Research On Intelligent Recognition And Location Algorithm Of Line Features Based On Machine Vision

Train positioning is the basis of train running control.At present,multi-sensor fusion methods such as ground transponder,track circuit and satellite navigation are widely used to realize high-precision train positioning.However,under extreme conditions such as ground equipment damage,vehicle ground communication failure and satellite signal shielding,there is still the possibility of positioning loss.Therefore,this paper proposes a virtual transponder technology based on machine vision,which can intelligently identify a large number of periodic markers along the railway,such as kilometer marker,hundred meters marker and catenary column number,and establish contact with the mileage of train running line,so as to realize high-precision positioning with similar transponder function,get rid of the dependence on external positioning equipment and improve the reliability of positioning system.Firstly,this paper analyzes the problems existing in the existing train positioning technology,and analyzes the feasibility of train positioning based on line visual features by investigating the actual scenes along the railway.Based on this,the overall scheme of virtual Transponder Based on line marker recognition and positioning is designed,and two positioning schemes based on dual camera and single camera are proposed.Through experimental analysis,the positioning scheme based on single camera is finally selected,and the technical indexes of each part of the algorithm are analyzed according to the overall accuracy requirements of train positioning.According to the single camera design scheme of virtual transponder,this paper focuses on the machine vision algorithm based on deep learning,including marker detection algorithm,marker recognition algorithm and marker distance estimation algorithm.Firstly,the line marker detection algorithm based on YOLOv5 is designed.In order to meet the needs of algorithm training and verification,the railway scene line marker target detection data set is established to solve the problem of category imbalance in the data set.Aiming at the problems of false detection and missing detection in the detection of line markers in the original network,targeted improvement and optimization are proposed,which achieves better detection effect than the original network,and achieves an overall average accuracy of 93.8% on the railway scene data set.Secondly,the line marker recognition algorithm based on CRNN is designed.In order to meet the needs of algorithm training and verification,the line marker text recognition image set is established,and the feature diversity of the data set is improved through data enhancement.Aiming at the problems of poor feature extraction effect and slow speed of the original network,by reducing the convolution layer and introducing the residual structure,the recognition effect is better than that of the original network,and the sample recognition accuracy of 96.75% is achieved on the railway scene data set.Finally,a marker distance estimation algorithm is designed,and the distance of line markers detected and recognized in the image is estimated by using the regression model of machine learning,which solves the problem of train position acquisition in the process of visual virtual response.In order to verify the designed algorithm,an embedded virtual transponder prototype system is built.The lightweight improvement of virtual transponder vision algorithm is completed by reducing the pruning of input channel and convolution channel.The network acceleration and deployment are carried out based on Tensor RT framework.The on-board experiment is carried out on the circular railway,and the whole process line feature recognition and positioning of 14 fps on the on-board prototype system are realized,The positioning accuracy on the experimental line is 1.88 m.