Driver Fatigue Detection Using YOLOv5 with Genetic- Algorithm-Based Hyperparameter Optimization and Fine-Tuning Strategies

Abstract

Driver fatigue is a major contributor to traffic accidents, injuries, and economic losses. This study proposes a computer-vision-based driver fatigue detection system using YOLOv5s to identify three facial states: normal, eye closed, and yawning. The model is fine-tuned from pre-trained COCO weights and evaluated under multiple training schemes, including full fine-tuning, partial layer freezing, and hyperparameter evolution based on a genetic algorithm. The dataset consists of 5,706 training images and 2,390 testing images compiled from self-recorded and public sources under varied lighting conditions. Experimental results show that the best model without evolution achieves an mAP of 0.887, while the best evolved model reaches an mAP of 0.888 and runs at 30 FPS on a CPU-based device. These findings indicate that the proposed model is sufficiently accurate, lightweight, and practical for early-warning deployment on low- resource edge devices in advanced driver assistance systems.