A Comparative Study of YOLO-Family Models for Driver Eye-State Detection under Mask and Glasses Wearing Conditions

Driver drowsiness is a major contributing factor to traffic accidents, and accurate and timely detection is crucial for enhancing road safety. With recent progress in deep learning-based object detection, YOLO-based models have been widely adopted for real-time driver condition monitoring using in-vehicle cameras. This paper does not directly classify driver drowsiness; instead, it focuses on robust eye-state (open/closed) detection, a key indicator for drowsiness assessment. To detect the driver’s eye state under diverse conditions (e.g., mask wearing, glasses wearing, and variations in facial angle), several YOLO-based models were comparatively evaluated under identical experimental conditions. YOLOv5, YOLOv7, YOLOv8, and YOLO-Former were employed and evaluated on publicly available image datasets. The trade-off between detection accuracy and real-time performance was systematically analyzed using mAP, precision, and recall. Experimental results indicated that performance was strongly affected by environmental complexity, and that lightweight models could achieve competitive accuracy while enabling faster inference. These findings provide practical insights into selecting an appropriate YOLO architecture for real-time driver drowsiness detection systems.