Global Registration of Multiview Unordered Forest Point Clouds Guided by Common Subgraphs

Abstract

To register multiview, unordered point clouds from forest scenes, we must establish how scans are associated. The proposed method, called RegisMUF, can register arbitrary forest point clouds from aiming scenarios without knowledge of initial position and orientation, without requiring artificial targets, and without recording the order of the scanning sequents. One of the novel contributions of the proposed method is the optimization of a scanning network. We exploit common subgraphs to connect data between spatial subsets and subsequently predict the overlapping areas between adjacent scenarios. In parallel, we propose a rapid coarse strategy and an accurate tree-oriented refinement strategy. Finally, all the scenarios converge to an anchor by combining the minimum loop expansion approach and a parallel merging approach. We experimentally evaluate three challenging data sets: one data set from the FGI benchmark Evo, Finland with five scans, and two other forest data sets from Jiangxi province, China, with 15 and 23 scans. For each data set, the network-building module of RegisMUF produced 100% correct connections to associate all the scans. Together, 108 pairwise cases are exploited to evaluate the matching module in RegisMUF, and the results reveal that the proposed method is superior to or on par with state-of-the-art practices in terms of registration accuracy and successful-registration rate. In end-to-end tests, RegisMUF performs impressively both in terms of registration accuracy and computational cost.