AccidentSim: Generating Vehicle Collision Videos with Physically Realistic Collision Trajectories from Real-World Accident Reports

The research team developed the AccidentSim framework, which uses physical clues from real-world accident reports to generate vehicle collision videos with physical authenticity. The framework first replicates the vehicle trajectories after the collision using a reliable physics simulator based on accident report information and constructs a dataset; then it fine-tunes the language model using this dataset to predict physically consistent trajectories in different scenarios; finally, it combines NeRF rendering for high-quality background and foreground vehicles to generate the final videos. Experimental results show that the videos generated by AccidentSim excel in both visual and physical authenticity.