基于改进YOLOv3的铁路关键作业流程自动鉴别系统  

Automatic identification system of railway critical operation process based on improved YOLOv3

作　　者：张志晨 李军[1] 何波[2] 郑文静 王昱凯 ZHANG Zhi-chen;LI Jun;HE Bo;ZHENG Wen-jing;WANG Yu-kai(School ofInformation and Automation,Qilu University of Technology(Shandong Academy of Sciences),Jinan 250353,China;School of Information Science and Engineering,Shandong University,Qingdao 26600,China)

机构地区：[1]齐鲁工业大学(山东省科学院)信息与自动化学院,山东济南250353 [2]山东大学信息科学与工程学院,山东青岛266000

出　　处：《齐鲁工业大学学报》2022年第4期14-22,共9页Journal of Qilu University of Technology

基　　金：国家自然科学基金(12005108);山东省自然科学基金(ZR2020QF016)。

摘　　要：针对铁路列车关键安全点的检测问题,提出一种改进YOLOv3网络的铁路关键作业流程自动鉴别系统,提高了对小物体特征的检测精度。使用残差网络对darknet53结构进行改进,采用FPN网络加强特征提取以及融合(104,104,24)的检测分支,通过YOLO_Head获得预测结果,从而提高小目标检测精度。在铁路数据集上的实验结果表明,改进的YOLOv3网络在较小的关键点的检测上有了较大突破,列车中电箱锁的置信度提升了0.33,精准度提升了21.39%,召回率提升了10.34%,F1值提升了0.16;3种关键特征的mAP提升了12.05%。Aiming at the detection problem of railway train critical safety points,an automatic identification system of railway critical operation process based on improved YOLOv3 network is proposed,which improves the detection accuracy of small object features.The residual network is used to improve the darknet53 structure,FPN network is used to strengthen feature extraction and fusion of detection branches(104,104,24),and YOLO_Head is used to obtain prediction results,so as to improve the accuracy of small target detection.The experimental results on the railway data set show that the improved YOLOv3 network has a great breakthrough in the detection of small key points.The confidence,accuracy,recall rate and F1 value of the train electric box lock are improved by 0.33,21.39%,10.34%and 0.16 respectively.MAP for 3 key features is improved by 12.05%.

关 键 词：深度学习 目标检测 YOLOv3 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]