Neighbor feature variance (NFV) based feature point selection method for three dimensional (3D) registration of space target

Abstract

Point clouds based three dimensional (3D) registration for the space target pose estimation requires high precision and noise robustness. To enhance the registration accuracy, a novel neighbor feature variance (NFV) based feature points selection method is proposed to provide the high precision input of point clouds based registration and promote the pose estimation accuracy of space targets. The proposed method generates more accurate correspondences by selecting the rank of all point’s NFV, for removing the redundant points and remaining more salient points for accurate registration. In coarse registration, the truncated least squares estimation and semidefinite relaxation (TEASER) algorithm is used to improve the accuracy of coarse registration and reduce the adverse effects caused by sparsity and noise of point cloud. In fine registration, the iterative closest point (ICP) algorithm is applied to estimate the pose transformation. The experimental results show that the maximum translation error and maximum rotation error are less than 0.019 m and 0.129°, and the mean translation and rotation errors can be reduced by 69.85% and 60.58%, respectively. Compared to the registration without the optimization algorithm, the proposed method can be used for variable scale and high noise points based registration.