Accident LLM: A multimodal large language model of accident description based on Inter-Frame Attention Mechanism

Addressing the technical bottlenecks in traffic accident video analysis and liability determination, this study proposes Accident LLM—a multi-modal large language model based on an inter-frame attention mechanism. This model generates interframe feature vectors by calculating the local correlation of adjacent frames, significantly reducing the complexity of spatial-temporal modeling, and optimizes the positional encoding strategy to enhance the recognition ability of spatial relationships such as traffic signs and vehicle trajectories. Experiments show that the model outperforms existing methods on various metrics (BLEU-4, METEOR, etc.) on the self-built accident datasets and demonstrates good generalization on general datasets (TVSum, MSRVTT). The research provides an efficient technical framework for intelligent liability determination systems, and future service closed-loop can be further optimized through legal knowledge injection and interactive dialogue.