Driver Fatigue Detection Method Based on Improved RT-DETR

To address the problem of poor robustness in driver fatigue detection under complex scenarios such as low illumination and motion blur, this paper proposes an optimization method based on hard sample augmentation. A hard sample dataset containing low illumination, motion blur, and their compound interference is constructed, and a progressive two-stage training strategy is designed to guide the model to adapt to various degradations step by step. Furthermore, a spatial-channel interleaved robustness module (SCIRM) is proposed to enhance the response of eye and mouth features, while the normalized Wasserstein distance (NWD) loss function is adopted to improve the localization accuracy of small targets. The improved RT-DETR model achieves mAP50 of 95.03%, 95.62%, and 94.56% under low illumination, motion blur, and compound interference scenarios, respectively, significantly outperforming the baseline model, while its performance on clear scenes remains almost unchanged. This method effectively enhances the robustness of fatigue detection under complex imaging conditions.