Research on a Lightweight Fatigue Driving Detection Algorithm Under Different Angles

To achieve both high real-time performance and detection accuracy in fatigue driving recognition, this study introduces an improved object detection model named FEL-YOLO11. The model incorporates three key innovations. First, FasterNet is integrated with the YOLO11 backbone, and an EMA mechanism is applied to enhance multi-scale feature fusion and attention modeling. This design enhances feature representation and robustness under complex illumination conditions while maintaining low computational overhead. Second, a novel C3K2-Faster neck module, constructed with Faster Block, replaces the traditional structure, significantly improving feature extraction efficiency and network performance. Third, a lightweight DELSCD detection head is proposed, which integrates cross-layer spatial and semantic features, employs Detail Enhancement Convolution (DEConv), and adopts a convolution sharing strategy. These improvements facilitate richer feature interactions across layers and enhance fine-grained representation, especially for small objects in complex scenarios. Experimental evaluations demonstrate that FEL-YOLO11 not only maintains high detection precision but also reduces computational cost and achieves superior robustness compared with mainstream models, highlighting its strong potential for fatigue driving detection.