Non-Uniformity of a Pattern and the Best Single View 3D Pose Estimator

                         Georgii Khachaturov

             Universidad Autonoma Metropolitana, Mexico

            Home page http://www.tarunz.org/~xgeorge
                 or : http://newton.uam.mx/xgeorge

                              Abstract

This work was motivated by necessity of robust, precise, and
algorithmically simple and fast estimation of full pose (six
parameters) of a target by a visually guided robot. There are a huge
number of approaches for such a problem. How to introduce an order for
them and choose a best solution?

Visual model of an object is presented as composition of a model-shape
function and a model-pattern function. Non-uniformity of a pattern is
defined as a quadratic form related to the gradient of intensity. For
a planar model, it is proved that a more precise estimation responds
to a greater non-uniformity. This is applied to develop a class of
single-view 3D-pose estimators based on Fourier technique able to
explore extremely non-uniform patterns. The estimators are of high
precision, fast, robust, and algorithmically simple. Possible
application is spacecraft docking, or any robot problem permitting
usage of a visual mark (target) on remote object.