Kullback-Leibler and Attention-guided Feature Distillation for Hazardous Driving Detection and a Unified Benchmark

Hazardous driving is the main cause of traffic accidents and thus fast and efficient detection of hazardous driving is meaningful. Various hazardous driving detection schemes have been presented, and the real-time performance need to be improved while it is cumbersome to deploy. In this paper, a Kullback-Leibler (KL) and attention-guided feature distillation method (KLA-FD) for hazardous driving detection are newly proposed. Compared with previous methods, we propose to obtain a soft probability map for each feature map after an attention module and normalization process. By simply minimizing the KL scatter between the soft probability maps of the two networks, the distillation process pays more attention to the most significant region of each feature map, which is valuable for hazardous driving detection. The experimental results show that compared with the original YOLOv8 network, our method shows better performance with higher speed and accuracy requirements.