A Novel One-step Method Based on YOLOv3-tiny for Fatigue Driving Detection

Recently, fatigue driving as the most significant factor for traffic accidents has increasingly attracted wide-range concerns. Although diverse fatigue driving detection pipelines based on multi-step eye feature extraction work to some degree, they suffer from lots of computation load resulting in hard deployment for real-time scenes. Additionally, current research lack of real-time scenes dataset. In this paper, we employ a well-designed YOLOv3-tiny model to overcome the above drawbacks. Furtherly, we contribute a novel real-driving scenes dataset named as REAL-Driving dataset which contains 28000 images and 40 real-driving videos in total. Experiments on both the CEW dataset and the self-made dataset are conducted to demonstrate the effectiveness of our approach. The mAP in CEW dataset is 0.999 and the one received in REAL-Driving dataset is higher than other well-known deep learning based object detection algorithms. In the detection of real-driving scenes, we achieve 97.1% accuracy with speed of 5.365ms per frame. The results obtained from sufficient experiments demonstrate the effectiveness and advancement of our method.