Non-linear filtering based on observations from Gaussian processes

We consider a class of non-linear filtering problems, where the observation model is given by a Gaussian process rather than the common non-linear function of the state and measurement noise. The new observation model can be considered as a generalization of the standard one with correlated measurement noise in both time and space. We propose a particle filter based approach with a measurement update step that requires a memory of past observations which can be truncated using a moving window to obtain a finite-dimensional filter with arbitrarily good accuracy. The validity of the conceptual solution is proved via simulations on a one dimensional tracking problem and implementation issues are discussed.