A lightweight YOLOv5 model used for safety helmet and reflective clothing detection

Abstract

Target detection models are widely used for helmet and reflective clothing detection under intelligent surveillance systems, but the current models are large in size, difficult to deploy in embedded devices with limited computing power and memory, and poor in robustness. To address these problems, a lightweight helmet and reflective clothing wearing detection model is proposed for security inspection robots, named YOLOSE. a dataset with complete categories and close to real construction scenes is constructed, and the generalization ability of the model is improved by various image enhancement methods; on the basis of YOLOv5s algorithm, the backbone network is replaced by ShuffleNet v2, a lightweight network, to reduce the number of parameters. reduce the number of parameters; introduce α-IoU loss function and h-swish activation function to improve the operation speed of the low computing power platform; and improve the size of the initial anchor frame to adapt to specific detection tasks. Experimental results show that YOLOSE has better robustness for different environments and maintains good accuracy with reduced volume compared to YOLOv5s.