A New Method of Point-Cloud Registration Based on Affine-invariant Features

Abstract

Considering the application of 3D environment reconstruction and aiming at the point-cloud registration problem when point clouds have a large view change, a 3D point-cloud registration method based on Affine-SIFT (ASIFT) is put forward in this paper. Firstly, RGB images and depth images of an object-of-interest or a complex environment with different viewpoints are collected by a Microsoft Xbox Kinect sensor. Secondly, affine-invariant features and efficient matches are obtained from RGB images by using ASIFT algorithm, and Optimal Random Sample Consensus (ORSA) algorithm is used to eliminate outlier matches. Thirdly, 3D feature point-clouds are generated by inlier matches and corresponding depth information, next unit quaternion method is adopted for computing initial transformation. Finally, the classic Iterative Closest Points (ICP) algorithm is used to achieve precise point-cloud registration. Experimental results indicate that the proposed registration method is more suitable to large angle changes and can achieve precise point-clouds registration results.