On the treatment of process noise in the unscented Kalman filter

A simple simulation of a turning target in two dimensions is used to show that the standard treatment of process noise in the Unscented Kalman Filter (UKF) can give rise to a potentially fragile tracking filter. In the example used, the turn rate is tracked as part of the state vector and additional state components are introduced to accommodate the process noise terms. This approach works well only while the underlying turn rate is constant or linear with time. With more complex turn rate dynamics, the filter breaks regardless of the process noise settings and such behaviour is indicative of a filter with zero process noise. It is found that a sequential Monte Carlo implementation of process noise gives rise to a much more robust tracking filter.