Research On Railway Obstacle Intrusion Detection Algorithm Based On Generative Model

With the increase of mileage of high-speed railway in China,the detection technology of railway foreign matter invasion is of great significance to ensure the safety of high-speed train operation.At present,the detection technology based on machine vision has become the main research method,especially the recognition algorithm based on supervised learning is widely used.The existing railway obstacle detection algorithm based on supervised learning needs a large number of labeled samples,however,in practical application it is unable to obtain all types of obstacle intrusion samples,and manual labeling of samples is time-consuming and laborious,which will also cause the missing of obstacle intrusion forms in training samples.In order to solve the above problems,this paper designs a detection algorithm of railway obstacle intrusion based on generated model.This algorithm can realize the detection of railway foreign body intrusion without training abnormal samples.It is of great significance to improve the accuracy of obstacle intrusion alarm and reduce the missing of unknown samples.Firstly,the actual railway scene video is used as the source to build the railway scene image data set based on different lighting conditions and different backgrounds.Due to the small proportion of obstacle objects in the railway scene image,it is difficult to extract features.The image subdivide algorithm is used to re segment the sample image,and the sample database containing about 20000 images is established by the segmented image for the training and testing of network model.Secondly,based on algorithm framework of generative model,a encoderdecoder-encoder structure is proposed,which combines Autoencoder and Generative Adversarial Network.The input image is mapped to a low-dimensional latent vector,and the resulting latent vector is used to reconstruct the generated image.In this paper,the specific design of each layer network structure is carried out,and the relevant network parameters are compared and selected.A new loss function is proposed to optimize the network model,and finally an optimal network structure is constructed for anomaly detection.In this paper,MNIST and cifar10 databases are used to verify the algorithm and compare with the existing generation model.The test results show that the proposed algorithm can generate more realistic images and improve the real-time performance of anomaly detection.Finally,according to the railway data set,the algorithm proposed in this paper isdesigned and tested for network structure parameters,and only normal samples are used for training to test the detection effect of railway obstacle invasion.By analyzing the experimental results of different railway scenes,the accuracy of foreign object detection is 95.42% without abnormal samples and label samples,and the rate of missing detection is only 0.3%.At the same time,a obstacle labeling algorithm is designed for the detected abnormal image to realize the obstacle object labeling,and the segmented image is reintegrated into the railway scene map to facilitate the application in the actual operation of the railway.In this paper,the generation model is applied to the railway environment anomaly detection system to avoid the shortcomings of the existing algorithm by an unsupervised learning method.It has some advantages in solving the problem of missing reports in the detection of railway obstacle matters and ensuring the real-time performance.It shows a good classification effect in the experimental results.