Particle Gaussian Mixture Filters for Nonlinear Non-Gaussian Bayesian Estimation

Nonlinear filtering is the problem of estimating the state of a stochastic nonlinear dynamical system using noisy observations. It is well known that the posterior state estimates in nonlinear problems may assume non-Gaussian multimodal probability densities. We present an unscented Kalman-particle hybrid filtering framework for tracking the three dimensional motion of a space object. The hybrid filtering scheme is designed to provide accurate and consistent estimates when measurements are sparse without incurring a large computational cost. It employs an unscented Kalman filter (UKF) for estimation when measurements are available. When the target is outside the field of view (FOV) of the sensor, it updates the state probability density function (PDF) via a sequential Monte Carlo method. The hybrid filter addresses the problem of particle depletion through a suitably designed filter transition scheme. The performance of the hybrid filtering approach is assessed by simulating two test cases of space objects that are assumed to undergo full three dimensional orbital motion. Having established its performance in the space object tracking problem, we extend the hybrid approach to the general multimodal estimation problem. We propose a particle Gaussian mixture-I (PGM-I) filter for nonlinear estimation that is free of the particle depletion problem inherent to most particle filters. The PGM-I filter employs an ensemble of randomly sampled states for the propagation of state probability density. A Gaussian mixture model (GMM) of the propagated PDF is then recovered by clustering the ensemble. The posterior density is obtained subsequently through a Kalman measurement update of the mixture modes. We prove the convergence in probability of the resultant density to the true filter density assuming exponential forgetting of initial conditions by the true filter. The PGM-I filter is capable of handling the non-Gaussianity of the state PDF arising from dynamics, initial conditions or process noise. A more general estimation scheme titled PGM-II filter that can also handle non-Gaussianity related to measurement update is considered next. The PGM-II filter employs a parallel Markov chain Monte Carlo (MCMC) method to sample from the posterior PDF. The PGM-II filter update is asymptotically exact and does not enforce any assumptions on the number of Gaussian modes. We test the performance of the PGM filters on a number of benchmark filtering problems chosen from recent literature. The PGM filtering performance is compared with that of other general purpose nonlinear filters such as the feedback particle filter and the log homotopy based particle flow filters. The results also indicate that the PGM filters can perform at par with or better than other general purpose nonlinear filters such as the feedback particle filter (FPF) and the log homotopy based particle flow filters. Based on the results, we derive important guidelines on the choice between the PGM-I and PGM-II filters. Furthermore, we conceive an extension of the PGM-I filter, namely the augmented PGM-I filter, for handling the nonlinear/non- Gaussian measurement update without incurring a large computational penalty. A preliminary design for a decentralized PGM-I filter for the distributed estimation problem is also obtained. Finally we conduct a more detailed study on the performance of the parallel MCMC algorithm. It is found that running several parallel Markov chains can lead to significant computational savings in sampling problems that involve multi modal target densities. We also show that the parallel MCMC method can be used to solve global optimization problems