Utilization of YoloV8 Algorithm for in-Vehicle Video-Based Driver Monitoring System

Abstract

Driving safety is a critical factor in reducing the risk of traffic accidents, which are often caused by unsafe driver behavior such as drowsiness, phone usage, or neglecting to wear seat belts. To address this, a real-time driver monitoring system is needed to detect and identify risky behaviors using the YOLOv8 algorithm. This study utilizes a secondary dataset titled “DMS Driver Monitoring System” from Kaggle, comprising 9,440 images of various driver behaviors. The dataset underwent preprocessing, including resizing 640x640 pixels and data augmentation to increase image diversity. The YOLOv8 model was trained for 100 epochs with a data split of 70% training, 20% validation, and 10% testing. Performance was evaluated using precision, recall, F1-score, and mean Average Precision (mAP). Results showed that the model achieved 89.6% precision, 87.2% recall, 88.0% F1-score, and 92.0% mAP50. The mAP50–95 score of 69.1% indicates room for improvement in more complex detection scenarios. Real-time video testing revealed the model could detect open eyes with 85% confidence and seat belt use with 35% confidence. The study concludes that YOLOv8 is effective for standard behavior detection but requires further optimization to handle varying lighting and camera angles for broader real-world deployment.