Real-Time Multi-Cue Driver Drowsiness Detection Using YOLOv8: A Deep Learning Framework for Intelligent Vehicle Safety

Abstract

This study presents a real-time driver drowsiness detection system using a YOLOv8-based architecture. The proposed framework continuously checks the driver's facial region to detect critical fatigue indicators, specifically eye closure (open/closed states) and yawning. A custom dataset was curated and annotated for these specific actions. The model was initialized with a pre-trained YOLOv8-nano backbone and then fine-tuned for 30 epochs on our dataset, employing data augmentation techniques including horizontal flipping, brightness variation, and Gaussian noise to enhance robustness. The data was partitioned into 70% for training, 15% for validation, and 15% for testing. The system proved high performance on the test set, achieving a mean Average Precision (mAP@.50-.95) of 86.4% and a mAP@.50 of 98.4%. Furthermore, it reached a precision of 94.7%, a recall of 96.5% and an F1-Score of 95.6%. These results show that the proposed model is a robust and effective solution, being a significant advancement for integration into real-world driver monitoring systems to enhance road safety.