Comparison of few particle filter varieties

Particle filters are very popular - number of algorithms based on Sequential Monte Carlo methods is growing. Paper describes and compares the performance of four of them: Auxiliary Particle Filter - this approach should reduce the high sensitivity to outliers values and poor posterior approximation, Rao-Blackwellised Particle Filter - this approach is recommended for objects with linear and nonlinear state variables, Bootstrap Filter - the first proposed Particle Filter which still can be used, because is very simple to implement, and some variety of SIR algorithm - this algorithm was chosen to show, that importance density also can be constant. The obtained results show that Bootstrap Filter and Rao-Blackwellised approaches give good results, but Bootstrap Filter works 10 times faster. The worst results gives SIR algorithm with unconditional importance function.