Driver activity detection and classification using deep learning-based YOLO models for intelligent transportation systems

Driver distraction is one of the major causes of road traffic accidents worldwide, resulting in significant economic losses, injuries, and fatalities. Intelligent driver monitoring systems have emerged as effective solutions for improving road safety through real-time detection of unsafe driving behaviors. This study proposes a deep learning-based framework for driver activity detection and classification using YOLO object detection architectures. A real-world driver activity dataset containing 5,422 annotated images was collected from drivers under natural driving conditions. The dataset consists of eight driver activity classes, including safe driving, drinking, eating, phoning, operating vehicle components, makeup/hair adjustment, object management, and talking to passengers. Four YOLO models, including YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv8n, were trained using transfer learning and evaluated based on precision, recall, AP, mAP@0.5. The results showed a consistent improvement in detection performance as model complexity increased. YOLOv5s achieved 90.1% precision, 88.7% recall, and 90.0% mAP@0.5, while YOLOv5m improved these values to 91.9%, 90.5%, and 92.5%, respectively. YOLOv5l further increased performance with 93.3% precision, 92.1% recall, and 95.0% mAP@0.5. YOLOv8n achieved the highest accuracy with 96.0% precision, 95.2% recall, and 96.3% mAP@0.5, demonstrating superior detection capability and computational efficiency. The findings indicate that advanced YOLO architectures provide accurate and efficient solutions for real-time driver activity recognition and can significantly contribute to intelligent transportation systems and road safety enhancement.