基于遮挡感知的安全帽细粒度穿戴检测算法

打开文本图片集

DOI：10.16652/j.issn.1004-373x.2025.18.027引用格式：，，，等.基于遮挡感知的安全帽细粒度穿戴检测算法[J].现代电子技术，2025，48（18）：177-186.

关键词：细粒度；遮挡感知；安全帽佩戴检测；目标遮挡检测；小目标检测；卷积神经网络中图分类号：TN911.23-34；TP391 文献标识码：A 文章编号：1004-373X（2025）18-0177-10

Abstract：Inordertocopewith thediversityandcomplexityofpoweroperationscenariosand solvethedificult problemof safetyhelmetsupervision，ansafetyhelmetfine-graineddetectionalgorithmbasedonoclusionperception （EHD-Net）is proposed.Inalusiontotheproblemof insuficientmulti-scalefeatureextractioncapabilityinocclusionscenarios，alarge separablekerelconvoutionmodule（EDKA）basedonocclusionperceptionisproposed，solvingtheisseofiaccuratedetection causedbytheocclusionofsafetyhelmet.Inalusiontotheproblemsoftheinsuffcientfeatureextractionandfusioncapabilityof themodel，aseparationandenhancementattentionmodule （DAAM）isproposed，andanewfeatureamplificationdetectiohead （FA-Head）isconstructed，resolvingtheproblemof poorsmall-objectdetectionperformancecausedbythelimitedmodel receptivefields.Toaddressthemodel'sinsuficientconvergencecapabilityalossfunctionbasedondistanceandscalefactors （DLS-IoU）isproposed，solvingtheproblemofslowconvergencespeedduring training.Inalusiontotheinsuficient generalizationcapability，aschemeoffine-graineddatasetparitioningisproposed，whichcandividethedatasetintofive diferentcategoriesbasedonthenormsofsafetyhelmet wearingandthestatusofthesafetyhelmetchinstrap，therebyenhancing thepracticalaplicationabilityofthemodel.Theexperimentalresultsshowthat，incomparisonwiththebaselinemodel （YOLOv8n），the average accuracy of the proposed algorithm can reach 94.5%，an improvement of 6.3% ：

Keywords：fine granularity;occlusion perception;safetyhelmet wearing perception;objectoccusion detection;smallobject detection; convolutional neural network

0 引言

近年来，随着中国电力行业的快速发展，对电力作业安全性的重视逐渐增强1。（剩余15661字）