Particle Filter Resampling Using Sequential Draining

The particle filter is known to be a powerful tool used for the estimation of latent dynamic target states from noisy measurements. The filter approximates the probability distribution of the target state using a set of weighted particles, called samples. The use of such a scheme makes the filter extremely suitable for target estimation within a general class of nonlinear and nonGaussian state space models. The filter involves a resampling step that overcomes degeneracy, a case when one particle grabs all the weight. Stochastic resampling methods like the systematic and stratified resamplers involve the use of random variables within a sequential search mechanism to determine samples that should be retained. These methods are known to be theoretically stable in terms of giving an unbiased estimate of the probability density. However the stochasticity involved in these resamplers makes the inter-particle communication volatile and unknown ahead in time for each time epoch, thus making it cumbersome for implementation at the processor level. Other resampling methods including the deterministic, partial deterministic and hardware architecture friendly resamplers aim at reducing the communication overhead involved within the operation, but suffer in estimation bias and/or poor tracking accuracy. This paper presents a fully deterministic resampling scheme that gives an unbiassed estimate of the probability density, and hence being hardware friendly.