Observer Reduction via Output Stabilization with Application to Visual Motion Estimation

Stefano Soatto, Control and Dynamical Systems, Caltech

Vision provides powerful and versatile sensing which is potentially useful for control systems. One of the main obstacles in using vision as a sensor is due to the massive amounts of sensory data which are necessary to process, while the parameters of interest are often few. For instance, the motion of a scene seen from a camera has been traditionally estimated using observers (Extended Kalman Filters) with a number of states on the order of few hundreds, while the rigid-motion parameters of interest are at most six. In this talk we discuss some general tools for reducing the order of an observer. We start by revisiting and generalizing the ideas underlying the "reduced-order observer", and then use output-injection in the observed state for restricting it to submanifolds of the state-space manifold of the original model (observer decoupling). Such guidelines can be applied to problems where the few parameters of interest are trapped inside a high-dimensional dynamical model, such as in mixed estimation/identification. We apply those for estimating rigid motion from a sequence of images, generating a chain of observers with 5, 4, 3, or 2 states.