基于改进 YOLOv5s的安全帽检测算法

打开文本图片集

摘要：针对于工业场所中密集场景下安全帽佩戴出现的漏检情况以及提高检测精度，提出了一种基于改进YOLOv5s的工业安全帽检测算法.首先，采用CIOU优化Soft-NMS对密集人群重叠的情况，减少重叠目标的漏检，从而提高了安全帽检测的准确性.其次，在网络的中间层添加辅助训练头引入丰富的梯度信息，最后，辅助训练头结合Optimal Transport Assignment 添加到Loss中，通过最优的目标匹配，减少模型的漏检和误检的情况，从而提升模型的准确率和召回率.实验结果表明，改进后的算法平均精确值（mAP@50-90）值为68.3%，相对于原YOLOv5s算法提升了3.8%，准确率为92.3%，相较于原YOLOv5s算法提高了0.7%.

关键词：YOLOv5s;安全帽检测；Soft-NMS;OTA

中图分类号：TP391文献标志码：A

Helmet Detection Algorithm Based on Improved YOLOv5s

XING Xuekai'， LIU Chengyi'， HU Guohua'， LIAN Shun2

（1.School of Advanced Manufacturing Engineering， Hefei University， Hefei 230601， China;2. iFLYTEK Co.， Ltd.， Hefei 230088， China）

Abstract： Aiming to the issue of missing detections of helmet wearing， a helmet detection algorithm based on improved version of YOLOv5s is proposed in order to enhance the accuracy of helmet detection in densely populated scenes within industrial environments. Firstly， CIOU is used to optimize Soft-NMS for handling overlapping instances， thereby reducing the occurrence of missed detections and improving the accuracy of helmet detection. Secondly， an auxiliary training head is introduced into the network's middle layer to incorporate rich gradient information. Finally， this auxiliary training head is integrated into Loss using Optimal Transport Assignment， which optimally matches targets to reduce cases of missing and false detections， enhancing both accuracy and recall rates of the model. Experimental results demonstrate that the improved algorithm achieves an average precision value（mAP@50-90） of 68.3%， surpassing the original YOLOv5s algorithm by 3. 8%. Additionally， the accuracy of the improved algorithm reaches 92.3%， which is 0.7% higher than that achieved by the original YOLOv5s algorithm.

Key words： YOLOv5s; helmet testing; Soft-NMS; OTA 收稿日期：2024-04-23

在采矿、机械工业、冶金高温作业场所、建筑工地等高危行业，安全生产始终是第一位.佩戴安全帽能够有效地预防作业过程中高空坠物对头部的伤害、防止物体打击的伤害、防止机械性的损伤，是有效保障作业工人人身安全的重要措施11.在上述的高危工作场所工作的人员佩戴安全帽尤为的重要，以往在上述的工作场景中，主要以人工监督工人是否佩戴安全帽，不仅浪费人力并且无法保证监管的效果。（剩余6720字）