A fully parallel particle filter architecture for FPGAs

Abstract

The particle filter is a nonparametric filter which approximates the posterior system state through a finite number of state samples i.e. particles drawn from a probability distribution. It consists of three steps which are motion update, sensor update and resampling. The first two steps are easily parallelized since the calculations do not depend on other particles. The resampling step however requires all particles to determine the particle set for the next iteration of the particle filter. In this paper, we introduce a novel FPGA optimized resampling (FO-resampling) approach to solve the parallelization problem of the resampling step by introducing virtual particles. Compared to multinomial resampling, FO-resampling achieves similar results with the added benefit of being able to completely parallelize all the steps of the particle filter. Additional to evaluating our approach with simulations, we implement a particle filter with FO-resampling on an FPGA.