基于改进Faster R-CNN和正交投影的无砟轨道板裂缝精细化测量  

Fine-Grained Measurement of Ballastless Track Slab Cracks Based on Improved Faster R-CNN and Orthogonal Projection

作　　者：王卫东[1,2] 张晨雷 胡文博 邱实 王万齐 李娜 王劲 WANG Weidong;ZHANG Chenlei;HU Wenbo;QIU Shi;WANG Wanqi;LI Na;WANG Jin(School of Civil Engineering,Central South University,Changsha Hunan 410075,China;MOE Key Laboratory of Engineering Structures of Heavy-Haul Railway,Central South University,Changsha Hunan 410075,China;Department of Civil and Environmental Engineering,The Hong Kong Polytechnic University,Hong Kong 999077,China;National Rail Transit Electrification and Automation Engineering Technology Research Center(Hong Kong Branch),The Hong Kong Polytechnic University,Hong Kong 999077,China;Institute of Computing Technology,China Academy of Railway Sciences,Beijing 100081,China;Guangdong Meilong Railway Co.,Ltd.,Guangdong Provincial Railway Construction Investment Group Co.,Ltd.,Guangzhou Guangdong 510101,China)

机构地区：[1]中南大学土木工程学院,湖南长沙410075 [2]中南大学重载铁路工程结构教育部重点实验室,湖南长沙410075 [3]香港理工大学土木与环境工程系,中国香港999077 [4]香港理工大学国家轨道交通电气化与自动化工程技术研究中心香港分中心,中国香港999077 [5]中国铁道科学研究院集团有限公司电子计算技术研究所,北京100081 [6]广东省铁路建设投资集团有限公司广东梅龙铁路有限公司,广东广州510101

出　　处：《中国铁道科学》2023年第6期46-56,共11页China Railway Science

基　　金：国家自然科学基金-高铁联合基金资助项目(U1734208);国家自然科学基金资助项目(52178442)。

摘　　要：裂缝的检测和宽度识别是无砟轨道板养护维修作业的重要依据。为此,提出一种基于改进Faster R-CNN和正交投影的裂缝宽度测量方法,并基于虚拟模型合成数据,充分训练深度网络,实现对复杂背景下无砟轨道板表面裂缝的精准检测,以提高裂缝几何特征量化的可靠性。首先,基于二维CAD图纸建立参数化的无砟轨道结构三维BIM模型,并通过UE5物理引擎实现真实裂缝特征与虚拟轨道模型的随机融合和真实巡检场景渲染;然后,通过配置虚拟摄像机输出,模拟真实巡检场景下的虚拟裂缝图像,充分训练改进后的Faster R-CNN网络,并在轨道巡检车采集到的原始图像上进行测试;最后,采用正交投影法逐像素地计算检测结果中裂缝的宽度,并与人工取点测量结果进行对比分析。结果表明:改进后的Faster R-CNN网络对于裂缝检测的平均精度提高约10%;网络性能随训练数据的虚实图像比例而变化,于4∶1时达到饱和,平均精度达95.12%;使用融合裂缝数据集训练出的网络模型能够在保持高精准率的同时,达到更高的召回率,有效减少了裂缝的错检、漏检;与人工测量相比,正交投影法测得的裂缝最小宽度与最大宽度分别增大了3.64%和22.40%,测量结果更加稳定且接近真实值,具有更高的可靠性。Crack detection and width identification is an important basis for the of maintenance and repair operations of ballastless track slab.To this end,a crack width measurement method based on improved Faster R-CNN and orthogonal projection is proposed.Based on the virtual model synthesis data,the depth network is fully trained to realize accurate detection of surface cracks on ballastless track slab in complex background in order to improve the reliability of crack geometric feature quantization.Firstly,a parametric 3D BIM model of ballastless track structure is established based on 2D CAD drawings,and a random fusion of real crack features and virtual track model and the rendering of real inspection scene are realized through UE5 physics engine.Then,the virtual camera output is configured to simulate virtual crack images of real inspection scenarios;the improved Faster R-CNN network is adequately trained and tested on the original images captured by the track inspection vehicle.Finally,the width of the cracks in the detection results is calculated pixel by pixel using the orthogonal projection method and compared with the manual point-taking measurement results.The results show that the average precision of the improved Faster R-CNN network for crack detection is increased by approximately 10%.The network performance varies with the proportion of virtual and real images of the training data and reaches saturation at 4∶1,with an average precision of 95.12%.In addition,the network trained with the fusion crack dataset can achieve higher recall while maintaining high accuracy,effectively reducing the missing detection of cracks.Compared with manual measurement,the minimum and maximum crack widths measured by the orthogonal projection method are increased by 3.64%and 22.40%respectively,and the measurement results are more stable and are close to the real value,which have higher reliability.

关 键 词：轨道板表面裂缝 虚拟数据 改进Faster R-CNN 正交投影法 裂缝宽度测量 

分 类 号：U216.3[交通运输工程—道路与铁道工程]