Helmet detection algorithm based on lightweight improved YOLOv8

Object detection technology enables real-time monitoring of helmet-wearing workers, overcoming manual limitations. However, scholarly improvements prioritize accuracy, complicating the model and rendering it unsuitable for embedded devices with limited resources. This paper presents a lightweight model enhancement approach rooted in YOLOv8. The objective is to minimize parameters and computational load while preserving high detection accuracy, aligning with the deployment constraints of embedded devices. We optimized YOLOv8's C2f module with partial convolution, creating a C2f-Light variant with fewer parameters and less computation. Additionally, there was a redesign of the detection head, which reduced both the number of parameters and the computational complexity. Introduction of the Wise-IOU as a replacement for the CIOU, thereby reducing the harm of low-quality samples. Furthermore, we employed a channel pruning algorithm to eliminate redundant channels to reduce the model size and expedite inference. Experiments results show that LS-YOLOv8n significantly reduces parameters and computations compared to YOLOv8n, without losing accuracy. The pruned LS-YOLOv8n model exhibits a 52%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$52\%$$\end{document} improvement in FPS and has a model size of 1.9 MB.