Research on Traffic Accident Risk Prediction Method Based on Spatial and Visual Semantics

Abstract

Predicting traffic accidents involves analyzing historical data, determining the relevant factors affecting the occurrence of traffic accidents, and predicting the likelihood of future traffic accidents. Most of the previous studies used statistical methods or single deep learning network model prediction methods while ignoring the visual effects of the city landscape on the drivers and the zero-inflation problem, resulting in poor prediction performance. Therefore, this paper constructs a city traffic accident risk prediction model that incorporates spatial and visual effects on drivers. The improved STGCN model is used in the model, a CNN and GRU replace the origin space–time convolution layer, two layers of a GCN are added to extract the city landscape similarity of different regions, and a BN layer is added to solve the gradient explosion problem. Finally, the features extracted from the time–space correlation module, the city landscape similarity module and the spatial correlation module are fused. The model is trained with the self-made Chicago dataset and compared with the existing network model. The comparison experiment proves that the prediction effect of the model in both the full time period and the high-frequency time period is better than that of the existing model. The ablation experiment proves that the city landscape similarity module added in this paper performs well in the high-frequency area.