Driver Distracted Behavior Detection Technology with YOLO-Based Deep Learning Networks

In order to develop a non-contact driving behavior detection system for the improvement of driving safety, in this study, the YOLO-based deep learning technology is utilized by setting up a webcam on the dashboard to detect the driver's behaviors. By RGB-channel images as inputs, the YOLO-based deep learning models, including YOLOv3-tiny, YOLOv3-tiny-3l, YOLO-fastest, YOLO-fastest-xl are adopted and trained as the candidate detectors. The detected behaviors involve normal driving, distracted head turning, drowsiness, eating, talking on the phone, etc. The experimental results show that when the same parameters are set, the YOLO-fastest-xl has the best performance with multi-category datasets, and its F1-score, false negative rate (FNR), and mAP are 91.84%, 6.94%, and 95.81%, respectively. By the software implementation, the proposed design performs 30 frames per second (FPS) on the NVIDIA GPU-based embedded platform.