Non-cooperative Target Extraction And Pose Estimation Based On Point Cloud

The issue of rendezvous and docking based on spatial non-cooperative goals has always been a common theme.Such as fuel filling,space debris cleaning and other space missions based on the technology emerge in an endless stream,and what we have to do is solving the relative navigation problem in the process of rendezvous and docking,that is,the pose estimation problem of target star relative to the service star.We use Kinect to simulate the fusion of monocular camera and scanning laser radar,the total number of points collected is limited by resolution,can only reach the order of magnitude in 610,in the verification of the algorithm,we take the help of Stanford bunny.In this paper,we construct a multidimensional tree structure for such a large point cloud data and sample it.Based on this,we will start the algorithm.First of all,we use convex-based segmentation program for point cloud to extract the target star without the need of data training for the purpose of the position of docking ring identifying in the docking process,achieving the separation of target and background,the target can be roughly cut up between components according to the physical structure,.The next step is to estimate the pose after the target is extracted.We use the solution based on point cloud registration,each step of the traditional ICP algorithm is optimized to overcome the initial point set of high degree of overlap requirements in the classical ICP algorithm,and complete the iterative process in a very short time.We extract the features using the fast point feature histogram for point cloud data,establish the corresponding point set with the help of FLANN,use SVD algorithm for initial registration,and complete the final pose estimation process in ICP.