Patch-guided point matching for point cloud registration with low overlap

Abstract

Point cloud registration is a classic and fundamental problem. Existing point cloud registration methods obtain correspondence point pairs by calculating the correlation between point features. However, the instability of point features makes the outlier rate of corresponding point pairs high, resulting in poor matching results, especially when facing low overlap point clouds. An obvious fact is that patch matching has higher reliability than point matching. To this end, we propose a patch-guided point cloud registration network. Specifically, we perform fusion on patches and points at both the feature and result levels to achieve the guidance of patch to point matching and improve the accuracy of predicted point pairs. At the feature level, we propose a Matching Pyramid Network (MPN) for multi-level patch/point matching. The core of the MPN is an attention-based cross-layer context aggregation (CCA) module, which is used for the fusion of matching features between upper and lower layers. At the result level, we design a matching consistency judgment module to ensure that the point pairs are consistent in the matching of each layer, which greatly reduces the outlier ratio. Based on the above design, the corresponding point pairs predicted by our network have a high inlier ratio, which makes our method perform well in the face of low overlapping point clouds. Extensive experimental results show that our method outperforms other existing methods for indoor and outdoor datasets.