Visual Navigation: From Biological Systems to Unmanned Ground Vehicles

Edward M. Riseman,¹ Allen R. Hanson,¹ J. Ross Beveridge,² Rakesh (Teddy) Kumar, 3 and Harpreet Sawhney3 ¹ University of Massachusetts, Amherst ² Colorado State University ³David Sarnoff Research Laboratory

Navigation in large-scale space, and the execution of a coherent set of actions in this space, requires the use of some knowledge of the environment. A long-range research program in computer vision at UMass has had as one of its primary goals the integration of a diverse set of research efforts into a knowledge-based system that ultimately achieves robust, real-time navigation. The focus of this paper is on robust landmark-based navigation in the UMass mobile robot navigation project, including issues of how to acquire, utilize, extend, and modify three-dimensional models of the domain that is being navigated. Much of this chapter discusses mechanisms and algorithms that can be utilized for navigation assuming a partial geometric model is available, but the general concepts can be applied more widely and are fundamental to applications such as object recognition and three-dimensional (3D) reconstruction.

Mobile robots that are to move through the world in a goal-oriented manner must have a variety of perceptual capabilities. Some of these are "low-level" in the sense that they are focused primarily on the immediate local results of perceptual processing, often for safety (e.g., emergency stops and obstacle avoidance), and might have only secondary importance for achieving the actual global goal. For example, stereo processing might be used during navigation for obstacle avoidance and have priority in the control of the vehicle whenever obstacles are in the chosen path.

-317-