Learning physically grounded traffic accident reconstruction from public accident reports

Accident reconstruction typically requires detailed scene measurements and expert analysis — costly and hard to scale. A team of researchers now proposes turning this into a parameterized multimodal learning problem using only publicly accessible crash reports and basic scene measurements. They built CISS-REC, a dataset of 6,217 real-world accident cases from the NHTSA Crash Investigation Sampling System. Their framework first extracts road topology and participant attributes from text, then reconstructs lane-consistent pre-impact motion, and refines collision interactions via localized geometric reasoning and temporal allocation. On CISS-REC, the method consistently outperforms representative baselines, achieving the strongest overall reconstruction fidelity — including improved accident point accuracy and collision consistency.

This work shows that accident reports — already routinely collected — can serve as scalable inputs for quantitatively verifiable reconstruction. The implications span traffic safety analysis (e.g., identifying common crash patterns), simulation (e.g., generating realistic accident scenarios for testing), and autonomous driving research (e.g., recreating edge cases to improve decision-making). By reducing reliance on expensive manual reconstruction, the approach could accelerate safety improvements across the transportation ecosystem.