Research On Railway Obstacles Recognition System Based On Deep Learning

With the rapid development of China’s railway system,the mileage of railway operation and the area covered by railway network have been greatly increased.Serious traffic accidents will occur when non-rail objects such as pedestrians and vehicles enter the range of railway tracks.There are a lot of defects in the traditional detection method,which is far from meeting the requirements of today’s railway system.The contact-based detection method has disadvantages such as large installation workload,high cost and limited detection types.The detection method based on traditional image adopts manual design for feature extraction,which cannot adapt to the diversity of target shape and background to be detected.Therefore,the railway system urgently needs a set of automatic detection system which can carry out high accuracy and intelligent recognition of obstacle images.Based on this,this paper proposes a railway track obstacle detection method based on deep learning.In this method,on-board monocular vision is adopted to realize the detection and recognition of obstacles in the track ahead of the train running.Based on the FASTER R-CNN model of VGG16 as the feature extraction network,adaptive improvement is made in combination with the actual requirements of railway track obstacle detection.In view of the large difference in the size of railway obstacles,the original Faster R-CNN model usually adopts a constant size for all training images.In this case,the detection effect of targets with different sizes is poor,leading to poor robustness of the target detection system.The multi-scale training is introduced,and the size of the image is adjusted randomly on the premise of keeping the original size of the obstacle image unchanged.The adjusted multi-scale image is sent into the network for training to improve the robustness of the system.Combining the reality of railway obstacle detection,image data on road track obstacles in acquisition video image cosco distance small obstacles the problem of lower recognition rate,regional advice network optimization,the original Faster-R-CNN model used in the three dimension ratio and three kinds of specifications of the anchor of anchor parameters generated by the candidate box in the improvement of 5 kinds of scale ratio and 4 specifications of 20 kinds of anchor candidate box in order to enhance small target detection accuracy.In view of the uncertainty factors such as multiple obstacles,the phenomenon of blur and occlusion exists in the detection image.If the target is detected directly,the recognition rate will be greatly reduced.The idea of using generative adversation network to generate occlusion to train detection network is proposed,and ASDN network is introduced into Faster R-CNN model to solve the occlusion detection problem of difficult samples.Based on the training test and result analysis of the model using ASDN network and the model not used,it is found that the model using ASDN network has better shielding detection effect.In order to verify the performance of the detection system.An experimental platform was designed to evaluate the overall network performance after improvement,and the experimental results met the expected accuracy requirements.