Y.R. Guo, R.N. Wu, S. Patnaik

Pages: 51-62

Abstract
Reducing the risk behaviors of drivers during driving is of great significance to improve the level of road traffic safety. Using intelligent technology to monitor unsafe behaviors of drivers can effectively reduce and control the occurrence of unsafe events of drivers, and avoid and reduce casualties and property losses. In this paper, we propose a multiple occurrence point recognition method for traffic unsafe driving behavior based on YOLOv5 visual perception. This method uses media filtering to remove the noise of driving image, divides the image in YOLOv5 network structure, objectifies the image features; uses visual perception to obtain the true aspect ratio, calculates the fatigue coefficient; uses threshold value to judge unsafe driving behavior, and determines the frequent occurrence of unsafe driving through low function. The results show that the data SNR of this method can reach 77.1dB, the recall rate can reach 99.0%, and the recognition time is less than 8s, indicating that this method has stronger recognition performance.
Keywords: unsafe driving behavior; feature recognition; YOLOv5 structure; feature fusion; loss function