In this thesis we propose a novel, fast and accurate 3D descriptor for use in
pose estimation of objects. To achieve this, we develop algorithm to identify
the pose of any instance of a particular object category without having its exact
model. Our aim is to estimate the object pose given only information about
its category. To achieve this, we focus on the impact of the choice of the reference
frame on the performance of a 3D descriptor. We propose a method that
estimates a reference coordinate system using a single viewpoint of the object
scene. We use an additional reference vector in order to define our reference
coordinate system uniquely. The main idea is to use the information retrieved
from a perception system collecting real world data. This information is neglected
in most of the state-of-the-art pose estimation methods. In this work,
we use the information which is given by the support plane that the objects
are lying on. Based on the reference coordinate system, we define the Pose Oriented
SEmi-global Feature Histogram (POSEFH) as an efficient 3D semi-global
descriptor which encodes the geometrical properties of the object surface. We
perform extensive qualitative and quantitative evaluation of our method on the
subset of the RGB-D Stereo Object Category database which is collected by the
7-joint Armar III robotic head with foveal and peripheral stereo cameras We
compare it with the state-of-the-art 3D descriptors, the Viewpoint Featrue Histogram
(VFH) and the Clustered Viewpoint Feature Histogram (CVFH). Our
method works for incomplete 3D pointclouds of objects. The results show a
significant improvement in the accuracy of one degree-of-freedom orientation
estimation while dealing with partial and real noisy data.
i