Research On Probability Hypothesis Density Filtering Alogrithm With Unknown Noise Statistics

Based on random finite set statistics,probability hypothesis density(PHD)filtering algorithm,as a multi-target tracking algorithm,has attracted much attention and has gained rapid development and recognition in recent years.The conventional PHD filtering algorithm needs to know noise statistics as a priori assumption.Unknown or inaccurate noise statistics will result in a decline in the tracking performance of the multi-target filtering algorithm.In the increasingly complex battlefield environment,however,the prior noise statistics are usaually not available and vary over time.Therefore,this thesis focused on multi-target PHD filter with unknown noise statistics.The main research and contributions are summarized as follows:i.A Sage-Husa adaptive noise statistics estimate based PHD filtering algorithm was proposed in order to solve the multi-target tracking problem with unknown noise statistics under linear condition.This algorithm used the adaptive method of Sage and Husa to estimate noise statistics,which were then used to modify posterior PHD update formulas.This algorithm implemented jointly estimation of multi-target states,numbers and noise statistics.Simulations demonstrate that with unknown noise statistics,this algorithm has a better tracking performance than conventional PHD filtering algorithm.Compared with variational Bayesian based adaptive PHD filtering algorithm,the proposed algorithm can make the estimation error decrease more than 30% and save up to 70% runtime.ii.A noise statistics adaptively estimates based UK-PHD filtering algorithm was proposed in order to solve the multi-target tracking algorithm with unknown noise statistics under nonlinear condition.This algorithm simultaneously estimated multi-target states and noise statistics in the nonlinear Gaussian system based on unscented transformation and maximum a posteriori estimate principle.This algorithm improved the multi-target tracking accuracy.Simulation results show that this algorithm achieves good tracking performance when measurement noise statistics are unknown.