Unsupervised Track Abnormal Object Detection

Railway is a very important mode of transportation in our country,and it is at the core of the entire transportation system.At the same time,as an important infrastructure,railway is also one of the main modes of transportation that people choose for daily travel.With the rapid development of railway construction,China’s railway has experienced a qualitative leap from ordinary-speed railway to the current high-speed railway.But at the same time,the safety problems of track lines have become more and more prominent,which has increased the demand for safety inspection of track lines,among which the detection of abnormal objects in the track is one of the important contents of the detection task.However,the current research on railway abnormal objects detection mainly focus on the intrusion of large objects,while there is very little research on the intrusion of small objects on the track bed.At present,the detection is mainly carried out by manual inspection,which is very inefficient.Aiming at the abnormal object on the ballastless track bed,this paper deeply explores the unsupervised track abnormal object detection method based on deep learning.The method proposed in this paper only trains on normal samples,and does not require abnormal object samples to participate in the training.This unsupervised method can overcome the difficulty of small number of abnormal samples and difficult to obtain,which has important theoretical and practical significance.The main works of this paper are as follows:(1)A detection method for track abnormal objects based on pixel-level Out-ofdistribution(OOD)detection is proposed.This method regards the detection of abnormal objects in the track as a pixel-level OOD detection problem,and OOD samples are detected by training with normal samples.Specifically,we first construct a ballastless track image dataset(BTD)that contains abnormal objects on multiple lines,and annotate the dataset at the pixel level.After that,a semantic segmentation model based on normal samples is constructed,and then the detection method is used to detect abnormal objects in the track,and the experimental verification is carried out on the constructed track abnormal objects dataset BTD.Experimental results show that this method can effectively detect the specific location of abnormal objects in the track image,and the detection effect of the same line track image as the training dataset is better than that of the different line track images.Finally,in order to solve the problem of low accuracy,this paper proposes three solutions based on probability ratio,combined class label and uncertainty analysis.Experiments have proved that the accuracy of abnormal objects in track images can be improved by analyzing the uncertainty of samples.(2)A GANomaly-based abnormal object detection method is proposed.Inspired by GANomaly,this method uses the method of reconstructing normal images to detect the specific location of abnormal objects in the test image by analyzing the difference between the reconstructed image and the original image.The training stage of this method only needs to use the normal image dataset without abnormal objects to train the generative adversarial networks,so that the output of the network is infinitely close to the "fake image" of the normal image,that is,the reconstructed image.When the track image with abnormal object is input in the test phase,because the generative adversarial network has not learned how to restore abnormal areas,it is difficult for the network to reconstruct the abnormal areas in the image,while the normal areas can be restored well.Therefore,the specific location of abnormal objects can be obtained by calculating the L1 distance between the original image and the reconstructed image.Experiments show that this method can effectively detect the specific location of abnormal objects in the image.At the same time,in order to solve the problem of unsatisfactory detection of abnormal objects with a color similar to the background,this paper further proposes two improvement strategies to better detect abnormal objects.