Fast particle flow particle filters via clustering

Particle flow filters, introduced in a series of papers by Daum and Huang, are an attractive alternative to particle filters for filtering tasks in high-dimensional spaces or with very informative measurements. Many variants of particle flow filters have been developed, but all require approximations in multiple stages of the implementation, which leads to particles deviating from the true posterior distribution. To preserve the statistical consistency of the filtering algorithm, some recent papers embed the particle flow techniques within a particle filter, using them to generate a proposal distribution. In recent work, we developed such a particle flow particle filter, modifying the flow mechanism to ensure that the implemented, approximate flow was an invertible mapping. This property allows efficient computation of the importance weights. In this paper, we strive to reduce the computational overhead of the particle flow particle filter by incorporating clustering of the particles. Results from a multi-target acoustic tracking simulation demonstrate that we can significantly reduce the computational cost of particle flow particle filters with a relative small sacrifice in tracking accuracy.