A Keypoint‐Assisted Deep Attention Network With Adaptive Weighting for Distracted Driving Behaviour Detection

Distracted driving behaviour is the main cause of traffic accidents, posing significant safety risks and economic losses. To address this effectively, this paper proposes a keypoint‐assisted deep attention Network (KDAN‐AW) for recognising distracted driving behaviour. The model first incorporates the driver's keypoint features as an auxiliary input, providing prior structural information and a holistic understanding about the driver's actions. Second, an attention module is embedded into the deeper layers to enhance the cognition of the image's global context. Furthermore, to address the imbalance between hard and easy samples, we adaptively increase the weight of hard samples during training by a modulator factor, forcing the model to achieve a better learning equilibrium. Experimental results on the constructed SAA13 dataset demonstrate that the proposed KDAN‐AW achieves a recognition accuracy of 90.70% while maintaining a processing speed of 85.4 frames per second (FPS) and outperforms other state‐of‐the‐art (SOTA) models, validating the model's potential and practical value for real‐time and accurate driver behaviour recognition in intelligent transportation systems.