Structure reparameterized YOLOV5s for improved traffic light detection

Abstract

The detection and recognition of traffic signals, which are important elements in the driving environment, are part of the intelligent driving domain, and the results of the detection and recognition affect the driving safety of intelligent vehicles. However, the classic traffic light signal detection algorithm model is too large, which has certain performance requirements on the devices deployed at mobile terminals. In this paper, we propose a structurally reparameterized network structure that can achieve the performance of classical mainstream algorithms and also significantly reduce the number of parameters and operations of the model. Firstly, the C3 structure of the yolov5s backbone network was replaced with MobileOne structure. Depthwise Convolution was used to reduce parameters and operation costs on the neck part of the model. The CIOU which calculated the location loss was replaced with the EIOU which focused on the length loss and width loss of the anchor boxes. Finally, SE attention module is introduced to further improve the characterization ability of the model. Experiments show that compared with the original Yolov5s algorithm, the parameter number and GFLOPs of the improved structure reparameterized Yolov5s decrease by 29.8%. the map0.5 performance index increased by 0.7% and reasoning speed increased by 19.2%, which effectively achieved a more real-time effect on traffic signal light detection and has certain application value. It can provide effective support for the driving safety of smart cars.