Research On Recognition Of Apparent Defects Of Ballastless Track Based On Machine Vision

Ballastless track is one of the main track structure types of high-speed railway in China.Due to long-term exposure to the complex operation environment,it is inevitable to produce apparent diseases such as cracks.These diseases need to be found,observed and rectified in time during daily maintenance to ensure the operation safety of high-speed railway.At present,the detection method for the apparent damage of ballastless track is still based on the traditional manual inspection,which has the problems of low efficiency,large amount of manpower and material resources.In recent years,with the development of machine vision technology,a new idea has been provided for the detection of apparent defects of ballastless track.This paper focuses on CRTS II slab ballastless track,one of the main structural types of ballastless track used in China’s high-speed railway,and introduces machine vision into the field of apparent defect detection,and carries out research around depth learning,image stitching,damage quantification calculation and system design.The main research contents of this paper are as follows:(1)This paper systematically summarizes and analyzes the common apparent diseases and structural characteristics of CRTS II slab ballastless track,and puts forward the unitization idea and main solutions for the detection of apparent diseases of ballastless track based on machine vision.This paper introduces the characteristics and causes of common apparent diseases of CRTSⅡ slab ballastless track,and designs a detection scheme of apparent diseases of ballastless track based on machine vision.The design of the scheme takes the unit image of a single sleeper as the disease detection unit,mainly including the image mosaic of ballastless track,the detection of typical features and apparent diseases,and the quantitative calculation of crack damage.(2)Research on image mosaic method of ballastless track based on unitization and smart phone.Image mosaic mainly includes four key steps: feature point detection and description,feature point matching,image registration and image synthesis.For feature point detection and description,SIFT feature detection operator and SURF feature detection operator are mainly analyzed;For feature point matching,BF Matcher algorithm and FLANN algorithm are analyzed.Finally,FLANN algorithm is selected as the feature point matching method.Aiming at the problem that a large number of feature point pairs are matched due to the high similarity between ballastless track images,the Lloyd’s ratio test method is proposed to further purify the matched feature point pairs;For image registration,affine transformation and perspective transformation are analyzed.Combined with the algorithm running speed,affine transformation is finally selected to build the transformation model of track image mosaic;For image synthesis,image arithmetic operation and weighted average fusion method are analyzed.Combined with the comparison of synthetic image effects,image arithmetic operation is finally selected as the method of orbital image synthesis.At the same time,the evaluation indicators of splicing results are introduced from qualitative and quantitative perspectives.Finally,combined with a specific project example,the image of a ballastless track line captured by a smart phone is taken as the material,and the actual effect of the splicing algorithm in this paper is verified by taking the image stitching of the side unit of ballastless track as an example.(3)Based on the deep learning technology,the typical characteristics and apparent disease detection model of ballastless track are constructed.Based on deep learning technology,the pre-crack detection model and apparent disease detection model of CRTS Ⅱ slab ballastless track are constructed.CA mortar layer and wide and narrow joints are selected as the research objects of apparent defects of ballastless track,the differences between the two algorithms of Two-Stage target detection and One-Stage target detection are introduced,the principle framework of YOLO series algorithm model is analyzed in detail,and finally YOLOv5 network is selected as the training network of typical characteristic pre-crack and apparent damage detection model of ballastless track.The result of model evaluation index shows that the two models trained have high detection accuracy,which lays a foundation for judging the integrity of unit image mosaic,apparent disease detection and damage quantification calculation.(4)The quantitative calculation method of crack damage of ballastless track is studied.Take the typical CA mortar layer off-joint damage as the research object,use the ballastless track apparent disease detection model and image processing technology to intercept the image of the off-joint damage area,use the image filtering algorithm to remove the noise in the intercepted image,so that the off-joint damage area is more smooth and complete,and then perform threshold segmentation on the image to obtain the connected region representing the off-joint damage,and solve the problem of fine burrs around the connected region,The morphological operation method is proposed for processing.At last,the off-seam skeleton of the deburred image is extracted,and the area,length,width and other relevant features of the off-seam are calculated by combining the calibration coefficient.Taking the calculation of the height of CA mortar layer from the joint as an example,the experimental verification shows that the error between the calculation results of the algorithm and the actual measurement results is small,indicating that the quantitative calculation method of the crack damage of ballastless track studied in this paper has certain feasibility and practicability.(5)The frame design of intelligent identification system for apparent defects of ballastless track based on machine vision.Based on the aforementioned research results of image mosaic,pre-crack detection,apparent disease detection and crack type quantitative calculation model of ballastless track,an intelligent identification system framework of apparent disease of ballastless track based on machine vision is designed.The system is mainly used to splice the unit image of the sleeper,identify and locate the damage based on the unit image,and then carry out targeted processing on the local image of the identified damage to achieve the quantitative calculation of the damage,and finally record the basic information of the disease.During the inspection,the staff can query the historical detection data of the damage under the same unit image,and check the change trend of the specific damage value and the apparent deterioration of the damage,Make appropriate maintenance decisions.