摘要
轨道结构作为高速列车行车的基础,必须保证其良好的在役服役性能,有效地实现无砟道床板间位移测量是实现铁路无砟轨道结构服役状态保持的重要措施。针对无砟道床板非接触式位移测量中感兴趣区域(Region of interest,ROI)自动提取问题展开研究,提出一种基于Faster R-CNN(Faster Regions with CNN)的ROI区域自动提取方法,阐明无砟道床板间位移机器视觉测量原理和实现方法的计算流程。基于Keras图像数据增强模型进行无砟道床板间位移目标库的增加,建立人工标靶数据集。通过CNN中卷积层对测量数据进行特征映射图提取,计算映射图中每个特征点的标靶概率,通过分类和边框回归,精确标记图像中的人工标靶。通过安装某250 km/h的双块式无砟轨道线路的典型测点进行Faster R-CNN算法的准确性和有效性验证。研究结果表明:道床板ROI自动提取算法的召回率为99.16%,准确率为98.91%,可以有效满足无砟道床板间位移测量中精度和准确率的要求;与其他的常用YOLO v3,SSD和Fast R-CNN等ROI算法相比,Faster R-CNN方法的计算效率较好、准确率最高,虽然计算效率上略有不足,但可满足铁路轨道状态实际监测的需求。建议在无砟轨道位移非接触式测量中采用基于Faster R-CNN的ROI自动提取方法,以有效地监测铁路基础设施服役状态。
As the foundation of high-speed train operation,the track structure must ensure its good in-service performance.Effectively realizing the displacement measurement between the ballastless track slabs is an important measure to realize the maintenance of the railway ballastless track structure in service.Aiming at the problem of automatic extraction of regions of interest(ROI)in non-contact displacement measurement of ballastless track slabs,a method for automatic extraction of ROI regions based on Faster R-CNN(Faster Regions with CNN)was proposed,and the calculation process of the machine vision measurement principle and realization method of the displacement between the ballastless track slabs were given.Based on the Keras image data enhancement model,the displacement target library between the ballastless beds was increased,and then an artificial target data set was established.The feature map of the measurement datawas extracted through the convolutional layer in CNN,andthe target probability of each feature point in the mapwas calculated,as well as the artificial target in the imagewas accurately marked through classification and border regression.The accuracy and effectiveness of the Faster R-CNN algorithm was verified by installing a typical measuring point of a 250 km/h double-block ballastless track.The results showthat the recall rate of the ROI automatic extraction algorithm of the ballast track slab is 99.16%and the accuracy rate is 98.91%,which can effectively meet the precision and accuracy requirements of the displacement measurement between the ballastless track slabs.Comparing with other commonly used ROI algorithms such as YOLO v3,SSD and Fast R-CNN,the Faster R-CNN method has better computational efficiency and the highest accuracy.Although the calculation efficiency is slightly insufficient,it can meet the actual monitoring requirements of the railway track status.It is recommended to use Faster R-CNN-based ROI automatic extraction method in the non-contact measurement of ballastless track displacement to effectively monitor the service status of railway infrastructure.
作者
王鲁明
李再帏
赵彦旭
路宏遥
何越磊
WANG Luming;LI Zaiwei;ZHAO Yanxu;LU Hongyao;HE Yulei(School of Urban Rail Transportation,Shanghai University of Engineering Science,Shanghai 201620,China;China Railway 21st Bureau Group Co.,Ltd.,Lanzhou 730070,China)
出处
《铁道科学与工程学报》
EI
CAS
CSCD
北大核心
2022年第2期310-318,共9页
Journal of Railway Science and Engineering
基金
国家自然科学基金资助项目(51978393)
甘肃省科技计划资助项目(19ZD2FA001)
中国铁建科技研发计划项目(2019-B08)。