3D Object Recognition and Registration Using Minimalist Descriptors

Abstract

This thesis presents a novel minimalist descriptor for 3D object recognition and registration problems called 3DLD (3D line descriptors). Building off of previous work on multiple point descriptors, this thesis introduces the first descriptor to apply lines to both of these problems. Each 3DLD is based on the depth information obtained between two 3D points. An efficient indexing scheme using multiple hash maps is introduced to efficiently retrieve the descriptors. From experimentation, 3DLD are very effective in both registration and recognition problems - achieving near perfect true positive rates in many of the tests. Particularly, on complex data, with cluttered and occluded scenes, 3DLD can provide superior efficiency.