Research On Inspection Algorithm Of High-Speed Railway Track State Based On Vision

Along with the advances of our country’s high-speed railways in terms of operating mileage and speed,the traditional inspection mode by manual can hardly meet the requirements of intelligence and modernization for current railways.By applying the technology of machine vision to the detection of high-speed railway conditions,the patrollers can be freed from the time-consuming and dangerous on-site patrol,which also helps to reduce inspection time and improve inspection frequency.In this way,the operating safety of high-speed trains can be further guaranteed.However,the existence of unfavorable factors such as the complex and changeable track environment and the small proportion of abnormal samples often results in poor adaptability and low accuracy of the algorithm,which seriously restricts its engineering application.Therefore,the research of efficient and robust vision-based high-speed railway track state inspection algorithm has important economic value and strategic significance.In order to deal with the influences of complex background,variable illumination and unbalanced data on the performance of high-speed railway detection algorithms,this paper tends to study high-speed railway detection algorithms with higher efficiency and better robustness.For this reason,this article has carried out research on the inspection algorithm of faulty fasteners of high-speed ballastless track and the inspection algorithm of foreign matter on the track slab.The main content of this paper is as follows:Firstly,aiming at the problems of numerous high-speed railway switch railway images,complex background,and the infeasibility of automatic software detection,this paper proposes an efficient switch railway image recognition algorithm combining convolutional neural networks and local binary pattern.The key characteristic of the algorithm is to combine the advantages of convolutional neural networks and local binary pattern,that is,the powerful feature representation ability of convolutional neural networks is utilized to achieve stable distinguished area in the original image,and the fast feature extraction ability of local binary pattern is leveraged to boost the feature extraction efficiency of the distinguished area.First,the original image is replaced with the distinguished area for feature extraction;then,the local binary pattern features extracted from the distinguished area are fed into the support vector machine to realize the switch railway image recognition.The experimental results show that,compared with existing convolutional neural networks method,the proposed algorithm improves the recognition efficiency by an order of magnitude while maintaining the recognition accuracy.Secondly,by making use of the position relationship among the different components on high-speed ballastless tracks,and exploring the specific illumination characteristics of fasteners,a robust component positioning and segmentation algorithm is proposed.First,observing the illumination characteristics of sleeper end lines,a robust sleeper end line positioning algorithm is proposed,which can realize the rough positioning of fasteners;Then,the proposed linear local binary pattern is used to encode fastener images,and fasteners are precisely located combined with template matching algorithms;Finally,by taking the precise position of fastener components as reference,the remaining track components are located according to the prior position relationship among components.The experimental results demonstrate that the proposed algorithm allows strong adaptability for track images having different backgrounds,and can accurately segment each component on the track.Thirdly,aiming at the problems of low accuracy and poor stability of existing faulty fastener recognition algorithms,a faulty fastener recognition algorithm based on local area features is proposed.First,ectopic fasteners are recognized by using the precise location of fasteners;Then,according to the relative and absolute gray values of the local area in fastener subimages,the minimum value area of fastener elastic strips is determined,from which the local area feature is extracted;Finally,the extracted local area features of fasteners are sent to the improved decision tree model to recognize deformed and missing fasteners.The experimental results show that the proposed algorithm achieved good performance in the recognition of faulty fasteners in different environments,which can meet the needs of engineering applications.Fourthly,aiming at the problems of various foreign matters on track slab and a small number of foreign matter samples,a track slab foreign matter detection algorithm based on transformation models is proposed.First,a large number of switch railway and track slab images are collected as source images,which leads to an auxiliary training set;Then,the edges of track slab are extracted by using structure learning-based edge detection operators,and the candidate area of foreign matters is determined according to the edge density of track slab;Finally,the VGG16 model learned on the auxiliary training set is used to recognize the candidate area of foreign matter.Experiments on real and synthetic images confirm the effectiveness and engineering applicability of the proposed algorithm.