Research On Detection Technology Of OCS Safety Problems Based On Machine Vision

The safety of the overhead contact system(OCS)is one of the key issues for the safety of high-speed railway transportation.The efficiency of manual detection is low,but the cost is high.It is necessary to explore automatic detection methods.Existing distance measurement and object detection methods include binocular vision,feature recognition,deep learning,etc.Which cannot be well adapted to the shooting and detection environment of the OCS.In order to realize the automatic detection of OCS safety risks,researches were carried out on issues include dangerous tree,hard bend,and arc etc.Detection technologies of OCS safety problems based on machine vision is proposed.The basis for judging dangerous trees is compare the distance between the branch of the tree and the OCS supports.Due to the long railway transportation lines,the image of dangerous trees can only be taken on trains,and stable binocular video or manual annotation cannot be obtained.Based on the above problems,a method of dangerous tree detecting based on longitudinal parallax is proposed.Which obtains the tree subregion,applying multi-scale color threshold segmentation and Fractal Dimension calculation based on the color and structural characteristics in tree area.Adopting the basic idea of the binocular visual ranging algorithm,using the displacement of the image area in the two frames and the movement of the camera to estimate the distance from the area to the center of the imaging plane,and use this distance to obtain the positional relationship between the tree area and the OCS to determine the dangerous tree.The hard bend of contact is bending with small radius and large radian on the linear structure.It is difficult to find and recognize the basic characteristics.To solve this problem,a method of hard bend detection of contact line based on line fitting and distance integration is proposed.This method is based on the structure feature of the contact line,using the distance transformation method to extract the image skeleton,using connectivity detection and straight line fitting to obtain the linear equation of the axis of the line in the image,and identifying the possible contact line according to the direction and radius characteristics of the contact line in the structure,the area of the curved part in the image is obtained by integrating the distance from the point-to-line fitting result,and the comparison of the bending area is used to realize the determination of the hard bend.Existing research on arc detection mainly analyzes the ultraviolet light image based on the optical characteristics of the arc,but some OCS components information will be ignored,and the arc classification cannot be achieved.Due to the uncertain shape of the arc,it is difficult to extract certain features.A method of arc detection and classification based on machine learning is proposed.According to the classification requirements of arc and OCS components,this method adopts two different network models,uses ResNet network to perform global and local arc detection on the images,and comprehensively determines the arc detection results,use DarkNet-53 network to classified the OCS components.Combined the results of the two experiments to achieve arc detection and classification.The experiments show that the critical tree detection,contact line hard bend detection,and arc detection and classification technologies have good accuracy,and the missed detection rate,false detection rate,and time complexity are all within acceptable ranges.The research results of this paper can basically realize automatic detection of OCS safety issues.