Driver Distraction Detection Based on Large Kernel Convolution and Attention Mechanism

ABSTRACT Vision‐based driver distraction detection plays an important role in improving road safety. However, existing methods often suffer from a suboptimal trade‐off between recognition accuracy and computational complexity, which limits their deployment on resource‐constrained in‐vehicle platforms. To address this issue, this paper proposes a lightweight hybrid detection framework, termed YOLOv8‐LarK‐CAFM‐RTDETR, which enhances global contextual perception and long‐range feature dependency while maintaining computational efficiency. Specifically, large‐kernel convolution (LarK) is introduced into the backbone to expand the effective receptive field, while a convolutional attention fusion module (CAFM) is integrated into the feature fusion stage to strengthen feature representation. In addition, an real‐time detection transformer (RT‐DETR) decoder with an auxiliary prediction head is employed to improve query optimization and classification robustness. Extensive experiments on the state farm distracted driver detection dataset demonstrate that the proposed model achieves an accuracy of 95.07% and a precision of 95.85%, outperforming the baseline YOLOv8n by 5.16% and 5.07%, respectively. Moreover, the proposed framework remains lightweight, requiring only 6.5 M parameters and 12.5 GFLOPs, indicating its suitability for practical deployment in computationally limited in‐vehicle systems. These results indicate that the proposed method provides an effective and extensible solution for driver distraction behavior recognition and has strong potential for practical in‐vehicle vision‐based safety systems.