Abstract:

To address the problem that distracted driving detection using classification methods can only identify a limited number of distracted driving behavior categories and ignore temporal information， we propose a distracted driving detection method based on improved YOLOv5. First， the Ghost module is introduced on the basis of YOLOv5， and linear transformation is used instead of partial conventional convolution for feature extraction to lighten the network model to achieve fast and accurate detection of cell phone， water bottle， driver's eyes and head region in the image； second， after obtaining the target detection results， the logic algorithm is designed to detect the presence of distracted driving in each frame by combining with head pose estimation. Second， on the basis of obtaining the target detection results， a logic algorithm is designed and integrated into YOLOv5 with head pose estimation to detect the presence of distracted driving in each frame from both cognitive distraction and visual distraction perspectives， which avoids the problem that the classification method is limited by the number of distracted driving categories， and then setting an appropriate time threshold， thus realizing real-time and effective distracted driving detection； finally， three sets of experiments are conducted on the collected driving behavior dataset of 18 drivers to verify the feasibility and effectiveness of the method in this paper.

Key words: distracted driving, YOLOv5, driving behavior, target detection, head pose estimation