Abstract. As a key enabling technology in 3D reconstruction and digital twin applications, multiview point cloud registration holds significant theoretical value and practical relevance. In real-world scenarios, although the relative poses between devices are typically unknown, the spatial positions and acquisition orientations of point cloud sensors are often obtainable, enabling the acquisition of prior spatial arrangement sequences for multiview point clouds. We propose a sliding-window hierarchical registration framework that leverages such prior knowledge to enhance efficiency and robustness. The framework seamlessly integrates 4PCS-based coarse alignment with a transformer-based fine registration module, and employs a sliding-window strategy to iteratively align and fuse spatially adjacent point cloud sequences. By recursively applying this process, a globally consistent, high-precision point cloud is constructed. By comparing various fine registration methods and multiview point cloud registration techniques, our method demonstrates superior performance across eight scenes from the tanks and temples, DTU, and ETH3D datasets, achieving the best registration results in six of the scenes while consistently requiring the least amount of processing time. These results demonstrate the advantages of our sliding-window hierarchical strategy, two-stage registration, and voxel-level fusion, providing an effective solution for multiview point cloud registration with prior spatial knowledge.
Paper
Full text
Hierarchical registration of multiview point clouds based on prior spatial arrangement
Semantic Scholar · Engineering · 2026
Abstract
Abstract. As a key enabling technology in 3D reconstruction and digital twin applications, multiview point cloud registration holds significant theoretical value and practical relevance. In real-world scenarios, although the relative poses between devices are typically unknown, the spatial positions and acquisition orientations of point cloud sensors are often obtainable, enabling the acquisition of prior spatial arrangement sequences for multiview point clouds. We propose a sliding-window hierarchical registration framework that leverages such prior knowledge to enhance efficiency and robustness. The framework seamlessly integrates 4PCS-based coarse alignment with a transformer-based fine registration module, and employs a sliding-window strategy to iteratively align and fuse spatially adjacent point cloud sequences. By recursively applying this process, a globally consistent, high-precision point cloud is constructed. By comparing various fine registration methods and multiview point cloud registration techniques, our method demonstrates superior performance across eight scenes from the tanks and temples, DTU, and ETH3D datasets, achieving the best registration results in six of the scenes while consistently requiring the least amount of processing time. These results demonstrate the advantages of our sliding-window hierarchical strategy, two-stage registration, and voxel-level fusion, providing an effective solution for multiview point cloud registration with prior spatial knowledge.