Visual Servoing in the Large

In this paper we consider the problem of maneuvering an autonomous robot in complex unknown environments using vision. The goal is to accurately servo a wheeled vehicle to a desired posture using only feedback from an on-board camera, taking into account the nonholonomic nature of the vehicle kinematics and the limited field-of-view of the camera. With respect to existing visual servoing schemes, which achieve similar goals locally (i.e. when the desired and actual camera views are sufficiently similar), we propose a method to visually navigate the robot through an extended visual map before eventually reaching the desired goal. The map comprises a set of images, previously stored in an exploratory phase, that convey both topological and metric information regarding the connectivity through feasible robot paths and the geometry of the environment, respectively. Experimental results on a laboratory setup are reported showing the practicality of the proposed approach.