Using monocular SLAM for position-based visual control

Most of the existing visual control algorithms make use of pairwise geometry constraints to define the relation between the control input of the robot and the dynamics of tracked features in an image. The assumption is that feature correspondences will be available between the current image and the goal image, which do not always hold. For example, if a non-holonomic robot has to turn a large angle to reach a goal image, it most surely will lose track of this goal image. As another limitation, the use of pairwise geometry needs to change its underlying model depending on the geometric configuration of the current pair of frames –usually, from fundamental to homography matrix. In order to cope with these two limitations, this paper proposes the use of geometric maps from SLAM (Simultaneous Localization and Mapping) for visual control. A SLAM map summarizes feature tracks by registering them, along with the camera position, in a 3D common reference frame. Even when a feature goes out of sight and the track is lost; it remains registered in the 3D scene and hence usable for the control. Using a map also makes the control independent of the geometric configuration of two particular frames. As a proof of concept, we present two experiments: In the first one, a low-cost robot (build with Lego NXT and equipped with a 320x240 black-and-white camera) navigates around an object only relying on monocular information and even when the object comes out of view in the first frames of the input image sequence. In the second one, the robot is able to go back to an initial position without presenting degeneracies.