Fast And Accurate Particle Filtering For Tracking Biomedical Signals

Abstract

The key interest of this thesis is in developing fast and accurate parti cle filters for tracking biomedical signals. The particle filter is a powerful Bayesian state estimation filter that has been successfully used for a general class of nonlinear and non-Gaussian models. The particle filter operates by approximating the required probability densities using a set of particles and their associated weights. The transition and congregation of these particles lead to a probable estimation of the true target state. The first step in the filter, termed the importance sampling guides particles into probability mass re gions. The filter is effective only when the the step guides particles are drawn from regions of higher probability mass. The direct consequence of not be ing able to draw from regions of high importance is called degeneracy and overcoming the said problem is still a challenge. This thesis develops a novel particle filter that cleverly leverages its sampling on the incoming observation so to overcome degeneracy. The second step in the filter, termed the resam pling overcomes degeneracy by eliminating low weight particles. However the step is computationally very expensive. This thesis develops two novel approaches to minimise communication within the resampling process and this accelerates the filtering operation. The developed approaches are applied to accurately tracking real electrocardiogram signals which consequentlyaids in diagnosing the biomedical features of the human heart. Keywords: Bayesianfiltering, Kalman filtering, particle filtering, importance sampling, resampling, biomedical signal tracking, mean squared error. newline newline