Improved feature descriptors for 3D surface matching

Abstract

Our interest is in data registration, object recognition and object tracking using 3D point clouds. There are three steps to our feature matching system: detection, description and matching. Our focus will be on the feature description step. We describe new rotation invariant 3D feature descriptors that utilize techniques from the successful 2D SIFT descriptors. We experiment with a variety of synthetic and real data to show how well our newly developed descriptors perform relative to a commonly used 3D descriptor, spin images. Our results show that our descriptors are more distinct than spin images while remaining rotation and translation invariant. The improvement in performance incomparison to spin images is most evident when an object has features that are mirror images of each other, due to symmetry.