Adaptive Context Based Road Accident Risk Prediction Using Spatio-Temporal Deep Learning

Traffic accidents are common urban events that pose significant risks to human safety, traffic management, and economic stability; consequently, the research community is paying increasing attention toward accident risk prediction. However, accident risk prediction is a challenging problem because accident occurrences are sparse and influenced by multiple contextual factors (e.g., POI, road structure, road type, hour of the day, and month). Therefore, in this article, we propose a novel architecture named Topographic-Weighted Context Category (TWCCnet) that adapts heterogeneous contextual category weights based on spatial–temporal correlations across sectors. Specifically, the framework consists of two parallel components: one uses convolution and stacked bidirectional gated recurrent unit (Bi-GRU) to capture spatial–temporal relationships between neighborhood sectors, while the other uses multiple graph convolution network (GCN) over resemblance graphs to capture spatial–temporal relationships between semantic sectors. At last, temporal attention is utilized on top of parallel components to learn important spatiotemporal features that have a substantial impact on traffic accidents. The extensive experiments on two publicly available citywide datasets, i.e., New York City and Chicago demonstrate the effectiveness of the proposed approach and showed
(7.21, 10.7)
Root Mean Squared Error (RMSE),
(34.09, 20.75)
Mean Average Precision (MAP), and
(0.19, 0.09)
Recall approximately over both datasets, respectively, outperforming baseline as well as state-of-the-art models.