“…A first such approximation is extended Kalman filtering [1], [7], based on a linearization of the model around its estimated trajectory in state space. The fact that this approach cannot accurately describe situations where the model is highly nonlinear and the pdf of the state is far from Gaussian has motivated the development of grid-based methods [4], [21], where the state space is partitioned a priori into cells and integrals are replaced by discrete approximations. This approach lacks flexibility, as it does not allow the partition to be adapted dynamically so as to get more resolution in regions with high probability.…”