Pairwise Registration of Low Overlapping Unorganized 3D Point Clouds Using Supervoxel Segmentation

Abstract

Since its introduction, the Iterative Closest Points (ICP) algorithm has led to developing a wide range of registration methods, most of these variations of ICP itself. Notwithstanding the efforts on improving the speed and accuracy of ICP, these variations cannot correctly align point clouds which overlapping ratio is considered low (under 40%) due to an inherited local minima convergence. Furthermore,more advanced registration techniques that rely on point descriptors also cannot overcome this problem because the tuning of their parameters tends to be volatile, which leads to making false point correspondences and consequently failing to perform an accurate registration. In order to solve this problem, we propose a pairwise registration approach that does not entirely rely on point descriptors and leverages the local minima convergence of ICP to correctly align 3D point clouds with overlapping ratios as low as about 20%.Our method uses the supervoxel segmentation technique to divide the point clouds into subsets and finds those which registration maximizes the overlapping ratio between correct correspondences in the full point clouds. We verified the effectiveness of the proposed method through tests in dense models and real-world scan datasets.