Research On Lidar Point Cloud Semantic Segmentation Of Overhead Catenary System Based On Deep Learning

High-speed railway is the gold business card made in China now,and it has important strategic significance on the development of China’s economy and politics.With the expansion of high-speed railway,higher requirements have been put forward for its operation and management.The overhead catenary system is an important part of the high-speed railway power supply system,and its working state relates to the normal and safe operation of high-speed railway.Due to being in an outdoor environment for a long time,the overhead catenary system is susceptible to outside influence and appear breakdown,so regular testing and maintenance are needed to keep it in normal working condition.The traditional manual inspection method is not efficient,reliable,safe,and so on,and it is hard to meet the demand of high-speed railway development.Lidar is an accurate and fast detection technology that can obtain three-dimensional spatial information,and it can be used to improve the efficiency and accuracy of catenary inspection.The overhead catenary system consists of many components,its structure is complex,and the data of the lidar point cloud is large.Therefore,how to efficiently and accurately segment point cloud data to identify catenary components is a key problem to be solved.To solve the problem of identifying multiple components of catenary based on the point cloud,this thesis takes the point cloud data collected by two-dimensional lidar installed on the catenary inspection system as the research object,and researches a point cloud semantic segmentation method based on deep learning to identify eight types of catenary components.The specific work is as follows:This thesis proposes a point cloud semantic segmentation framework.To deal with the problems of the training and recognition caused by large-scale point cloud in the catenary scene,this thesis divides the point cloud semantic segmentation task into three parts: single frame point cloud classification,single frame point cloud combination,and multi-frame point cloud semantic segmentation.This thesis uses CNN to classify every frame,then takes a combination algorithm to combine adjacent single frames to multi-frame point cloud data based on classification results.After that,a segmentation model is adopted to segment the multi-frame point cloud data for recognition of the components.This thesis proposes a semantic segmentation network based on the KNN search algorithm.Due to the insufficient ability of local feature extraction,the Point Net can not identify the components with sparse data in the multi-frame point cloud semantic segmentation task.For this problem,this thesis uses the KNN search algorithm to create a local area for each point,constructs the local features based on the features of point cloud and the distance between the points to enhance the local feature extraction ability of the network,and combines local features of multiple layers and global features to identify the category of each point.This thesis collects the real catenary point cloud dataset and labels it,and verifies the feasibility and validity of the proposed methods based on it.The segmentation network based on the KNN search algorithm performs segmentation of each component than Point Net and achieves average segmentation precision of 97.01% and average segmentation Io U of 93.71%.Besides,the local feature extraction method can improve the segmentation accuracy of the components with sparse data.The precision of dropper and steady arm is increased by 4.78%、3.08%,and the Io U of them is increased by 10.88%、3.53%.