PNGOR: A Novel Guaranteed Outlier Removal Method Ensuring Robust Rotation Estimation From Planar Normals

Abstract

In this letter, we propose a guaranteed outlier removal method based on computational geometry consistency checks, named PNGOR, effectively leveraging planar normals from 3D scenes to estimate rotation. The challenge of estimating rotation can be conceptualized as a maximum consensus problem and we address it by approximating the computation through the resolution of subproblems associated with each plane correspondence. To efficiently remove outlier correspondences, our proposed method consists of three crucial modules: first, we present and validate rotation-invariant constraints to coarsely filter outlier correspondences; second, we propose a rotation-based guaranteed outlier removal method to further prune outlier correspondences, computing rotational upper and lower bounds for each inlier correspondence; finally, in the stage of computing maximum consensus, we employ an interval stabbing method to estimate the rotation. The efficacy and robustness of our proposed method were empirically validated across three extensive LiDAR point cloud datasets. Through the implementation of efficient outlier removal mechanisms and the effective suppression of rotational errors, our approach ultimately achieves state-of-the-art performance.