A novel particle filtering approach and its application to target tracking

Particle filters provide asymptotically optimal numerical solutions in problems that can be cast as estimation of unobserved time-varying states of dynamic systems. Such methods rely on knowledge of the prior probability distributions of the initial state and the noise processes that affect the analyzed system, and require ability to evaluate the likelihood function and the state transition density. We describe a class of particle filtering methods that aim at the estimation of the system state from the available observations without a priori knowledge of any probability density function. When compared to the popular auxiliary bootstrap filter, in a problem consisting of tracking of a moving target in a 2D space, computer simulation results illustrate the robustness and excellent performance of the proposed algorithm.