Neural network in the truncated point clouds registration problem

Abstract

Point cloud registration searches an orthogonal transformation to align two point clouds. Such registration is a basic task of computer vision and is common approach in 3D environment reconstruction, simultaneous localization and mapping (SLAM), and robotics. The common method to registering point clouds is the Iterative Closest Points (ICP) algorithm. Recently, registration methods have begun to use deep learning. Several types of neural net algorithms of point clouds registration have been proposed. In particular, the Deep Closest Points (DCP) network has been described. In this article, we apply a modified version of DCP to the problem of registering incongruent point clouds. Using computer simulations, we show the quality of the proposed neural network algorithm in the ModelNet40 database.