Traffic Accident Risk Forecasting using Contextual Vision Transformers with Static Map Generation and Coarse-Fine-Coarse Transformers

We propose an enhancement to our previously proposed novel model called Contextual Vision Transformer (ViT) to address the problem of traffic accident risk forecasting. This framework combines spatial and temporal information using a data-driven approach. By treating the problem as a computer vision task, we can predict traffic accident risk as the next frame in a video sequence. Specificaly, we extend the ViT network with a Static Map generation (named XViT) for even better results on the Chicago dataset. Furthermore, we propose a Coarse-Fine-Coarse transformer architecture as an alternative approach to enhance traffic accident risk prediction.