Verification of multi-view point-cloud registration for spherical harmonic cross-correlation

Abstract

Spherical harmonic cross-correlation is a robust registration algorithm that brings two point-clouds of the same scene into coarse rotational alignment. The found rotation however may not give the desired alignment, as misalignments can occur if there is not enough overlap between point-clouds, or if they contain a form of symmetry. We propose a verification method whose purpose is to determine if registration has failed for a priori unknown registration. The rotational transformation between multiple clouds must satisfy internal consistency, namely multiple rotational transformations are transitive. The rotation verification is performed using triplets of images, which are cross-referenced with each other to classify rotations individually. Testing is performed on a dataset of a priori known registrations. It is found that when the number of images or the percentage of correct rotations is increased, the number of correct rotation classifications improves. Even when tested with only four images and a correct rotation percentage of 17%, the rotation verification is still considered a viable method for classifying rotations. Spherical harmonic cross-correlation is benefited by rotation verification as it provides an additional approach for checking whether found rotations are correct.