Vision Based Ballastless Track Abnormal Object Detection

As of 2019,the total mileage of China's railways has reached 132,000 kilometers,of which the total mileage of high-speed railways has exceeded 35,000 kilometers.The railway has become the significant part of transportation for the Chinese people's daily business trips.It is an important way of transporting materials between regions.With the rapid development of China's railway,the problem of railway safe operation has become increasingly prominent,which has increased the demand for inspection of rail infrastructure.The detection of railway abnormal object is one of the important tasks in track safety detection.At present,most of the research on abnormal object detection in railways are aimed at some large intrusive objects.The detection of abnormal objects on the track bed mainly depends on manual detection.We focus on abnormal detection on the ballastless track bed and explore the vision-based automated detection technology in railway abnormal object detection tasks.We argue the different learning methods in abnormal object detection.We propose railway abnormal object detection method based on similarity symmetry metric learning in this paper.Our proposed method improves the effectiveness and generalization ability of the detection system,which has important theoretical significance and practical value.The main work of this paper is as follows:(1)A railway abnormal object detection method based on supervised learning is proposed.In order to tackle the problem that there is no large-scale railway abnormal object dataset,we have collected track image from four long railway line.We employ data argument technology to enriching abnormal data.We use multiple supervised deep learning methods to detect abnormal object in railway.The experimental results verify the effectiveness of the supervised deep learning method.(2)A method for detecting abnormal objects in ballastless track based on unsupervised learning is proposed.In the real ballastless track image data,there are few real abnormal object samples,and the types cannot be exhaustive.Therefore,the supervised learning method has problems of being unable to detect unknown abnormal object and lacking enough training data.In response to these problems,this paper proposes an unsupervised abnormal object detection method based on the reconstruction loss of the auto-encoder.This method only uses the data of normal objects to train an auto-encoder model,and uses the difference in reconstruction loss distribution between normal objects and abnormal objects to identify anomalies of proposals.(3)Aiming at the shortcomings of low accuracy based on auto-encoder method,we propose an unsupervised method based on symmetry metric learning to detect abnormal objects.Because the track image is geometrically symmetrical about the centerline of the track,normal objects appear repeatedly in pairs,and the difference from their symmetrical regions is large.So that we propose a similarity model using the track image without any abnormal data,which learn the similarity between the proposals and the corresponding symmetric regions.The method judges the abnormality of proposes by measuring the similarity score.The experimental results verify that the unsupervised method based on symmetry model achieves a good accuracy and recall rate.(4)An easy-to-use platform for railway abnormal object detection system is built.The system incorporates two types of supervised and unsupervised models based this paper,as well as image preprocessing and data storage technologies.