A Novel Adaptive Marginalized Particle Filter for Mixed Linear / Nonlinear State-Space Models

Abstract

The speed with which the state parameters can be estimated with acceptable variance is an important issue in any target tracking application. In most of these problems, a linear Gaussian substructure is present in the model which is made use of and estimation is usually done using Marginalized Particle Filter (MPF) than the most general Particle Filter (PF). An important parameter that affects the performance of such filters is the amount of noise present in the measurement which may not remain constant, varying between a low value and a high value. The main aim of this paper is to introduce a new filter known as Adaptive Marginalized Particle Filter (AMPF) which exploits this property of noise thereby adapting the number of particles whenever possible and hence improving the estimation speed. The proposed filter is able to estimate the state parameters much faster than the existing filters. The performance of AMPF is compared with that of MPF and PF using a typical target tracking example. From the simulation, it can be concluded that AMPF is better than MPF and PF in terms of execution time, without much affecting its performance in terms of RMSE.