Analytic fusion of visual cues in model-based camera tracking

Abstract

Model-based camera tracking is a technology that estimates a camera pose by tracking visual cues, i.e. points and edges on a known 3D scene model, in camera images. In model-based camera tracking, it has been a main challenge how to use the visual cues effectively for better performance. In this paper, we carefully analyze the dependency of the visual cues on tracking conditions (or environments) and propose a formula for integrating the visual cues cooperatively into a single framework based on the analysis. Then, we demonstrate that the analytic integration outperforms separate use of either cue and expedient integration of visual cues in arbitrary environments through experiments with synthetic camera images for which ground truth camera poses are given.