Driver drowsiness detection using YOLO based deep learning models

Abstract

The incidence of traffic accidents in Indonesia has been escalating, predominantly attributed to human factors such as fatigue and drowsiness. This study presents the implementation of a deep learning-based drowsiness detection system utilizing the you only look once (YOLO) architecture to enhance vehicular safety. Three YOLO model variants (YOLOv5, YOLOv8, and YOLOv10) were evaluated using a dataset comprising 1,000 annotated images across four classes: alert, low vigilance, drowsy, and microsleep. A quantitative experimental methodology was employed, with performance assessed through precision, recall, accuracy, and F1-score metrics. Experimental results demonstrate that YOLOv8 (medium and small variants) achieved superior overall performance, exhibiting a balanced optimization across all evaluation metrics. YOLOv5 yielded the highest recall, suggesting its suitability for comprehensive detection tasks, whereas YOLOv10 demonstrated enhanced computational efficiency without significant performance degradation. Based on these findings, YOLOv8 is recommended as the most effective model for real-world deployment, while YOLOv5 and YOLOv10 offer viable alternatives depending on specific operational requirements. This study contributes to the advancement of early warning systems for driver drowsiness detection, with the broader aim of mitigating traffic accident risks.