Perception planning and execution control for a visual and landmark-based navigation task

Abstract

This paper deals with indoor navigation based on visual landmarks. The robot is equipped only with a PTZ camera (pan, tilt and zoom control), used in a learning step, to detect and learn landmarks such as posters, signs or other characteristic objects in the environment. To reach a goal the robot generates a path, so it could always locate itself with respect to one or two landmarks. For this work we propose a strategy for optimally control of the PTZ camera, thus saccadic motions are minimized during the transition steps between the trajectory stages controlled by different landmarks. Depending on the position of learnt landmarks a perception planner selects along the path, the better landmarks to be tracked and to be used to control the robot motion. During the execution of this plan, some unforeseen events (occlusions, tracker disruptions, obstacle avoidance) make mandatory to modify either the perception plan or the generation of a new plan. Several tracking modalities can be selected according to the landmark characteristics, also to the current illumination conditions. The strategy we are proposing is initially validated, using a simulation environment, then, on a robot equipped with a Sony PTZ camera.