Gaussian Mixture Model-Based Registration Network for Point Clouds with Partial Overlap

Mainstream methods of point cloud registration can be divided into two categories: strict point-level correspondence, which is commonly used but incompatible with real-world data; and statistical calculations, which compensate for the shortcomings of point-level methods but are inflexible, mainly when applied to scenes containing partial overlap. This paper proposes a novel registration network (poGMM-Net), the first statistical registration method to successfully align two partially overlapping point clouds. Specifically, our model modifies the registration problem to involve the minimization of Kullback-Leibler divergence in Gaussian mixture models (GMMs), focusing on overlapping regions. In poGMM-Net, the GMMs are associated with points in the point clouds by the learned potential correspondence matrix. The fitting of nonoverlapping points and outliers is avoided by fusing learned secondary feature sets. Application of models to ModelNet40 datasets demonstrated that poGMM-Net achieves state-of-the-art performance under various registration conditions, outperforming both point-level-based and statistical methods.