Title: TrafficRAG: A Multimodal RAG Framework for Traffic Accident Liability Determination

Abstract:

Determining liability in traffic accidents remains a pivotal yet complex challenge within the realms of intelligent transportation systems and legal support services. Current methodologies frequently struggle with inefficiency, subjective interpretations, and inconsistent outcomes. Furthermore, the application of large language models in this domain is hindered by the presence of noisy video data and a lack of specialized legal expertise. To overcome these limitations, this study introduces TrafficRAG, a multimodal retrieval-augmented generation system designed for the automated analysis of traffic incidents and the subsequent generation of reports.

The proposed architecture utilizes a vision-language model to convert accident scenarios into structured textual descriptions, which function as precise queries for information retrieval. Leveraging these textual inputs, the system employs a hybrid retrieval mechanism that combines BM25 sparse retrieval with dense embedding techniques to locate pertinent traffic regulations and analogous historical case studies. Subsequently, a large language model synthesizes the retrieved legal information alongside multimodal accident evidence to perform comprehensive reasoning, ultimately producing standardized liability analysis reports that are firmly grounded in legal principles.

Extensive experimental evaluations demonstrate that TrafficRAG consistently surpasses baseline approaches. The framework achieved a Legal Norm Adaptation Accuracy of 77.32%, a Factual Faithfulness score of 81.71%, and a Liability Ratio Mean Absolute Error (MAE) of 5.48%. These findings confirm that augmenting multimodal factual evidence with legal clauses through retrieval mechanisms significantly enhances the reliability and precision of traffic accident liability determinations.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC