Application of Unscented Transformation for Nonlinear State Smoothing

Motivated by the well-known fact that the state estimate of a smoother is more accurate than that of the corresponding filter, this paper is concerned with the state smoothing problem for a class of nonlinear stochastic discrete systems. Firstly, a novel type of optimal smoother, which provides a unified theoretical framework for the solution of state smoothing problem no matter that system is linear or nonlinear, is derived on the basis of minimum mean squared error (MMSE) estimation theory. Further, in the case that the dynamic model and measurement functions are all nonlinear, a new suboptimal smoother is developed by applying the unscented transformation for approximately computing the smoothing gain in the optimal smoothing framework. Finally, the superior performance of the proposed smoother to the existing extended Kalman smoother (EKS) is demonstrated through a simulation example.