Together with Dr. Richard Dearden, I have developed the Gaussian particle fiter
(GPF) (see my papers).
It generalizes the Rao-Blackwellised particle filter (RBPF) to the non-linear
case, approximating the belief state for every particle by a Gaussian that is
propagated using an unscented Kalman filter for each particle.
For linear Gaussian systems, GPF is equivalent to RBPF.
Obviously, it is a biased filter for non-linear, non-Gaussian systems, but our
experiments show that this bias is very small in the domains we study, and that
the results of GPF are much better than for standard PF. The best-performing
variant in our experiments is look-ahead GPF (la-GPF), the generalization of
la-RBPF.
Here's the Matlab code for GPF,
including a demo that compares GPF, la-GPF, PF and the unscented particle filter
(UPF).
It includes models for real data, a non-linear toy-model, and a linear model
used in an application of RBPF to rover data.