改进YOLO v5n的作业人员着装规范性检测方法  

Detection Method for Dress Suitability of Operating Personnel based on Improved YOLO v5n

作　　者：程文冬[1] 刘超 权程 叶旺盛 王洋 CHENG Wendong;LIU Chao;QUAN Cheng;YE Wangsheng;WANG Yang(School of Mechatronic Engineering,Xi’an Technological University,Xi’an 710021,China)

机构地区：[1]西安工业大学机电工程学院,西安710021

出　　处：《西安工业大学学报》2024年第5期647-655,共9页Journal of Xi’an Technological University

基　　金：国家自然科学基金项目(52302504);陕西省自然科学基础研究计划项目(2022JQ-488);陕西省大学生创新创业训练计划项目(S202310702077)。

摘　　要：为了解决当前作业人员着装规范性(DSOP)检测存在小目标识别率低、环境鲁棒性差等问题。针对DSOP检测中小目标、多重叠、背景复杂等特点,提出基于改进YOLO v5n的算法来实现DSOP检测,首先以E-YOLO架构优化设计提升特征提取和融合效能;其次以Dynamic Head提升多尺度、多视角下DSOP检测精度;最后以OTA和Soft-NMS算法来改善目标堆叠及背景构图复杂的不利影响。实验结果表明:相较YOLO v5n算法,参数量和浮点运算量分别下降31%和16%,精度值提升了0.2%,召回率值提升了1.7%,mAP@0.5:0.95值提升5.9%。可以为各类复杂场景的着装规范性检测提供可行的技术参考。To solve the problems such as low recognition rate of small targets and poor environmental robustness in the current DSOP detection methods,the paper presents an algorithm based on YOLO v5n for DSOP detection characterized by small targets,multiple overlaps and complex backgrounds.Firstly,the E YOLO architecture is optimized to enhance feature extraction and fusion efficiency;secondly,the Dynamic Head is adopted to improve DSOP detection accuracy across multiple scales and perspectives;lastly,the OTA and Soft-NMS algorithms are employed to mitigate the adverse effects of target stacking and complex background composition.The experimental results demonstrates that,compared with with YOLO v5n,for the improved detection method,the parameters and GFLOPs are reduced by 31%and 16%,the accuracy and recall rate are improved by 0.2%and 1.7%,and mAP@0.5:0.95 is increased by 5.9%.This study can provide feasible technical reference for DSOP detection in complex scenarios.

关 键 词：DSOP检测 YOLO v5n效能提升 Dynamic Head OTA Soft-NMS 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]