Point Data Registration With the Multi-Object, Cardinalized Optimal Linear Assignment Metric

Point cloud registration is a critical component of many tasks including the estimation of sensor motion and 3D reconstruction. The Iterative Closest Point (ICP) algorithm and its variants were initially used to solve such problems and since then, various global methods have been proposed to improve registration when data outliers exist. Realistic point cloud datasets contain detection errors (outliers) as well as spatial errors, making correct registration fragile when the outlier rejection and point data association methods fail. Therefore, in this article, a registration technique based on the multi-object Cardinalized Optimal Linear Assignment (COLA) metric is presented, which jointly penalizes both detection and spatial errors in a meaningful manner. This allows robust scan registration to take place in the presence of unknown point correspondences and inter-scan translation and orientation as well as point cloud detection and spatial errors. The resulting Particle Swarm Optimization (PSO)-COLA registration algorithm is capable of determining inter-scan point correspondences, but can also run based on point correspondences determined by other algorithms, such as the application of Fast Point Feature Histograms (FPFH) descriptors. In both cases, the PSO-COLA algorithm is shown to outperform current local and global point cloud registration algorithms in the presence of data outliers and spatial uncertainty. Multiple experiments and comparisons are based on the Stanford Bunny and Dragon, ModelNet40 and ETH Apartment datasets.