煤矿井下暗光环境人员行为检测研究  

Research on mine worker behavior detection in low-light underground coal mine environments

作　　者：董芳凯 赵美卿 黄伟龙 DONG Fangkai;ZHAO Meiqing;HUANG Weilong(Department of Mechanical Engineering,Shanxi Institute of Technology,Yangquan 045000,China;School of Mechanical Engineering,North University of China,Taiyuan 030000,China;Quanzhou Institute of Equipment Manufacturing Haixi Institutes,Chinese Academy of Sciences,Quanzhou 362000,China)

机构地区：[1]山西工程技术学院机械工程系,山西阳泉045000 [2]中北大学机械工程学院,山西太原030000 [3]中国科学院海西研究院泉州装备制造研究中心,福建泉州362000

出　　处：《工矿自动化》2025年第1期21-30,144,共11页Journal Of Mine Automation

基　　金：山西省教育厅2022年度高等学校科技创新项目(2022L704);阳泉市科技计划项目(2022JH051)。

摘　　要：煤矿井下环境复杂,对部分作业现场人员行为进行检测时易出现漏检与误检问题。针对该问题,提出了一种煤矿井下暗光环境人员行为检测方法,包括暗光环境图像增强和行为检测2个部分。暗光环境图像增强基于自校准光照学习(SCI)进行改进,由图像增强网络和校准网络构成。人员行为检测通过引入Dynamic Head检测、跨尺度融合模块和Focal-EIoU损失函数来改进YOLOv8n模型。SCI+网络增强后的图像作为人员行为检测模型检测的对象,完成井下暗光环境人员行为的检测任务。实验结果表明:(1)井下暗光环境人员行为检测方法的m AP@0.5为87.6%,较YOLOv8n提升了2.5%,较SSD,Faster RCNN,YOLOv5s,RT-DETR-L分别提升了15.7%,11.5%,0.9%,4.3%。(2)井下暗光环境人员行为检测方法的参数量为3.6×106个,计算量为11.6×109,检测速度为95.24帧/s。(3)在公开数据集EXDark上,井下暗光环境人员行为检测方法的mAP@0.5为74.7%,较YOLOv8n提升了1.5%,表明该方法具有较强的泛化能力。The underground coal mine environment is complex,leading to missed and false detections when monitoring behaviors of mine workers under certain operational conditions.To address this issue,a method for detecting mine worker behaviors in low-light underground environments is proposed,which includes two parts:a low-light image enhancement and a behavior detection.The low-light image enhancement(SCI+)was improved based on self-calibrated illumination(SCI)learning,which consists ofan image enhancement network and a calibration network.The behavior detection improved the YOLOv8n model by incorporating the Dynamic Head detection,a cross-scale fusion module,and the Focal-EIoU loss function.Enhanced images from the SCI+network were used as inputs to the behavior detection model to complete the tasks of mine worker behavior detection in low-light underground environments.Experimental results showed that:①the method for mine worker behavior detection in low-light underground environments achieved an mAP@0.5 of 87.6%,representing an improvement of 2.5%over YOLOv8n,and improvements of 15.7%,11.5%,0.9%,and 4.3%compared to SSD,Faster RCNN,YOLOv5s,and RT-DETR-L,respectively.②The method had a parameter count of 3.6×106,a computational complexity of 11.6×109,and a detection speed of 95.24 frames per second.③On the public EXDark dataset,the method achieved an mAP@0.5 of 74.7%,which was 1.5%higher than YOLOv8n,demonstrating strong generalization capability.

关 键 词：暗光环境 井下人员行为检测 自校准光照学习 图像增强 SCI+网络 Dynamic Head 跨尺度融合模块 Focal-EIoU损失函数 YOLOv8n 

分 类 号：TD67[矿业工程—矿山机电]