Using the Bhattacharyya distance in Functional Sampling Density of Particle Filter

The particle filter for nonlinear state estimation of discrete time dynamic stochastic systems is treated. The functional sampling density of the particle filter strongly affecting estimate quality is studied. The density is given by weighted mixture of the transition probability density functions. The weights are calculated using distance of two reference variable probability density functions representing prior and measurement information. The aim is to find a suitable distance that does not suffer from problems with its numerical computation and that can be computed for a large set of systems analytically. It seems that the Bhattacharyya distance is feasible for evaluation of such a distance. Quality and computational demands of the functional particle filter with primary weights computed using the Bhattacharyya distance are illustrated in a numerical example.