Dynamic spatiotemporal graph network for traffic accident risk prediction

Abstract

Traffic accidents remain major public safety concerns, often causing severe injuries, deaths, and economic costs, especially in rapidly urbanizing areas. Accurate traffic accident risk prediction is crucial for developing effective strategies to reduce accidents and enhance urban mobility. However, predicting traffic accident risks is challenging due to the relationships among factors such as weather, traffic conditions, and road characteristics, along with capturing spatial correlations of traffic accidents across different time scales. To address these challenges, we propose the dynamic spatial-temporal accident risk network (DSTAR-Net). Our model uses channel-wise convolutional neural networks to detect spatial accident patterns across weekly, daily, and hourly time scales with automatic weight learning, simultaneously employing graph convolutional networks to process road network features, population feature while integrating external data like weather and dates. The dynamic learning of spatial correlations, combined with the integration of road characteristics and contextual variables, significantly enhances the accuracy of traffic accident predictions. Experiments in Perth show the DSTAR-Net outperforms state-of-the-art models with RMSE 24.901, Recall 21.59%, and MAP 0.0721. Notably, the weights learned by our model indicate that hourly patterns have the highest weight at 0.390, while weekly trends carry the lowest weight at 0.255, suggesting that recent traffic conditions have the most significant influence on accident risks. This study provides a foundational framework for predicting traffic accident risks, aiding urban planners and policymakers in enhancing road safety and traffic management in cities.