摘要
基于现有的铁路侵限异物检测方法只能识别出静态障碍物且识别速度较慢的问题,提出一种基于YOLO算法的铁路侵限异物检测方法.针对铁路侵限异物检测的特殊性,合理设计YOLO模型结构.采用24个卷积层、4个最大池化层及2个全连接层完成异物图像的提取、降维、识别输出,并使用实拍疑似侵限异物图片对YOLO模型进行预训练,达到学习及降低过拟合的目的,最终实现了异物定位及识别的功能.实验中通过对1660张单幅疑似铁路侵限异物图片进行检测,结果表明,该方法对正常曝光的侵限异物图片检测准确率较高,而且在识别速度方面较AlexNet及Adaboost算法具有较大的优越性.
Since the existing railway intruding obstacle detection method can only recognize static obstacles and its recognition speed is slow,an obstacle detection method based on YOLO(You Only Look Once)algorithm was proposed in this paper.The structure of YOLO model is designed reasonably according to the particularity of the railway intruding obstacle.24 convolutional layers,4 maximum pooling layers and 2 full connection layers were used to complete the images extraction,dimensionality reduction and recognition output.Real pictures of suspected intruding obstacle were taken to pre-train the YOLO model,so as to achieve the purpose of studying and overfitting reduction,finally the function of realizing the localization and recognition intruding obstacles from images has been realized.The detection experiment was carried out with 1660 single pictures of suspected intruding obstacles of railway invasion.The results show that this method is more accurate than AlexNet and Adaboost algorithms in detecting the images of suspected railway intruding obstacles under normal exposure,and its recognition speed is superior to AlexNet and Adaboost algorithms as well.
作者
于晓英
苏宏升
姜泽
董昱
YU Xiao-ying;SU Hong-sheng;JIANG Ze;DONG Yu(School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Key Laboratoryof Opto technology and Intelligent Control,Ministry of Education,Lanzhou Jiaotong University,Lanzhou 730070,China;Electrification&Communication Inst.,China Railway Design Corporation,Tianjin 300308,China)
出处
《兰州交通大学学报》
CAS
2020年第2期37-42,共6页
Journal of Lanzhou Jiaotong University
基金
国家自然科学基金(61867003,61763023)
兰州交通大学青年科学基金(2015038)。
关键词
铁路侵限异物
YOLO算法
目标识别
列车运行安全
railway intruding obstacle
YOLO algorithm
target recognition
train operation safety
