A review of resampling techniques in particle filtering framework

A particle filtering (PF) is a sequential Bayesian filtering method suitable for non-linear non-Gaussian systems, which is widely used to estimate the states of parameters of interest that cannot be obtained directly but still relate to noisy measured data with probability masses. Possible values of targeted parameters (or particles) are sampled according to the related prior knowledge, with their probabilities (or weights) evaluated from the likelihood of being the true values of those parameters. However, most have negligible weights. The standard PF algorithm consists of three steps as particle generation, weight calculation or updating and particle regeneration, which is called resampling. The performance of PF depends greatly on the quality of particle regeneration. Resampling preserves and replicates particles with high weights, while those with low weights are eliminated. However, particle impoverishment is a side effect that reduces the diversity of particles used in the next time steps. Therefore, efficient resampling have to guarantee high likelihoods particles. This paper reviews the classification and qualitative descriptions of recent efficient particle weight-based resampling schemes and discusses their characteristics, implementations, advantages and disadvantages of each scheme.