Automatic Point Clouds Registration Based on Regions

We present a region-based algorithm for the automatic registration of 3D point clouds. Most existed algorithms align the two scans globally, making them unsuitable when the overlapping ratio is low or the input shapes do not have strong features. We notice that rigid transform is low-dimensional and two overlapped regions of the point clouds are enough to recover it. Thus, we align each pair of regions directly, and then solve an energy optimization to obtain the global transform from a series of region registrations by introducing confidence term and consistency term. Finally, sparse ICP algorithm is used for refinement. Experiments show that under premise of robustness to noise and outliers, our algorithm can align scans with lower overlapping ratio and more general shapes.