EJRGF: Efficient Joint Registration of Multiple Point Clouds Using Fast Gaussian Filter

Abstract

Joint registration plays a critical role when it comes to aligning multiple point clouds. Despite its capacity to obtain unbiased solutions, current joint registration approaches face substantial computational challenges, particularly regarding processing speed and resource consumption, which impede their practical implementation with large-scale and long-sequence data. Accordingly, we present a novel probabilistic framework called EJRGF for joint registration that achieves substantially higher efficiency, reduced resource consumption, and state-of-the-art accuracy. We assume that each data point is generated from a Gaussian Mixture Model (GMM) with isotropic components and the joint registration is then reformulated as a maximum likelihood estimation problem. To solve this optimization problem efficiently, we formally derive an innovative Expectation-Maximization (EM) algorithm accelerated by a modified fast Gaussian filter based on permutohedral lattice that estimates the transformations and the GMM parameters without compromising accuracy. Furthermore, we extend our method to address global registration tasks through feature augmentation. Comprehensive experiments demonstrate that our approach outperforms state-of-the-art methods by a large margin in terms of efficiency and accuracy, achieving at least an order of magnitude acceleration compared to JRMPC.