Model extension and refinement using pose recovery techniques

Abstract

Abstract Visual measurements of modeled 3‐D landmarks provide strong constraints on the location and orientation of a mobile robot. To make the landmark‐based robot navigation approach widely applicable, it is necessary to automatically build the landmark models. A substantial amount of effort has been invested by computer vision researchers over the past 10 years on developing robust methods for computing 3‐D structure from a sequence of 2‐D images. However, robust computation of 3‐D structure, with respect to even small amounts of input image noise, has remained elusive. The approach adopted in this article is one of model extension and refinement. A partial model of the environment is assumed to exist and this model is extended over a sequence of frames. As will be shown in the experiments, prior knowledge of the small partial model greatly enhances the robustness of the 3‐D structure computations. The initial 3‐D model may have errors and these are also refined over the sequence of frames. © 1992 John Wiley & Sons, Inc.