TRACER Reconstructs Traffic Accidents with Training-Free AI

Traffic accident reconstruction is a complex forensic task requiring the recovery of physically consistent motion from limited evidence. Traditional learning-based methods often prioritize semantic plausibility or visual realism, sometimes at the expense of quantitative accuracy. This new framework, TRACER, approaches the problem differently by formulating it as a closed-loop structured inference process. Instead of directly generating full trajectories, TRACER constructs and refines event-anchored motion hypotheses. It operates under strict geometric, kinematic, and interaction constraints, guided by a structured case memory and a consistency-driven diagnostic system. This design allows for incremental, interpretable corrections, mirroring how human experts might approach such a task. Evaluations using real-world accident data demonstrate that TRACER significantly enhances geometric fidelity, velocity consistency, and collision accuracy. It outperforms both data-driven and physics-based baselines, offering a more robust and interpretable solution for forensic analysis.