CPCRNet: Cross-source Point Cloud Completion-registration Network for Intelligent Factory Virtual Scene Construction

Point cloud registration is a fundamental problem in computer vision, extensively used in applications such as 3D reconstruction and synchronous positioning. Furthermore, it offers a viable solution for constructing virtual scenes in factories. Despite the emergence of some point cloud registration algorithms with excellent performance, their effectiveness deteriorates substantially when dealing with point clouds from disparate sources. To address this issue, we propose a cross-source point cloud completion-registration network(CPCRNet) tailored for robotic arms. We give a preprocessing method to solve the cross-source point cloud problems and leverage an encoder-decoder network to complete the original point cloud, then use the global optimal algorithm to register. The shape completion module of our network generates training datasets to rectify and augment incomplete point clouds. It obviates the need for collecting extensive datasets, facilitates the construction of standard virtual scenarios. In scenarios involving cross-source point cloud registration with limited overlap, our proposed method effectively enhances the registration accuracy.