Numerically Robust SVD-based Kalman Filter Implementations

Abstract

The so-called factored-form Kalman filter (KF) implementations are designed to deal with the problem of numerical instability of the conventional KF. They include Cholesky factorization-based, UD-based and singular value decomposition (SVD) algorithms. The SVD-based estimators are the most recent developments in this realm. They were shown to be more robust with respect to roundoff than the classical KF implementation and the previously derived factored-form methods. This paper discusses further improvements in estimation accuracy and numerical robustness of the recently proposed SVD-based estimators.