6DOF decoupled roto-translation alignment of large-scale indoor point clouds

Abstract

We address the problem of 6DOF alignment of large-scale point clouds of indoor spaces such that extensive 3D models can be assembled out of multiple point clouds. We present an algorithm that it is fast, insensitive to initial alignment and tolerates very low overlap. The algorithm is designed to exploit inherent characteristics of indoor spaces. It loosens the tight coupling between translation and rotation estimation such that these can be performed in consecutive steps with estimation problems of reduced complexity which can be reliably solved using strong features characteristic to indoor spaces. First, the point clouds are rotationally aligned to gravity using PCA. Then, the translation along gravity is computed through floor matching. Subsequently, the ground-plane rotation is determined by cross-correlating histogram signatures of surface normal directions. Finally, the ground-plane translation is determined by seeking the location where the bi-variate shift histogram of point pairs with high curvature values (called CPSHs) peaks. This voting-like approach avoids establishing correspondences through computationally demanding feature extraction and matching processes. To support very low overlap cases, the CPSH-based alignment is furthermore cast into a probabilistic framework that involves computing the CPSHs on segments of the point clouds and finally fused using an ML estimator. The results show that the proposed approach succeeds in the alignment of datasets for which general-purpose algorithms fail while being at least as efficient as the fastest methods previously proposed.