Qualitative visual servoing for navigation

We propose in this article a novel approach for vision-based control of a robotic system during a navigation task. This technique is based on a topological representation of the environment in which the scene is directly described within the sensor space, by an image database acquired off-line. Before each navigation task, a preliminary step consists in localizing the current position of the robotic system, This is realized through an image retrieval scheme, by searching within the database the views that are the most similar to the one given by the camera. Then a classical shortest path finding algorithm enables to extract from the database a sequence of views that visually describe the environment the robot has to go through in order to reach the desired position. This article mainly focuses on the control law that is used for controlling the motions of the robotic system, by comparing the visual information extracted from the current view and from the image path. This control law does not need a CAD model of the environment, and does not perform a temporal path planning. Furthermore, the images from the path are not considered as successive desired positions that have to be consecutively reached by the camera. The qualitative visual servoing scheme proposed, based on cost functions, ensures that the robotic system is always able to observe some visual features initially detected on the image path. Experiments realized in simulation and with a real system demonstrate that this formalism enables to control a camera moving in a 3D environment.