Skip to main content
Back

TrafficRAG Turns Dashcam Footage Into Legally Grounded Accident Reports

TrafficRAG combines dashcam video understanding with legal knowledge retrieval to automatically assign fault in traffic accidents, hitting 77% legal norm accuracy without any fine-tuning.

Paper, Full Access
  • TrafficRAG is a new multimodal RAG framework that turns dashcam video into legally grounded accident liability reports automatically.
  • A VLM converts video frames into structured text descriptions, which are then used as retrieval queries -- avoiding direct video-to-vector search.
  • Hybrid BM25 + dense retrieval pulls from both a traffic regulation corpus and a judicial case database, then reranks across sources.
  • Results: 77.32% Legal Norm Accuracy, 81.71% Factual Faithfulness, and a fault-ratio error of just 5.48% MAE, outperforming all baselines.
  • Key limitation: system quality depends on VLM scene description accuracy; low-quality or occluded footage degrades the entire pipeline.
  • No code or models released; the paper is available at arxiv.org/abs/2606.01737.

Determining who is at fault after a traffic accident is a slow, expensive, and often inconsistent process. Traffic police, insurance adjusters, and legal teams must manually review footage, cross-reference traffic regulations, and dig through precedent cases -- all while under time pressure and with room for subjective bias. A new paper introduces TrafficRAG, a multimodal Retrieval-Augmented Generation (RAG) framework that automates this entire pipeline: from raw dashcam video to a structured, legally grounded liability report.

The gap nobody closed

Vision-language models (VLMs) have gotten remarkably good at describing what happens in a traffic video. But describing a scene and legally adjudicating it are two very different things. Efficiently and reliably converting accident videos into responsibility attribution and concrete statutory provisions has become a key challenge for traffic police and courts -- the sheer scale and diversity of cases makes manual analysis and statute lookup extremely costly.

Existing methods mostly remain at the perceptual or semantic level -- answering "what type of accident is this?" -- and have not been elevated to the legal level of responsibility attribution and statute application. Meanwhile, legal LLMs are almost entirely built on single-modality "case texts + statutes", without incorporating video evidence into the legal reasoning process.

This creates a clear gap: video models understand the scene but not the law; legal models know the law but cannot watch the footage. TrafficRAG is designed to close that gap.

TrafficRAG framework flowchart showing dual-path retrieval from traffic regulations and judicial case knowledge bases, through structured accident description, to generated liability report

How it works, step by step

The pipeline has three major stages that convert raw video into a legal report:

  1. Video-to-text translation. A vision-language model processes sampled frames using CLIP feature encoding, then performs temporal phase parsing to understand the accident sequence. The output is a structured textual description of the accident scenario -- not a vague caption, but a precise, machine-readable summary of what happened, who was involved, and how.