A Registration Method for 3D Point Clouds with Convolutional Neural Network

Abstract

Viewpoint independent 3D object pose estimation is one of the most fundamental step of position based vision servo, autopilot, medical scans process, reverse engineering and many other fields. In this paper, we presents a new method to estimate 3D pose using the convolutional neural network (CNN), which can apply to the 3D point cloud arrays. An interest point detector was proposed and interest points were computed in both source and target point clouds by region growing cluster method during offline training of CNN. Rather than matching the correspondences by rejecting and filtering iteratively, a CNN classification model is designed to match a certain subset of correspondences. And a 3D shape representation of interest points was projected onto an input feature map which is amenable to CNN. After aligning point clouds according to the prediction made by CNN, iterative closest point (ICP) algorithm is used for fine alignment. Finally, experiments were conducted to show the proposed method was effective and robust to noise and point cloud partial missing.