Particle filtering for systems with unknown noise probability distributions

In recent years particle filtering has become a powerful tool for tracking signals and time-varying parameters of dynamical systems. These methods require a mathematical representation of the dynamics of the system evolution, together with assumptions of probabilistic models. In this paper, a new class of particle filtering methods that do not assume an explicit mathematical form of the probability distributions of the noise in the system is presented. As a consequence, the proposed techniques are more robust than standard particle filters. Besides the theoretical development of a specific method in the new class, experimental results that demonstrate its performance in the problem of target tracking are provided.