Split-Gaussian particle filter

This paper is concerned with the use of split-Gaussian importance distributions in sequential importance resampling based particle filtering. We present novel particle filtering algorithms using the split-Gaussian importance distributions and compare their performance with several alternatives. Using a univariate nonlinear reference model, we compare the performance off the importance distributions by monitoring the effective number of particles. When using adaptive resampling, the split-Gaussian approximation has the best performance, and the Laplace approximation performs better than importance distributions based on unscented and extended Kalman filters. In addition, we also consider a two-dimensional target-tracking example where the Laplace approximation is not available in closed form and propose fitting the split-Gaussian importance distribution starting from an unscented Kalman filter based approximation.