CatMat: 3D Object Recognition Using Catenarian Matching

Abstract

Object recognition in 3D point clouds is an important and widely researched topic. We propose a novel method based on local point descriptors. We detect edge points on the scene and object clouds, and construct a weighted edge graph on the object clouds. We find point chains on the objects based on the constructed graph, and seek similar point chains on the scene cloud using local descriptor matching and geometric constraints. We estimate transformations using corresponding point chains, and validate the transformations with a voxel-based method. Our method is capable of multi-instance object recognition. In this paper we present our method and compare it with a similar solution. Based on our evaluation, the proposed method is able to find various objects on scene clouds and robust to noise.