Globally consistent rigid registration

In this paper we present a novel approach to register multiple scans from a static object. We formulate the registration problem as an optimization of the maps from all other scans to one reference scan where any map between two scans can be represented by the composition of these maps. In this way, all loop closures can be automatically guaranteed as the maps among all scans are globally consistent. Furthermore, to avoid the incorrect correspondences between the points in the scan, we employ a parametric bi-directional approach that generates invertible transformations in pairwise overlapping regions. With the parameter information in use and the consistency taken into consideration, we are able to eliminate the drift that often occurred in multi-view registration process. Our approach is fully automatic and has performed better than existing approaches by various experi-