Research On Detection Method Of Railway Intrusion Obstacles Based On Deep Learning

With the rapid development of rail transit industry,the requirements of new train control system for equipment automation are constantly improving.As one of the important topics for railway researchers,the detection of foreign bodies has always been an important concern in the field of railway operation.The random intrusion of foreign matters such as vehicles,people and objects into the train limit poses a serious threat to the train operation safety.The existing conventional protective measures such as installing protective nets and video monitoring in key areas can not meet the requirements.Therefore,the research on the technology that can realize the continuous detection in front of the train is the practical demand to ensure the safety of train travel.At present,intrusion detection based on deep learning computer vision technology has gradually become the mainstream research direction.Based on the safety protection requirements of train operation,this paper mainly studies the application technology of in-depth learning in the detection of foreign objects in railway intrusion,and makes targeted improvements according to the characteristics of the research object of this subject to meet the specific requirements.The main research contents are as follows:(1)Preprocessing of intrusion image.The noise generated by the heating of electronic components in the process of digital image imaging and transmission is random and unpredictable,which greatly reduces the quality of the image.The removal of noise has always been an important link in the process of digital image processing.In order to reduce the influence of noise on track detection and foreign object detection,this paper compares three widely used denoising algorithms: mean filter,Gaussian filter and bilateral filter,and selects the gauss filter algorithm as the sample image preprocessing method in this paper.(2)Track detection.In view of the unsatisfactory effect of the traditional edge detection method on the rail detection,the sliding window method needs to be used to further screen the characteristic points of the track when establishing the rail linear model,and the processing process is too complex.Therefore,a semantic segmentation network UNet based on deep learning is proposed to detect the rail.By predicting the semantic category of pixels,the rail can be directly segmented from the picture as the semantic category to be detected.Finally,the track feature points are extracted by directly traversing the rail pixels,and the rail linear model is established by the least square method.(3)YOLO v4 realizes the detection of foreign matters invading the railway limit.YOLO v4 enhanced feature extraction network realizes the fine extraction of picture features and forms multi-scale features;PANet fully integrates the features of different scales;The detection heads of different scales can accurately classify and regress objects of different sizes.Through experimental verification,YOLO v4 has achieved good detection effect for most foreign bodies studied in this paper,but the detection effect for objects with small contrast with the environment and less obvious characteristics is not ideal.Therefore,this paper makes the following improvements to the YOLO v4 network:(1)The attention mechanism network senet is added between the feature extraction network and the feature fusion network to further strengthen the utilization of effective features;(2)The initialization method of clustering center in K-means algorithm is improved,and the improved method is used to cluster on the self-made data set and modify the anchor size;(3)Using the method of transfer learning to train the network model and improve the generalization performance;(4)The drop block layer is added to the feature extraction network structure to suppress the over fitting phenomenon of the detection model.The experimental results show that the improved YOLO v4 algorithm not only realizes the accurate detection of foreign bodies,but also performs well in real-time,and has a certain application value.