Inlier extraction for point cloud registration via supervoxel guidance and game theory optimization

Abstract

As a key step in Six-Degree-of-Freedom (6DoF) point cloud registration, 3D keypoint technique aims to extract matches or inliers from random correspondences between the two keypoint sets. The major challenge in 3D keypoint techniques is the high ratio of mismatched or outliers in random correspondences in real-world point cloud registration. In this paper, we present a novel inlier extraction method, which is based on Supervoxel Guidance and Game Theory optimization (SGGT), to extract reliable inliers and apply for point cloud registration. Specifically, to reduce the scale of keypoint correspondences, we first construct powerful groups of keypoint correspondences by introducing supervoxels, which involves 3D spatial homogeneity. Second, to select promising combined groups, we present a novel ‘fit-and-remove’ strategy by incorporating 3D local transformation constraints. Third, to extract purer inliers for point cloud registration, we propose a grouping non-cooperative game algorithm, which considers the relationship between the combined groups. The proposed SGGT, by eliminating the mismatched combined groups globally, avoids the false combined groups that lead to the failed estimation of rigid transformations. Experimental results show that when processing on large keypoint sets, the proposed SGGT is over 100 times more efficient compared to the stat-of-the-art, while keeping the similar accuracy.