Research And Implementation Of Railway Track Detection Based On Structure Representation And Global Self-Attention Mechanism

In recent years,with the advancement of urbanization,China’s railway transportation industry has developed rapidly,and it is increasingly important to ensure the safety of railway transportation operations.The application of deep learning algorithms in vehicle automatic driving has become increasingly consummate,and railway transportation is also gradually moving towards automation and intelligence.In railway autopilot system,railway track detection is a core function to locate the track region of interest to facilitate subsequent tasks of track deformation analysis and foreign object encroachment alerting.The traditional methods use hand-crafted feature based on prior information,which is susceptible to factors such as lighting,weather,and occlusion,and the generalization performance of the features is insufficient,leading to poor detection robustness.As a public railway track dataset has not yet been released,research on deep learning-based railway track detection has just started.Most of the current deep learning methods segment track pixels by semantic segmentation and detect track by clustering or fitting operations,which improves the detection robustness,but the inference rate of models is low,which is difficult to meet the real-time requirements of practical applications.To address the difficulties of railway track detection,firstly,we set a set of row anchors based on the position distribution characteristics of the track,and divide each row area into several grids uniformly,transforming the railway track detection task into a selection and classification task of grid positions in the row anchor direction.Secondly,we considerate the ability of the self-attention mechanism to excel in capturing the information association within different features,and bridging the self-attention coding network behind the feature extraction network to extract features containing the global relevance of position.Then we design a semantic segmentation network to use the position information of track pixels to assist in predicting the track grid positions.Finally,we consider that the tracks are paired and the predicted positions in adjacent row directions are similar,and design the corresponding loss functions to represent the structural characteristics of the tracks.The thesis proposes a railway track detection method based on structure representation and global self-attention mechanism with above considerations,which combines the advantages of traditional methods and deep learning methods to detect railway track in a balanced robustness and real-time manner.In this thesis,we propose an end-to-end trainable deep neural network model(Structure Representation and Global self-Attention Network,SRGA-Net),which consists of five components:a feature extraction network,a global self-attention coding network,a classification and prediction network,an auxiliary segmentation network,and a structure representation module.The track images and semantic segmentation images are input to the model,the feature extraction network is used to extract texture and position features,the global self-attention coding network is used to construct global position correlation,and the loss function of auxiliary segmentation network and structure representation module constrains the model.We detect railway track by predicting the track grid positions in the row anchor direction.Sufficient ablation experiments are conducted on the proposed SRGA-Net model to verify the necessity of the network modules designed in this thesis.The model performance is verified on the new railway track dataset.The experimental results show that the SRGA-Net model has good detection accuracy and the real-time performance has met the practical application requirements.