Structure and motion estimation in perspective systems using a dynamic vision parametrization

Abstract

Estimation of 3D structure and motion from 2D images in computer vision systems can be performed using a nonlinear dynamic system, often referred to as a perspective dynamic system. In this paper we describe how a specific parametrization of the perspective dynamic system can be utilized when formulating estimation problems for structure and motion. The parametrization allows for a single estimation problem formulation which is applicable to structure estimation as well as motion estimation. The parameters to be estimated appear explicitly in the resulting dynamic system, and available partial knowledge of parameters can be taken into account in a straightforward manner. The problem formulation allows estimators for structure and motion to be derived using available methods from nonlinear and adaptive control. We demonstrate how estimators for structure and motion can be constructed based on the parametrization, and illustrate the estimation performance by simulations. In this way, it is demonstrated how a nonlinear observer can be used for motion estimation as well as recovery of three-dimensional position in a monocular vision system, using measurements from two-dimensional images.