Abstract: Matching local geometric features on real-world depth images is a challenging task due to the noisy and low-resolution of 3D scan captured by depth cameras like Kinect.At present,most of the solutions to this problem are based on the feature histogram method,which requires a large amount of calculation and strict requirements on the rotation of the scene.This paper proposed a method based on data-driven.From a large number of well-reconstructed RGB-D data sets,a self-supervised deep learning method is used to construct a model that can describe the geometric correspondence between three-dimensional data.Then,corresponding approximate points of two parts of the point cloud are botained by using KD-Tree-based K Nearest Neighbor (KNN) algorithm.Through removing erroneous matching point pairs using RANSAC,a relatively accurate set of feature point pairs is obtained by estimating the geometric transformation.By regis-tering and comparing the models in the Stanford University point cloud library and the David plaster model collected in the real environment,the experiments show that the proposed method can not only extract the local geometric features of unknown objects for registration,but also show good performance when dealing with large changes in spatial angle.

Key words: Deep learning, KD-Tree, Large angle transformation, Point cloud feature registration, Self-supervised