Dynamic stereo vision.

Abstract

An important problem in dynamic vision is integration of the spatial information that is available to a camera system as it moves in a stationary environment. Like most other problems, the representation of information is the dominant factor in forming a solution. Many previous approaches in dynamic vision were formulated using a viewer-centered representation. This dissertation presents the foundations of a new general approach that exploits the relative properties of stationary objects. As part of this approach we propose a new representation scheme called Referent-centered representation which describes the spatial arrangement of the scene in terms of fixed referents in the environment rather than the moving viewer. Our approach is presented in terms of a computational model for dynamic vision, and is supported by mathematical, computational and psychophysical evidence. The most important aspect of this model is that it allows the integration of spatial information over time without the explicit computation of camera motion. We describe both monocular and binocular dynamic vision systems, designed based on this model, which analyze time-varying images and integrate spatial information over time. These systems are implemented and tested on real stereo image sequences, and numerical and visual proofs are presented showing that the computational model is both practical and robust.