A hierarchical multiview registration framework of TLS point clouds based on loop constraint

Automatic registration of multiple point clouds is a significant preprocessing step for 3D computer vision tasks including semantic segmentation, 3D modelling, change detection, etc. Many methods were proposed to deal with this problem and yet most of them are not fully utilizing the redundant information offered by multiple common overlaps among point clouds. The existing methods are also inefficient when dealing with large-scale point clouds. In this paper, a novel automatic registration framework is presented to align point clouds efficiently and robustly. First, the overall number of scans is grouped into several scan-blocks by a proposed blocking strategy, and we build the pairwise relationship among scans through a fully connected graph in each scan-block. Second, perform loop-based coarse registration in each scan-block using a proposed false matches removal strategy. The proposed strategy can effectively identify grossly wrong scan-to-scan matches. Third, the minimum spanning tree is extracted from the graph, and ICP is applied along its edges. Moreover, the Lu-Milios algorithm is used to further optimize all poses at once by utilizing all redundant information in each scan-block. Finally, global block-to-block registration aligns all scan-blocks into a uniform coordinate reference. We test our framework on challenging WHU-TLS datasets, ETH datasets, and Robotic 3D Scan datasets to evaluate the efficiency, accuracy, as well as robustness. The experiment results show that our method achieves the state-of -the-art accuracy, while the time performance is improved by more than 30% compared with the state-of-the-art algorithms. Our source code is made available at https://github.com/WuHao-WHU/HL-MRF for benchmarking purposes.