Traffic accident severity prediction and analysis via spatio-temporal deep learning: A ConvLSTM-transformer approach

Traffic accidents pose a significant challenge to public safety, and predicting their causation and severity is crucial to traffic safety research. Traditional machine learning methods, however, mainly depend on direct records from traffic datasets, therefore ignoring the spatial and temporal heterogeneity of traffic accidents. Unfortunately, most existing deep learning algorithms, which excel at extracting spatiotemporal features of accidents, often overlook semantic features like driver and vehicle characteristics and fail to differentiate accident severity levels. To address the limitations of existing models, we propose Spatio-Temporal RiskFormer (ST-RiskFormer), a deep learning model designed to accurately predict traffic accident severity (slight, serious, fatal) by integrating ConvLSTM for capturing spatio-temporal correlations with a Transformer-based module for processing semantic features, including driver demographics and vehicle attributes. Moreover, Gradient-weighted Class Activation Mapping (Grad-CAM) is applied to data visualization for ST-RiskFormer. Simulation results show that medium-term traffic data can uncover the most significant spatial-temporal information of traffic accidents. More importantly, as compared to other typical machine learning algorithms, ST-RiskFormer, which highlights the spatio-temporal distribution of traffic accidents, performs better in predicting traffic accident severity. We hope to provide some insights into deeper-level traffic data mining through this work.