| Currently,train is the most common tool for medium and long travel in our country.In recent years,with the growth of railway mileage and the increase of train speed,the requirements for the train driving environment have become higher and higher,and the corresponding requirements for the railway inspection system have become more and more stringent.In terms of railway in section,traditional inspection method facing such problems as lower efficiency,and lower safety factor.on the contrary,drone have the advantages of flexibility,easy operation,and high safety factor,with the power to overcome any kinds of poor working environment,so the new railway inspection method based on drones has gradually become the mainstream.This thesis uses drone as a carrier to design a track recognition system based on monocular vision,and identify various obstacles in the railway track environment,providing an efficient inspection scheme for railway tracking system.There are four main aspect included in this thesis.(1)Track image preprocessing.In order to realize the track identification of the railway rail in the future,the railway image is preprocessed first.Aiming at the track image,the image filtering algorithm is used to eliminate the noise.Secondly,the local contrast enhancement algorithm is used to enhance the metal features of railway rails;Finally,the contour features of the track are further highlighted by improved adaptive morphological processing.(2)Positioning and identification of rails in railway lines.Aiming at the problem of rail identification,the rail lines are divided into two types,straight rails and curve rails,and corresponding algorithm are designed.Facing the straight rail,the inverse perspective transform is used to process the track image,and then the least square method based on the pixel spacing constraint clustering is designed to fit the straight rail line.Then the mathematical model of the rail is constructed,and the data filtering algorithm of the track line is designed,finally achieve stable line recognition.For the curved rail,the curved rail is divided into near filed and far filed,the near filed is fitted with linear model and the far field of view is fitted with a 3-degree Bspline curve model,then the curved track line is finally obtained.(3)Track obstacle recognition based on deep learning.In order to detect the obstacle target information in the track,Yolov5 s target detection model is used for track obstacle detection.Based on the poor detection effect on small targets of Yolov5 s target detection model,the CBAM self-attention mechanism is added subsequent to the feature extraction module to improve the feature extraction ability of the network for small targets.(4)The construction of track inspection drone system and verification of overall scheme.A drone physical verification platform was built,then to verify video data and verify the algorithm in the actual environment,and finally analyze the data to get the actual feasibility of the algorithm.In this thesis,a new drone inspection scheme equipped with vision module is proposed,and a straight and curved track identification algorithm based on computer vision is designed to improve the accuracy and stability of track straight line and curve identification.At the same time,the track obstacle identification algorithm based on neural network is adopted to solve the problems of low efficiency and low safety factor in the current inspection.Finally,the effectiveness of the proposed algorithm is verified by a real platform. |
PDF Full Text Request