An Outlier-Robust Kalman Filter Based on the Generalized <tex>$t$</tex> Distribution with a Mixed Gaussian Approximation

Xueyi Song,Enbin Song,Dunbiao Niu,Zhujun Cao,Tingting Wang
DOI: https://doi.org/10.23919/CCC58697.2023.10240760
2023-01-01
Abstract:In this paper, a novel robust Kalman filter (KF) based on the generalized <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$t$</tex> distribution is proposed to deal with measurement outliers. By ultilizing the relationship between the generalized <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$t$</tex> distribution and the Gauss-Inverse-Wishart distribution, a latent random variable is introduced, which helps us derive the accurate analytical form of the posterior density. After that, a mixed Gaussian approximation method is proposed to obtain an accurate approximation of the posterior mean and covariance. We also establish the best Gauss distribution to approximate the posterior distribution under the Kullback-Leibler (KL) divergence. Compared with the current robust KFs, our method is more implementation-friendly, and can achieve a higher accuracy due to the adoption of the latent matrix random variable and the best Gaussian approximation for the posterior distribution under the KL divergence.
What problem does this paper attempt to address?