Unsupervised Learning Grouping-Based Resampling for Particle Filters

Conventional resampling for particle filters suffers from discarding much potentially useful information due to using less information of spatial distribution of sampling particles set. An unsupervised learning grouping based resampling for particles filter, which could further exploits the relevant spatial information hidden in sampling particles set, was proposed. The core of our approach is that the k-means clustering algorithm is used to automatically form clusters of similar particles, instead of simple heuristic partitioning of the particles set in the spatial domain. Based on the time series analysis between the particle path and observation path for all particles in each cluster, the particle called important particle that has highest temporal correlation is chosen. The diversity of resampling particles set is strengthened by adding important particles to improve the residual resampling. Additionally, by the Kolmogorov-Smirnov two-sample test, we proof that the particles set resampled by the proposed algorithm and the original particles set come from the same distribution function. Extensive experimental results on two numerical models show that the estimation accuracy of particle filter can be improved in this way.