A deep spatiotemporal network for forecasting the risk of traffic accidents in low-risk regions

Junjie Zheng, Jing Wang, Zhilin Lai, Cheng Wang, Huizhen Zhang

Abstract: It is admirably significant to forecast real-time traffic risks in the future, which protects people’s lives and improves the safety of the road. Most previous work uses the grid method to divide the entire city, which destroys the city’s inherent geospatial attributes and may lead to invalid prediction results. On the issue of sparse accidents, although previous studies have considered the imbalance in the number of accidents and normal events, the imbalance in the number of accidents between different regions caused by inner-city heterogeneity is ignored, resulting in unsatisfactory predictions for low-risk regions. Moreover, such a model is not suitable for citywide traffic accident prediction. To solve the above problems, firstly, we combine taxi division map, census track partition map and road network to divide the entire city, which retains the geographical spatial attributes, makes the regional division more reasonable and interpretable, and avoids the possible invalid prediction caused by grid method. Secondly, we propose the concept of double imbalance in traffic accident data, which is addressed by an improved cost-sensitive loss function, enabling the model to better predict accidents in low-risk regions. Finally, a deep spatiotemporal network that fuses local and global features (DSTFLG) based on a self-attention mechanism is proposed to forecast traffic accident risk. Extensive experiments on two real-world datasets demonstrate that the proposed framework improves the prediction accuracy over baseline approaches.