A Generalization of Iterative Closest Point Registration Algorithms Using a Local Reference Frame-Based SE(3) Augmentation

Point cloud registration is a key computer vision task that must be solved in most modern pipelines for 3-D data processing. Iterative Closest Point (ICP) algorithms are one of the most important and widely used approaches for solving the point cloud registration problem. In this paper, we tackle the problem of increasing the convergence basin of ICP algorithms, which often require a rough initial guess to obtain good results. We propose a novel family of ICP-based algorithms that perform registration over the augmented SE(3) space by including additional information about the local geometry through Triple Orthogonal Local Depth Images (TOLDI) local reference frame (LRF) technique. The key step in our proposed method is a new correspondence search step that computes the nearest neighbors by comparing distances in the SE(3) space while effectively balancing the importance of the rotation and translation components.We thoroughly compare the proposed methods with 16 other algorithms over diverse environments, including synthetic, indoor RGBD, and outdoor LiDAR data. The three algorithms from the proposed family that we implement, namely the SE(3)-pt2pt, SE(3)-pt2pl, and SE(3)-GICP methods achieve a consistent improvement in robustness to large initial displacements compared to the related iterative closest point algorithms. Our SE(3)-GICP algorithm variant substantially outperforms competitive fine registration algorithms on synthetic problems with large initial displacements and indoor RGBD experiments, whereas it remains on par with the best methods for the outdoor LiDAR dataset