Approximate Noise-Adaptive Filtering Using Student-t Distributions

We present an adaptive method for Bayesian filtering of linear state-space models with unknown noise statistics. The proposed method makes use of separation of the state and parameter posterior at each time step recursively for subsequent approximate inference. The filter exploits properties of the inverse-Wishart and the Student-t distributions and has relations to recent results from outlier-robust filtering. The method is well suited to platforms with limited computational resources because of its simplicity. Simulation results show that the proposed method can correctly estimate the measurement-noise statistics under large initial errors, in addition to being robust to outliers in the measurement and process noise.