基于改进YOLOv5s的综采工作面人员检测算法  

Personnel detection algorithm in fully mechanized coal face based on improved YOLOv5s

作　　者：张磊[1,2] 李熙尉[1] 燕倩如 王浩盛 雷伟强 ZHANG Lei;LI Xiwei;YAN Qianru;WANG Haosheng;LEI Weiqiang(School of Coal Engineering,Shanxi Datong University,Datong Shanxi 037003,China;Intelligent Mining Equipment Industry Technology Innovation Research Institute(Industrial College),Shanxi Datong University,Datong Shanxi 037003,China)

机构地区：[1]山西大同大学煤炭工程学院,山西大同037003 [2]山西大同大学智能采矿装备产业技术创新研究院(产业学院),山西大同037003

出　　处：《中国安全科学学报》2023年第7期82-89,共8页China Safety Science Journal

基　　金：山西省研究生教育创新项目(2021Y739);山西大同大学研究生教育创新项目(21CX02,21CX37);山西大同大学2022年度校级揭榜招标项目(2021ZBZX3);山西大同大学2021年度产学研专项研究项目(2021CXZ2)。

摘　　要：为了智能监控井工煤矿综采工作面危险区域人员闯入和安全帽佩戴问题,避免监控视频受粉尘干扰、光照不均等因素影响图像检测精度的问题,提出一种基于改进YOLOv5s的目标检测算法(简称YOLOv5s-DPE),并建立相关模型。首先,在颈部网络部分,采用深度可分离卷积(DwConv)替换普通卷积,降低参数量和计算量;然后,引入改进的路径聚合网络(PANet)提升特征提取能力,替换边界框损失函数完全交并比(CIOU)为有效交并比(EIOU),提升检测准确率;最后,选取综采工作面视频中的人员图像进行检测,选取煤矿井下人员闯入和安全帽佩戴监控视频作为检测数据集,并进行训练和验证。结果表明:对比初始YOLOv5s算法模型,YOLOv5s-DPE算法模型的参数量下降14.2%,浮点数计算量下降7.6%,算法网络模型大小下降12.5%,均值平均精度(mAP)@0.5提升到93.7%,mAP@0.5∶0.95提升到65.8%,YOLOv5s-DPE模型对小目标检测效果更好,误检漏检等情况有所减少。In order to intelligently monitor the intrusion of personnel entering dangerous areas and the wearing of safety helmets in the fully mechanized mining face of underground coal mines,an improved object detection algorithm based on YOLOv5s was proposed to solve the problem of dust interference and uneven illumination affecting image detection accuracy in monitoring videos.Firstly,in the neck network section,DwConv was used to replace ordinary convolutions,which reduced parameters and computational complexity.Then,an improved PANet was introduced to improve the feature extraction capability,replacing the bounding box Loss function CIOU(Complete-Intersection Over Union)with EIOU to improve the detection accuracy.Finally,the personnel images in the video of the fully mechanized mining face were selected for detection,and monitoring videos of coal mine underground personnel entering and wearing safety helmets were selected as the detection dataset for training and verification.The results show that,compared with the initial YOLOv5s algorithm network model,the number of parameters of the YOLOv5s-DPE algorithm network model is decreased by 14.2%,the number of floating-point arithmetic calculations is decreased by 7.6%,the size of the algorithm network model is decreased by 12.5%,mAP@0.5 is increased to 93.7%,and mAP@0.5∶0.95 is increased to 65.8%.The YOLOv5s-DPE model has better detection performance for small targets,with a reduction in false detections and missed detections.

关 键 词：YOLOv5s 综采工作面 检测算法 深度可分离卷积(DwConv) 有效交并比(EIOU) 路径聚合网络(PANet) 

分 类 号：X936[环境科学与工程—安全科学]