Process-noise adaptive particle filtering with dependent process and measurement noise

Knowledge of the noise distributions is typically key for reliable state estimation. However, in many applications only the measurement noise can be determined a priori, since only this correspond to measurable quantities. Moreover, modeling of physical systems often leads to nonlinear state-space models with dependent noise sources. Here, we design a computationally-efficient marginalized particle filter for jointly estimating the state trajectory and the parameters of the process noise, assuming dependent noise sources. Our approach relies on marginalization and subsequent update of the sufficient statistics of the process-noise parameters. Results and comparisons for a benchmark example indicate that our method gives clear improvements.