Research on multi-feature fusion fatigue driving detection algorithm based on improved YOLOv8

Fatigue-driving detection based on facial cues is challenging because the key fatigue-related regions, especially the eyes and mouth, are small, easily affected by illumination and occlusion, and must be detected in real time. To address insufficient small-target feature extraction and the trade-off between detection accuracy and model complexity, this paper proposes FWC-YOLOv8, an improved fatigue-driving detection framework built on YOLOv8s. First, an additional 160×160 detection head is introduced to preserve high-resolution spatial information and improve localization of small eye and mouth regions. Second, a Coordinate Attention (CA) module is embedded in the backbone to enhance position-aware feature representation. Third, the feature-fusion structure is refined to strengthen the interaction between shallow spatial details and deep semantic information. Finally, Wise-IoU, which uses a dynamic focusing mechanism, is adopted as the bounding-box regression loss to reduce the influence of low-quality samples and improve convergence. Experiments on the GX_FatigueDriving2020 dataset show that FWC-YOLOv8 achieves an mAP@0.50:0.95 of 0.640, corresponding to a 4.2 % improvement over the original YOLOv8s baseline. The model contains 17.0 M parameters and requires 35.5 GFLOPs, indicating a favorable balance between accuracy and computational cost. In addition, detected eye and mouth states are further converted into four fatigue indicators, namely PERCLOS, blink frequency, continuous eye-closure duration, and yawn frequency, and fused through a weighted multi-index assessment strategy to classify driver fatigue into alert, mild-fatigue, and severe-fatigue states. The framework combines task-specific facial-state detection with interpretable fatigue-level assessment.