Multiview range image registration using the surface interpenetration measure

This paper presents a new method for the precise registration of multiple range images with low pairwise overlap. The method is based in enhanced genetic algorithms (GAs). The proposed technique minimizes the alignment error within the common overlap area among a set of views, which is computed by a novel robust figure of merit called the surface interpenetration measure (SIM). The key idea behind this measure is the observation that mean squared error alone is insufficient to evaluate the quality of a registration solution because it fails to account for the spatial distribution of the error. The new measure explicitly forces the solution to distribute the errors across the overlap area, producing more stable, reliable solutions that limit propagation and amplification of error in the multiview problem. Our approach is not dependent on GAs as the search mechanism, but because they search in a space of transformations, GAs are capable of registering surfaces with no need for prealignment. The need for prealignment is a major weakness of methods based on the iterative closest point (ICP) algorithm, the most popular family of methods to date. The experimental results confirm that the new method ensures more precise global alignments than combined sequential pairwise alignments for registration.