A multi-regional spatio-temporal network for traffic accident risk prediction

Traffic accident risk prediction is a cornerstone of intelligent mobility, and the main challenge is adequately capturing dynamic spatial and temporal characteristics. However, most existing approaches ignore two critical features of traffic accident risk prediction. First, they introduce traffic conditions only at strictly periodic times, thus weakening the role of temporal continuity. Second, most methods consider the spatio-temporal characteristics of local or global regions, ignoring the effects of multi-regional convergence. Considering the aforementioned issues, we put forward a model of fused multi-regional spatio-temporal characteristics under periodic translation (ATCGCN). The model fuses adjacent and multi-adjacent regions to accurately predict road accident risk by capturing dynamic spatio-temporal correlations. Inspired by the first feature, this paper combines traffic data from adjacent and cycle panning periods, captures their temporal characteristics through GRU, and uses the attention mechanism to characterize the space-time features of multi-adjacent regions dynamically. To achieve the second feature, ATCGCN introduces CNN and GCN to capture the spatial correlations of adjacent and multi-adjacent regions, respectively, and then fuses the spatial and temporal interdependence of the two different regions to enhance the accuracy of traffic accident risk prediction effectively. Through executing trials on a pair of actual datasets from the real world, we highlight the efficacy of each component in ATCGCN and demonstrate that its predictive performance surpasses that of several existing approaches. © 2023, International Association of Engineers. All rights reserved.