复杂作业环境下安全帽实时检测算法研究  

Research on real-time detection algorithm of construction workers hard hats in complex operating environments

作　　者：胡启军 潘学鹏[1] 余洋 刘瑞 潘莉 HU Qijun;PAN Xuepeng;YU Yang;LIU Rui;PAN Li(School of Computer Science,Southwest Petroleum University,Chengdu 610500,China;School of Civil Engineering and Geomatics,Southwest Petroleum University,Chengdu 610500,China)

机构地区：[1]西南石油大学计算机科学学院,成都610500 [2]西南石油大学土木工程与测绘学院,成都610500

出　　处：《安全与环境学报》2024年第5期1904-1912,共9页Journal of Safety and Environment

基　　金：国家自然科学基金项目(52178357);四川省“天府万人计划”天府科技菁英项目(川万人第658号);四川省青年科技创新研究团队项目(22CXTD0087);西南石油大学自然科学“启航计划”项目(2022QHZ013)。

摘　　要：为解决建筑工地安全帽背景复杂时检测精度不高、安全帽目标太小不易检测等问题,以YOLOv5框架为基础,提出了一种复杂作业环境下安全帽实时检测算法。首先,在网络中添加坐标注意力机制模块,以抑制无效背景对目标的干扰并提高网络对目标特征的提取能力;其次,在特征融合层引入自适应空间特征融合模块,使网络能自动学习不同特征层的权重,从而增强特征融合能力;最后,采用缩放交并比损失替代完整交并比损失作为边界框损失函数,以解决预测框在回归时的随意匹配问题,进一步提高模型的检测精度并加速收敛速度。结果表明,相较于原始YOLOv5模型,改进后的网络精度提升了2.6百分点,平均精度均值提高了2.3百分点,达到了95.6%,有效提高了复杂环境下安全帽的检测能力。To solve problems such as low detection accuracy when the safety helmet background is complex and the safety helmet target is too small to be detected,this paper proposes a real-time detection algorithm for safety helmets in complex working environments based on the YOLOv5 framework.In this paper,a new hard hat data set is generated for model training by selecting a public data set that conforms to the construction scene and combining images shot at the construction site and photos crawled from the network.In terms of the network model,the coordinate attention mechanism module is added after each C3 module in the YOLOv5 backbone network to make the network pay more attention to the specific location of the target area,improve the backbone network s ability to extract target feature information and prevent the interference of invalid background on the target.The adaptive spatial feature fusion module is introduced in the feature fusion layer so that the network may automatically learn the weights of multiple feature layers,enhancing feature fusion ability and further improving detection accuracy.Finally,the CIoU loss function is replaced by SIoU to solve the random matching problem of the prediction box during regression,improve the model s detection accuracy,and accelerate the convergence speed.The experimental results show that the modified algorithm outperforms other standard algorithms in detection accuracy and detection speed.Compared with Faster R CNN and SSD,the average precision mean is improved by 16.09%and 13.59%,respectively.The improved network s accuracy is improved by 2.3%,and the average precision mean is improved by 2.1%,reaching 95.6%when compared with the original YOLOv5.The improved network pays more attention to the characteristic information of the safety helmet in the detection of the safety helmet,effectively improves the detection ability of the safety helmet in complex environments,and provides a more reliable solution for safety management in complex scenes such as construction sites.This

关 键 词：安全工程 安全帽识别 YOLOv5 坐标注意力 特征融合 损失函数 

分 类 号：X924[环境科学与工程—安全科学]