Efficient tracking of objects with arbitrary 2D motions in space-variant imagery

Abstract

In the last few years, the biologically-motivated log-polar foveal imaging model has attracted the attention from the robot and computer vision communities. The advantages that log-polar images offer for target tracking and other activevision tasks have been explored and exploited. However, despite the progress made, the proposed systems have some limitations, or rely on some assumptions which may be violated in practical real scenarios. Estimating complex parametric motion models in space-variant imagery is an example of one problem still deserving further investigation, and just the focus of this paper. The proposed algorithm learns to relate a set of known motions with their corresponding visual manifestations for a given reference pattern. After learning, motion estimation becomes a very efficient, search-free process. Synthetic and real experiments provide evidence on the effectiveness of this approach for motion estimation and visual tracking using log-polar images.