Convolutional neural network for 3D point clouds matching

Geometric registration is a key task in many computational fields, including medical imaging, robotics, autonomous driving. The registration involves the prediction of a rigid motion to align one point cloud to another, potentially distorted by noise and partiality. The most popular point cloud registration algorithm, Iterative Closest Point (ICP), alternates between estimating the rigid motion based on a fixed correspondence estimate and updating the correspondences to their closest matches. Recently, the success of deep neural networks for image processing has motivated an approach to learning features on point clouds. Adaptation of deep learning to analyze point cloud data is far from straightforward. Most critically, standard deep neural network models require input data with regular structure, while point clouds are fundamentally irregular: Point positions are continuously distributed in the space, and any permutation of their ordering does not change the spatial distribution. Several neural networks have recently been proposed for analyzing point clouds data such as PointNet and DGCNN. In this paper, we propose a permutation invariant neural network to identify matching pairs of points in the clouds. Computer simulation results are provided to illustrate the performance of the proposed algorithm.