KEYPOINT-BASED POINT-PAIR-FEATURE FOR SCALABLE AUTOMATIC GLOBAL REGISTRATION OF LARGE RGB-D SCANS

摘要

A method, system, and apparatus provide the ability to globally register point cloud scans. A first and a second three-dimensional (3D) point cloud are acquired. The point clouds have a subset of points in common and there is no prior knowledge on an alignment between the point clouds. Particular points that are likely to be identified in the other point cloud are detected. Information about a normal of each of the detected particular points is retrieved. A descriptor (that only describes 3D information) is built on each of the detected particular points. Matching pairs of descriptors are determined. Rigid transformation hypotheses are estimated (based on the matching pairs) and represent a transformation. The hypotheses are accumulated into a fitted space, selected based on density, and validated based on a scoring. One of the hypotheses is then selected as a registration.