Robust symmetric iterative closest point

Abstract

Point cloud registration (PCR) is an important technique of 3D vision, which has been widely applied in many areas such as robotics and photogrammetry. The iterative closest point (ICP) is a de facto standard for PCR. However, it mainly suffers from two drawbacks: small convergence basin and the sensitivity to outliers and partial overlaps. In this paper, we propose a robust symmetric ICP (RSICP) to tackle these drawbacks. First, we present a new symmetric point-to-plane distance metric whose functional zero-set is a set of locally-second-order surfaces. It has a wider convergence basin and higher convergence speed than the point-to-point metric, point-to-plane metric, and even original symmetric metric. Second, we introduce an adaptive robust loss to construct our robust symmetric metric. This robust loss bridges the gap between the non-robust ℓ2 cost and robust M-estimates. In the optimization, we gradually improve the degrees of robustness via the decay of a robustness control parameter. This loss has a high “breakdown” point or low computational overhead compared with recent work (e.g., Sparse ICP and Robust ICP). We also present a simple but effective linearization for the alignment function based on Rodrigues rotation parameterization with the small incremental rotation assumption. Extensive experiments on challenging datasets with noise, outliers or partial overlaps show that the proposed algorithm significantly outperforms Sparse ICP and Robust ICP in terms of both accuracy and efficiency. Our source code will be publicly available in https://ljy-rs.github.io/web.