The Automatic Point Clouds Registration Based On Regions

The automatic registration of3D point clouds is a fundamental problem in digital geometry processing, and has been widely used in reverse engineering and shape modeling. It brings a scan into the same coordinate of the other by a rigid transform. The two point clouds are overlapped, but neither the overlapping ratio nor the overlapped region is known. Also, their initial relative positions are usually arbitrary. Most existed algorithms align the two scans globally, making them unsuitable when the overlapping ratio is low or the input shapes have no strong features.We notice that rigid transform is a low-dimensional linear transform, having only six degrees of freedom, and it can be recovered by two overlapped regions of the scans. Thus, we propose a region-based algorithm for the automatic registration of3D point clouds. Firstly, we segment the two scans into multiple regions respectively, and align each pair of regions directly. Then, we construct an energy function by introducing confidence term and consistency term, and then maximize it to obtain the global rigid transform from a series of region registrations. Finally, sparse ICP is used for refinement.Experiments show that as our algorithm registers the point clouds on the level of regions and does not need to compute the point’s local feature, it is robust to noise and outliers. Furthermore, because the energy function we define makes the overlapped regions have greater role in the registration process, our algorithm can align scans with less overlapping ratio and more general shapes.