Multi-view visual servoing using epipoles

We explore the benefits of multiple views for visual servoing (VS) by commanding a single camera to first "peer" at a scene from two vantage points, thus acquiring a set "reference" images, prior to executing a visual position task. Our approach completely decouples the translational and rotational components of our controller: epipoles from the reference views drive the translational error to zero, while the rotational degrees of freedom maintain all of the features in the field of view (FOV). We furnish a simple Lyapunov stability proof that demonstrates a large domain of attraction while maintaining all features in the FOV. Finally, we present simulated experiments that suggest robustness to measurement noise and large variations in the baseline between the reference views.