3D Matched Manifold Detection for Optimizing Point Cloud Registration

Point cloud registration is usually performed by matching key points to obtain an approximate global alignment, followed by a local optimization algorithm such as the iterative closest point (I CP) and its variants, to refine the initial estimate. These refinement algorithms, however, converge in many cases to a false local extremum. We propose a new matched manifold detection approach over the group of rigid 3-D transformations, by employing a novel correlation operator between functions defined on sparsely and non-uniformly sampled point clouds. Correlation between point clouds is evaluated using a method-ology inspired by the definition of the Kernel Point Convolution (KPConv), but instead of performing convolution with a kernel, the inner-product of feature vectors evaluated on the points in the two point clouds are aggregated. The proposed approach is shown to outperform state of the art local registration methods in terms of accuracy on challenging data sets.