Research On Recognition And Positioning Method Of Spacecraft Based On Point Cloud

With the continuous advancement of the exploration of outer space,the number of spacecraft launches continues to increase,but at the same time,more and more spacecrafts are scrapped and stranded in orbit due to the end of their service life or long-term work,which constantly threatens other normal conditions.The missions of spacecraft in space mainly include on-orbit maintenance,interactive docking and salvage of space junk,etc.The premise of these missions is to identify the target object and know the exact pose information for subsequent operations.For a variety of aerospace vehicle shapes,the capture positions are generally typical structures such as solar panels,docking rings,and nozzles.These components are all rectangular or circular,and image-based methods are usually used to identify such simple geometric shapes..However,the lighting conditions in the space environment are more complicated,so a sensor that is not sensitive to lighting is required to obtain the information of the target object,while sensors such as depth cameras or lidars do not depend on the lighting conditions and can directly obtain the coordinates of the threedimensional points on the surface of the object,which becomes a big research hotspots in the past few years.This article takes spacecraft as the research object,introduces the measurement methods of spacecraft at home and abroad,and aims at the recognition and pose estimation issues during operation,the deep neural network recognition framework for point clouds,key point extraction and description,and registration methods are studied.The specific content is as follows:First,in order to better store and search point cloud data,the data structures of Kdtree and Octree are studied,and the difference and efficiency of their radius search and nearest neighbor search methods are analyzed,and the reasons are compared and analyzed through experiments.The process of point cloud data collection is often accompanied by noise,and the amount of data acquired by lidar can usually reach millions.For the efficient operation of the subsequent algorithm,preprocessing operations such as outlier removal and point cloud downsampling are required to optimize the algorithm.Secondly,due to the disorder of the 3D point cloud,even the same point cloud data will have different permutations and combinations.According to the characteristics of the point cloud data,the neural deep network suitable for 3D points is studied,and the network structure of Point Net++ is introduced.On this basis,the important geometric structure normal vectors in the point cloud data are attached,and the existing 3D model of spacecraft is used to make The data set trains the neural network structure.Again,point cloud registration requires a good initial pose,and the method based on key point description can provide good initial pose information.Due to the defects of the equipment and objective factors,even the point cloud data of a flat object will not be absolutely flat,and it is possible to produce convex parts on the plane,and then these points on the plane will be extracted during the key point extraction process.Describing key points is unfavorable.In order to solve this problem,a key point extraction method based on the normal vector histogram is proposed.According to the information of the angle between the normal vector of the neighboring point,the key point extraction on the plane is reduced,and the data set and other commonly used key point extraction algorithms are used.Compare.For the initial registration algorithm for random sampling consistency,the wrong matching points in the registration process are eliminated through the conditions of polygon geometric constraints to improve the registration efficiency.Finally,build a software and hardware experiment platform according to the entire algorithm flow,obtain the accurate values of the position and attitude of the spacecraft before and after the movement according to the precise control of the target simulator,and compare with the calculated values obtained through the algorithm and analyze the possible causes of the error.