Research And Validation Of Turnout Detection Method Based On Deep Learning

The western region of China has a vast territory,complex environment and imperfect transportation infrastructure.In order to adapt to complex environment along the railroad in the western region,it is required that train has autonomous perception ability of operation status.In the process of train running,real-time and accurate perception of the track and turnout in front of the train is a prerequisite for ensuring the safety of train operation.At the same time,turnout detection can assist train in positioning correction and further reduce ground and trackside equipment.Traditional track and turnout detection methods are limited by manual design features and large calculation volume,which cannot meet real-time and accuracy requirements.To address the above problems,this thesis carries out research on track and turnout detection based on deep learning.On the basis of existing algorithm structure of convolutional neural network,the algorithm is improved and optimized by combining with the actual application scenarios,rail segmentation algorithm based on deep learning and turnout reference position(frog and blade)detection are proposed to realize the recognition and classification of turnout.The main research of this thesis is as follows.(1)Conduct theoretical research related to deep learning such as semantic segmentation and target detection.Combined with the research needs,the turnout recognition scheme based on U-Net network and YOLOv5 network is constructed.The actual image data is collected and annotated to build rail segmentation dataset and turnout reference position detection dataset.(2)The rail segmentation algorithm based on RGBD fusion is proposed to achieve rail detection under different environments.The depth image is introduced as the supplementary information of RGB image to compensate for the lack of lighting conditions.Meanwhile,in order to better utilize the spatial information contained in the depth image,convolutional block attention module and atrous spatial pyramid pooling module are introduced in the convolutional neural network to further optimize the rail segmentation model based on depth fusion.The effectiveness of the improved rail segmentation algorithm is verified through relevant algorithm comparison experiment.(3)Based on the result of rail segmentation,YOLOv5 algorithm is improved by introducing the Ghost module.Under the condition of ensuring the detection accuracy,it improves the detection speed of the network for turnout reference position,reduces the parameters and computation of the model,and optimizes the overall performance of the network.Experimental test shows that the algorithm can achieve fast and accurate identification and localization of turnout reference position.(4)Based on the result of turnout reference position detection,the classification method of simple turnouts is proposed.simple turnouts are classified by extracting the track center line,combining turnout reference position and the distribution of the track center line in space.The proposed algorithm is validated using the constructed dataset.The experiments prove that the track center line extraction algorithm can be applied to different types of turnout scenarios and the turnout classification algorithm has certain feasibility.Figure 74,Table 17,References 65.