HarSoNet: a two-stage point cloud registration method integrating soft and hard matching

The purpose of point cloud registration is to determine the transformation parameters among multiple partially overlapping point clouds, and it plays an important role in various scenarios such as simultaneous localization and mapping (SLAM), scene reconstruction, industrial manufacturing and so on. However, due to the unordered and irregular nature of point clouds, accurately establishing correspondences poses a significant challenge. Coarse-to-fine methods, consisting of coarse and fine matching stages, have become popular in point cloud registration due to their effectiveness in handling repeatable keypoints. However, these methods are highly sensitive to the correspondences generated during the coarse-matching stage, where low-quality correspondences can lead to complete registration failure. Furthermore, the hard matching approach employed in coarse and fine matching stages often introduces a large number of outliers into the established correspondences. To overcome these challenges, this study introduces HarSoNet, a two-stage Hard-to-Soft Network designed for end-to-end point cloud registration. In the hard matching stage, the model incorporates a hybrid similarity fusion module, which combines similarity scores obtained from different algorithms to establish superpoint correspondences. These superpoint correspondences, along with their neighboring points, are then grouped into fuzzy patch correspondences. In the soft matching stage, patch correspondences are refined into point correspondences by calculating and adjusting the similarity matrix for each fuzzy patch. Finally, all local point correspondences are aggregated into global correspondences, and the transformation parameters are computed using the weighted singular value decomposition (SVD) algorithm. Experimental results demonstrate that HarSoNet achieves Error(R) = 1.376 and Error(t) = 0.015 on noisy, partially overlapping point clouds, demonstrating high registration accuracy and strong generalization performance.