CARLA) to generate a high-fidelity vehicle motion trajectory dataset. The simulation process considers physical parameters (such as friction, gravity, and road slope) to ensure the trajectory conforms to real-world dynamics. Pre-collision trajectory planning consists of three sub-steps:

[0014] (a) Initial state extraction: Analyze vehicle dynamic information and collision parameters to determine the initial position, speed and collision target of each vehicle;

[0015] (b) Lane and path selection: Based on the scene map and vehicle dynamics, generate candidate lane combinations that meet collision constraints;

[0016] (c) Trajectory Generation and Validation: Dynamically feasible vehicle trajectories are generated based on the selected lane combinations, and their effectiveness is validated. The generated trajectory dataset records complete vehicle motion information, providing a foundation for subsequent optimization and video generation.

[0017] Step S3: Vehicle trajectory prediction. A large language model (such as Llama) is used to fine-tune the trajectory dataset using the low-rank adaptation (LoRA) method to optimize trajectory prediction accuracy and ensure that the generated dynamic traffic scene conforms to physical constraints.

[0018] Step S4: Background Reconstruction. Based on the contextual information in the accident report, a 3D scene reconstruction is performed to generate a high-quality background image. The reconstruction process includes two stages: viewpoint adjustment and background rendering.

[0019] (a) Viewpoint adjustment: Parse the viewpoint information in the context information to generate the corresponding position and angle transformation matrix.

[0020] (b) Background rendering: Multi-camera alignment and lighting consistency rendering techniques are used to ensure seamless matching between the background image and the vehicle's viewpoint. If a new viewpoint is required, a scene reconstruction method (such as NeRF) is used; otherwise, the original background frame is used.

[0021] Step S5: Foreground and Background Blending. Using channel compositing technology, the foreground vehicle trajectories are blended with the background image, with precise adjustments to transparency and lighting consistency to generate a visually coherent traffic simulation video. The blending process ensures a high degree of spatial and temporal coordination between the dynamic trajectories of the foreground vehicles and the background scene.

[0022] Step S6: Video Output. Output the final traffic simulation video. The video follows the dynamic laws of the real world in terms of physical realism and presents high-quality scene details in terms of visual realism. It is widely applicable to high-fidelity application scenarios such as autonomous driving testing, traffic flow analysis, and intelligent transportation system development.

[0023] The method of this invention can significantly improve the diversity and realism of traffic scenarios, reducing reliance on real data collection. The traffic simulation videos generated by this method can effectively compensate for the lack of scarce scenario data in autonomous driving testing, support more accurate modeling in traffic flow analysis, and are particularly suitable for intelligent transportation applications requiring high-fidelity traffic simulation.

Claims

1. A traffic scenario simulation method based on accident reports, characterized in that, The method comprises the following steps: (a) Extracting physical clues from real traffic accident reports, including vehicle speed, collision angle, vehicle mass, and contextual information such as road type and environmental conditions; (b) Using a physical simulation engine to process the extracted physical clues and contextual information to generate a high-fidelity vehicle motion trajectory dataset; (c) Generating dynamic traffic scenarios based on the vehicle motion trajectory dataset through pre-collision trajectory planning algorithms; (d) Using a large language model to optimize trajectory prediction for the dynamic traffic scenarios, ensuring that the simulated scenarios meet real-world physical constraints; (e) Using three-dimensional scene reconstruction techniques to generate high-quality background images; (f) Generating visually and physically realistic traffic simulation videos by fusing foreground vehicle trajectories with background images.

2. The traffic scenario simulation method according to claim 1, characterized in that, The physical simulation engine combines vehicle dynamic parameters (such as friction, gravity, etc.) and road characteristics (such as slope, surface properties, etc.) to ensure that the generated vehicle motion trajectories are physically realistic.

3. The traffic scenario simulation method according to claim 1, characterized in that, The extraction of physical clues uses a large language model combined with a few-shot learning strategy to automatically analyze pre-collision environmental conditions, collision dynamics, and post-collision results in accident reports, improving the accuracy and adaptability of information extraction.

4. The traffic scenario simulation method according to claim 1, characterized in that, The pre-collision trajectory planning includes the following sub-steps: (a) Initial state extraction, analyzing vehicle dynamic information and collision parameters to determine the starting position and collision target of each vehicle; (b) Lane and path selection, generating candidate lane combinations that meet collision constraints based on scene maps and vehicle information; (c) Trajectory generation and verification, generating dynamic feasible pre-collision trajectories based on selected lane combinations and verifying that they meet collision standards.

5. The traffic scenario simulation method according to claim 1, characterized in that, The three-dimensional scene reconstruction step uses multi-camera alignment and lighting consistency rendering techniques to ensure that the generated background images seamlessly match the vehicle perspective and environmental conditions.

6. The traffic scenario simulation method according to claim 1, characterized in that, The foreground and background fusion step uses channel synthesis technology to accurately adjust the transparency between the foreground vehicle trajectory and the background image to generate visually coherent traffic simulation videos.

7. The traffic scenario simulation method according to claims 1 to 4, characterized in that, The large language model is fine-tuned using a fine-tuning method, using collision trajectory datasets based on accident reports to optimize trajectory prediction accuracy.

8. The traffic scenario simulation method of any one of claims 1 to 7, characterized in that, This method is suitable for applications such as autonomous driving testing, traffic flow analysis, and intelligent transportation system development that require high-fidelity traffic simulation, reducing the reliance on real data collection while ensuring the diversity and physical realism of simulated scenarios.

Images