An adaptive particle filter based on the mixing probability

In the stochastic system whose state is described by the multiple model particle filter, the true dynamic system motion cannot be reflected precisely by the predicted measurement from each model in time because of random sampling. It causes the probability of each model inaccuracy and falls down the performance of estimation. In the interacting multiple model particle filter (IMMPF) algorithm, the dominant model should be paid more attention in order to close the posterior. So it should have much more sampling particles. Meanwhile, it is not necessary to utilize too many particles in other models which perform weak. Considered the above problem, an improved method for the IMMPF with an adaptive particle number strategy is proposed. The sampling number in each sub-filter of the IMMPF algorithm is adaptively changed, according to the value of the mixing probability. When the mixing probability exceeds the designed threshold, which is about 5∼8 times of the initial mode transition probability, an appropriate strategy is designed by making a decision of the dominant model in the mode set. Then, the sampling number is increased in the dominant model and decreased in non-dominant models respectively. The simulation result shows that this method has a prior performance than the general IMMPF with a fixed particle number and a similar computational cost.