Reliable Mobile Robot Navigation from Unreliable Visual Cues

Amy J. Briggs
Daniel Scharstein
Stephen Abbott

Abstract

Vision-based mobile robot navigation requires robust methods for planning and executing tasks due to the unreliability of visual information. In this paper we propose a new method for reliable vision-based navigation in an unmodeled dynamic environment. Artificial landmarks are used as visual cues for navigation. Our system builds a visibility graph among landmark locations during an exploration phase and then uses that graph for navigation. To deal with temporary occlusion of landmarks, long-term changes in the environment, and inherent uncertainties in the landmark detection process, we use a probabilistic model of landmark visibility. Based on the history of previous observations made, each visibility edge in the graph is annotated with an estimated probability of landmark detection. To solve a navigation task, our algorithm computes the \emph{expected} shortest paths between all landmarks and the specified goal, by solving a special instance of a Markov decision process. The paper presents both the probabilistic expected shortest path planner and the landmark design and detection algorithm, which finds landmark patterns under general affine transformations in real-time.
A gzip'ed postscript version of this paper is available.

Back to Amy Briggs' home page