Abstract
Driver drowsiness remains a critical factor in road accidents, accounting for thousands of fatalities and injuries each year. This paper presents a comprehensive evaluation of real-time, non-intrusive drowsiness detection methods, focusing on computer vision based YOLO (You Look Only Once) algorithms. A publicly available dataset namely, UTA-RLDD was used, containing both awake and drowsy conditions, ensuring variability in gender, eyewear, illumination, and skin tone. Seven YOLO variants (v5s, v9c, v9t, v10n, v10l, v11n, v11l) are fine-tuned, with performance measured in terms of Precision, Recall, mAP0.5, and mAP 0.5-0.95. Among these, YOLOv9c achieved the highest accuracy (0.986 mAP 0.5, 0.978 Recall) while YOLOv11n strikes the optimal balance between precision (0.954) and inference efficiency, making it highly suitable for embedded deployment. Additionally, we implement an Eye Aspect Ratio (EAR) approach using Dlib's facial landmarks, which despite its low computational footprint exhibits reduced robustness under pose variation and occlusions. Our findings illustrate clear trade offs between accuracy, latency, and resource requirements, and offer practical guidelines for selecting or combining detection methods in autonomous driving and industrial safety applications.
Abstract (translated)
驾驶员疲劳依然是道路事故中的一个关键因素，每年导致数千人死亡和受伤。本文对实时、非侵入式的疲劳检测方法进行了全面评估，重点关注基于计算机视觉的YOLO（You Look Only Once）算法。我们使用了一个公开的数据集UTA-RLDD，该数据集中包含了清醒和困倦两种状态，并且在性别、眼镜佩戴情况、光照条件以及肤色上具有多样性。我们对七种YOLO变体（v5s, v9c, v9t, v10n, v10l, v11n, v11l）进行了微调，性能评估指标包括精度、召回率、mAP0.5和mAP 0.5-0.95。在这其中，YOLOv9c在准确性上表现出最佳（mAP 0.5为0.986，Recall为0.978），而YOLOv11n则在精度（Precision为0.954）和推理效率之间达到了最优平衡，因此特别适合嵌入式部署。此外，我们还实现了一种基于Dlib面部特征点的Eye Aspect Ratio (EAR) 方法，尽管其计算开销较低，但在姿态变化和遮挡情况下表现出了较差的鲁棒性。我们的研究结果表明了在准确性、延迟性和资源需求之间存在明显的权衡，并为自动驾驶及工业安全应用中选择或结合检测方法提供了实用指南。
URL
https://arxiv.org/abs/2509.17498
