Research On Registration Of Unorgnized Point Clouds

3D reconstruction technique is widely used in reverse engineering, digital heritagepreservation, 3D animation, 3D motion capture etc. This technique will be demandedmore and more as digital technology and computer technology developing. Point cloudregistration is one of the key problems in the 3D reconstruction. The application ofautomatic algorithms has been restricted because of poor stability, slow convergencespeed. Current commercial products almost use the method of manual registration,putting sign points on reconstructed object, or calibrating the movement between twoscans. For the development of desktop laser scanner, this dissert aims on the research ofautomatic algorithm for point cloud registration, focus on local and global registration,and presents automatic registration scheme with high-precision, rapid convergence, wideconvergence region.Global registration of two point sets at any position provides initial values for thelocal registration. In this paper, 4-points Congruent sets method (noted as 4-PCS),presented by Mitra in 2008 is realized for global registration. 4-PCS method is based onRANSAC scheme, and fastened by inducing the ratio between segments generated by thediagonal intersection of these four planar points. Manual registration is also realized forhard registration of low overlapping point sets.Local registration starts with an estimation of two point clouds position, and thenrefines this estimate to better register the two point clouds. We analyze and comparesome improvement strategies of the classical ICP (iterated closest point) algorithm, andpresent HT-ICP to improve the accuracy, convergence rate and convergence region basedon Hong-Tan positioning algorithm. And then, we present GHT-ICP algorithm to widenthe convergence region of HT-ICP by using the local geometric features to reject wrongcorrespondent point pairs..Abundant experiments show that the presented scheme for automatic registrationperforms well when overlapping area of tow point clouds is over 40%. Manualregistration gives the solution for other hard problems witch can not be registered byautomatic algorithm.