Accurate and robust registration of low overlapping point clouds

Abstract

Point cloud registration has various applications within the computer-aided design (CAD) community, such as model reconstruction, retrieving, and analysis. Previous approaches mainly deal with the registration with a high overlapping hypothesis, while few existing methods explore the registration between low overlapping point clouds. However, the latter registration task is both challenging and essential, since the weak correspondence in point clouds usually leads to an inappropriate initialization, making the algorithm get stuck in a local minimum. To improve the performance against low overlapping scenarios, in this work, we develop a novel algorithm for accurate and robust registration of low overlapping point clouds using optimal transformation. The core of our method is the effective integration of geometric features with the probabilistic model hidden Markov random field. First, we determine and remove the outliers of the point clouds by modeling a hidden Markov random field based on a high dimensional feature distribution. Then, we derive a necessary and sufficient condition when the symmetric function is minimized and present a new curvature-aware symmetric function to make the point correspondence more discriminative. Finally, we integrate our curvature-aware symmetric function into a geometrically stable sampling framework, which effectively constrains unstable transformations. We verify the accuracy and robustness of our method on a wide variety of datasets, particularly on low overlapping range scanned point clouds. Results demonstrate that our proposed method attains better performance with higher accuracy and robustness compared to representative state-of-the-art approaches.