Locally-linearized Particle Filter

Abstract

Nonlinear dynamic systems with Gaussian noise are widely used in applied fields. Extended Kalman filter (EKF) and particle filter (PF) are two traditional methods in estimating these systems. In the thesis, two algorithms called locally-linearized particle filters are proposed. Both of them combine the idea of EKF and PF. In the first algorithm, process noises are split into two parts. One part is realized in random samples, leaving out a remaining system on which EKF is applied. In the second algorithm, same procedure is conducted, but results from EKF are not the final estimates. Instead, they are used to construct a sampling scheme that generates samples from the target distribution. Results from simulation studies show that both methods gain improvements over EKF and PF.