Learning Compact Transformation Based on Dual Quaternion for Point Cloud Registration

Accurately estimating 3-D rigid body transformation is a critical step for correspondences-free point cloud registration method. However, recently proposed methods have faced challenges in effectively estimating rigid body transformation due to issues related to parameters redundancy and singularity. In this article, we propose a new framework to estimate rigid transformation by dual quaternion which provides a compact representation for rigid transformation information. Different from traditional methods which generate dual quaternion utilizing prior knowledge, the multiscale features association network (MFANet) is introduced to adaptively learn transformation parameters of dual quaternion for accurately estimating rigid transformation. In addition, MFANet enhances data interaction between feature maps of low-dimensional and high-dimensional, which can potentially promote the learning of transformation parameters and reduce the appearance of preference features. Finally, our method demonstrates superior precision and robustness through comprehensive experiments conducted on synthetic dataset ModelNet40 and real-world dataset 3DMatch.