Robust Bayesian Filtering and Smoothing Using Student's t Distribution

Michael Roth,Tohid Ardeshiri,Emre Özkan,Fredrik Gustafsson

DOI: https://doi.org/10.48550/arXiv.1703.02428

2017-03-07

Methodology

Abstract:State estimation in heavy-tailed process and measurement noise is an important challenge that must be addressed in, e.g., tracking scenarios with agile targets and outlier-corrupted measurements. The performance of the Kalman filter (KF) can deteriorate in such applications because of the close relation to the Gaussian distribution. Therefore, this paper describes the use of Student's t distribution to develop robust, scalable, and simple filtering and smoothing algorithms. After a discussion of Student's t distribution, exact filtering in linear state-space models with t noise is analyzed. Intermediate approximation steps are used to arrive at filtering and smoothing algorithms that closely resemble the KF and the Rauch-Tung-Striebel (RTS) smoother except for a nonlinear measurement-dependent matrix update. The required approximations are discussed and an undesirable behavior of moment matching for t densities is revealed. A favorable approximation based on minimization of the Kullback-Leibler divergence is presented. Because of its relation to the KF, some properties and algorithmic extensions are inherited by the t filter. Instructive simulation examples demonstrate the performance and robustness of the novel algorithms.