Rigid registration of noisy point clouds based on higher-dimensional error metrics

Abstract

Methods based on distance error metrics, such as the iterative closest point (ICP) algorithm and its variants, do not efficiently register noisy point clouds. In this paper, we propose a novel method for registering noisy point clouds by extending the ICP algorithm. The proposed method, which is based on higher-dimensional error metrics minimization, has two variants: One variant is based on area error metric, and the other is based on volume error metric. For the registration of point clouds, triangles or tetrahedrons are constructed between the point clouds by using an optimal vertices selection algorithm. To reduce computational complexity, the method is linearized by assuming that the rotation angle is small. The main advantage of the proposed method is its robustness for the registration of noisy point clouds. In particular, the volume minimization-based registration variant exhibits good robustness in the presence of strong noise. The proposed method was compared with the variants of ICP algorithm in experiments conducted on many types of point clouds, such as noisy point clouds with different noise levels. The experimental results obtained show that the robustness of the registration is increased by using higher-dimensional error metrics.