Lightweight and Real-Time Improvement of Yolov8n Model for Driving Fatigue Detection

To address the challenges of high computational cost, large parameter requirements, and insufficient facial keypoint extraction capability in fatigue driving detection models, this paper proposes an optimized algorithm based on an improved YOLOv8n. A lightweight FCOS architecture is integrated into the Detect head, leveraging shared convolution to significantly reduce the number of parameters while enhancing object localization and classification performance, thereby lowering computational overhead. The model was trained for 100 epochs under the same hardware configuration. Experimental results show that the improved model achieves a mean average precision (mAP) of 80.5 %, representing a notable improvement over the baseline model. Additionally, the computational cost (GFLOPS) and model size (MB) are reduced by 19.7 % and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2 2. 2 \%}$</tex>, respectively, while latency decreases by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 0. 2 \%}$</tex>, and detection speed increases by 148 FPS. These results demonstrate that the proposed algorithm holds significant application value in driver facial object detection tasks.