A probabilistic observer for visual tracking

This paper describes an observer for estimation of the rigid pose and shape states associated to an object being tracked in an image sequence. The defined observer utilizes standard estimation strategies for the finite-dimensional rigid pose sub-states, and a novel strategy for the shape sub-states. In particular, the shape sub-state observer utilizes an implicit probability field, where the 50% probability iso-contour defines the object shape. A general purpose second-order model and a corresponding correction scheme are defined for the observer state. The observer is applied to recorded imagery and its performance is examined using objective error metrics.