摘要
研究目的:异物侵限是铁路面临的最严重威胁之一,近年来屡屡发生列车脱轨和人员伤亡事故。当前异物侵限主要依靠人员盯控,无法实时进行监测和预警,且存在很多盲区。为了实现异物侵限的实时智能化监测,研究视频图像融合U-Net孪生网络模型的监测技术,建立不同类型异物的样本库,通过与传统方法进行对比分析,验证本文方法的有效性,为该技术工程应用奠定基础。研究结论:(1)提出的引入ResNet-101构造的U-Net孪生网络模型,融合了不同尺度的语义信息,极大程度地减少了目标物的漏检问题,相比传统算法优势明显;(2)U-Net孪生网络模型在不同种类异物识别方面具有泛化性,对列车、人员、树枝、石头和轻飘物具有较高的识别率;(3)视频图像融合U-Net孪生网络技术的精确率和召回率分别达到0.95和0.96,误检率和漏检率分别为0.05和0.04,满足铁路监测要求;(4)本研究结果可为铁路异物侵限智能识别提供科学方法,可作为人工巡检的重要补充,极具推广价值。
Research purposes:Foreign object intrusion is one of the most serious threats faced by railways,and in recent years,train derailments and casualties have occurred repeatedly.At present,foreign object intrusion mainly relies on personnel monitoring,but personnel cannot monitor and warn in real time,and there are many blind spots.In order to achieve real-time intelligent monitoring of foreign object intrusion,the monitoring technology of U-Net siamese network model was studied,and sample libraries of different types of foreign objects were established.By comparing and analyzing with traditional methods,the effectiveness of our method was verified,laying the foundation for the engineering application.Research conclusions:(1)The proposed U-Net siamese network model constructed by ResNet-101 integrates semantic information of different scales,greatly reducing the problem of missed detection of target objects,and has obvious advantages compared to traditional algorithms.(2)The U-Net siamese network model has generalization ability in recognizing different types of foreign objects,with high recognition rates for trains,personnel,tree branches,stones,and light floating objects.(3)The accuracy and recall of U-Net siamese network model reached 0.95 and 0.96,respectively,with false detection and missed detection rates of 0.05 and 0.04,meeting the requirements of railway monitoring.(4)The research results can provide a scientific method for intelligent recognition of railway foreign object intrusion,which can serve as an important supplement to manual inspection and has great promotion value.
作者
王飞
刘桂卫
陈则连
张璇钰
孙琪皓
张瑞
薛双纲
WANG Fei;LIU Guiwei;CHEN Zelian;ZHANG Xuanyu;SUN Qihao;ZHANG Rui;XUE Shuanggang(China Railway Design Corporation,Tianjin 300251,China;Southwest Jiaotong University,Chengdu,Sichuan 610031,China;China Railway Guangzhou Group Co.Ltd,Guangzhou,Guangdong 510088,China)
出处
《铁道工程学报》
北大核心
2025年第4期88-92,共5页
Journal of Railway Engineering Society
基金
国铁集团重大课题(K2023G032)
中国铁设科技开发重点课题(2023A0226403,2022A02538003)。
关键词
异物侵限
图像识别
U-Net孪生网络
foreign object intrusion
image recognition
U-Net siamese network
