Title: Stochastic Particle Flow for Nonlinear High-Dimensional Filtering Problems

Abstract: A series of novel filters for probabilistic inference called particle flow filters, which propose an alternate convenient way of performing Bayesian updates, has been much in evidence lately. Those filters lay out a procedure to approximately solve nonlinear filtering problems by which the prior probability density is updated continuously between two subsequent time steps. We propose a novel filter inspired by the particle flow methods to address the shortcomings of sequential Monte Carlo methods in solving important nonlinear high-dimensional filtering problems that arise in multi-target multi-sensor tracking. The novel filter establishes a Monte Carlo procedure for generating a sequence of samples that guide local solutions of the Fokker-Planck equation to compose a Gaussian mixture that approximates a filtering density. We show that the new filtering technique, called Stochastic Particle Flow, is adequate for nonlinear high-dimensional problems constituting a credible estimator of the filtering density. In addition, we extend the original particle flow filters for tackling multi-target multi-sensor tracking problems to enable a comparison with the new filter.