Risk Adjusted Set Membership Identification of Wiener Systems

Abstract

This technical note addresses the problem of set membership identification of Wiener systems. Its main result shows that even though the problem is generically NP-hard, it can be reduced to a tractable convex optimization through the use of risk-adjusted methods. In addition, this approach allows for efficiently computing worst-case bounds on the identification error. Finally, we provide an analysis of the intrinsic limitations of interpolatory algorithms. These results are illustrated with a non-trivial problem arising in computer vision: tracking a human in a sequence of frames, where the challenge here arises from the changes in appearance undergone by the target and the large number of pixels to be tracked.