CAFFT: Cross-Attention Feature Fusion Transformer for Risky Driving Behavior Recognition

According to the results of a survey administered by the World Health Organization, the primary cause of the majority of traffic accidents is risky driving behavior. Consequently, the identification of risky driving behaviors is imperative for ensuring road safety. In order to effectively address the limitations of existing methods in modeling long-term dependencies, integrating external scenario information, and adapting to different driver personalities, this paper proposes a Transformer-based cross-attention feature fusion model for risky driving behavior recognition, referred to as CAFFT. This model utilizes a hierarchical Transformer encoder structure to capture fine-grained changes and long-term dependencies in driving behavior. Furthermore, the cross-attention mechanism, a cross-modal attention mechanism, is introduced to effectively fuse driving behavior features with weather-related external scenario features, thereby rendering the research scenario more aligned with real-world driving conditions. Finally, through cross-validation and cross-test set generalization experiments, the universality and stability of the method are verified. The experimental results on real driving datasets demonstrate that the CAFFT model achieves substantial performance enhancements in recognizing risky driving behavior. This improvement leads to enhanced accuracy and reliability in recognizing risky driving behavior, thereby providing substantial support for the development of intelligent driving assistance systems.