A Traffic Accident Prediction Model Based on Spatio-Temporal Depth Map and Transformer

As a real-time data sensor in vehicle networks, dash cams capture video with both static (object positions, scene layouts) and dynamic (motion paths, speed changes) spatio-temporal data. Analyzing this information helps predict scene changes and prevent accidents. To improve spatio-temporal feature representation and prediction accuracy, we propose a traffic accident prediction model based on spatio-temporal depth maps and Transformer, named the STDMP model. The model creates 3D depth maps using monocular estimation, aligns spatial-temporal features through two-stream architecture, and integrates GRU's sequential processing with Transformer's attention mechanism for precise short-term tracking and long-term dependency modeling. Experimental data shows the STDMP significantly outperforms baseline models on the DAD dataset. The key metrics (AP:64.3, mTTA:4.25) confirm the effective combination of two techniques: depth-based spatial modeling and Transformer's temporal feature capture.