FAMT: Fusion of Feature Attention Mechanisms and Multiscale Temporal Relationships for Traffic Accident Prediction

Travel safety continues to be a major concern, with traffic accidents having a significant negative impact on both society and individuals. It is essential to develop a method for accurately predicting traffic accidents. However, previous studies are deficient in several ways: model learning only considers information at global scale and ignores implicit temporal relationships at multiple scales; the frequency of traffic accidents is significantly influenced by weather conditions, traffic dynamics, and domain-specific events. Previous studies have failed to collect comprehensive datasets from multiple sources or to enable models to effectively incorporate the influence of these external factors on traffic accidents. To address the above limitations, this study proposes a traffic accident prediction approach that incorporates feature attention mechanisms and multiscale time series relationships (FAMT). The proposed FAMT has the following two distinctive characteristics: (i) In the time domain, explicit time series relationships are not only mined, but also implicit temporal relationships are captured through multiscale modelling; (ii) In the feature domain, the extraction of the effects of multiple heterogeneous data sources on traffic accidents is accomplished through the use of a feature attention mechanism. The effectiveness of the proposed method is confirmed by experiments on Manhattan and Los Angeles datasets.