CrashRecast: Agentic Crash Reconstruction from Accident Reports via World Models

Overview

Real crashes are documented in written accident reports rather than in sensor data. CrashRecast reconstructs those crashes as photorealistic driving scenes, so that safety-critical events from the real world can be replayed and studied.

Contributions

From accident reports to crash trajectories

An agentic workflow infers initial and collision positions from NHTSA/TIMS accident reports and road topology. Through traffic simulation and iterative agent review, it produced approximately 3,500 valid crash trajectories for scene reconstruction.

From trajectories to photorealistic scenes

A pipeline converts crash trajectories into photorealistic 4DGS scenes: it reconstructs crash sites from location-matched street-view imagery, inserts prepared vehicle assets along the inferred trajectories, and refines lighting and appearance with diffusion priors.

Context

Co-first author, developing the visual reconstruction and video enhancement pipeline. The paper is ready for submission to CVPR 2027.

Advisor: Prof. Henry Liu, Mobility Transformation Lab, University of Michigan. Jul 2026 – present.