Traffic Accident Risk Prediction Method of Urban Road Network Based on Multi-source Spatiotemporal Data

Abstract

Traffic accident risk prediction is used to study the historical accidents, identify the relevant factors and predict accident risk in the future. The existing prediction methods mainly obtain predicted unit by regularly gridding the road areas, resulting in a decrease in accuracy and low practical value. To improve the prediction accuracy, this paper takes urban roads as the prediction unit, adopts graph convolution neural network and gated recursive unit, and proposes a spatiotemporal gated graph convolutional neural network model (STGG-CnovNet) fusing multi-source spatiotemporal data features. The model consists of spatial convolution, temporal convolution, and spatiotemporal convolution. In the spatial convolution module, the spatiotemporal map data is constructed, and the spatial correlation is captured. In the temporal convolution module, a gated cycle unit is used to model the time correlation of traffic accidents. In the spatiotemporal convolution module, constructing a road similarity map can capture the spatiotemporal correlation of nodes. On real datasets, experimental results demonstrate our method is better than other baselines.