Lattice Particle Filters

Abstract

A promising approach to approximate inference in state-space models is particle filtering. However, the performance of particle filters often varies significantly due to their stochastic nature. We present a class of algorithms, called lattice particle filters, that circumvent this difficulty by placing the particles deterministically according to a Quasi-Monte Carlo integration rule. We describe a practical realization of this idea, discuss its theoretical properties, and its efficiency. Ex~ perimental results with a synthetic 2D tracking problem show that the lattice particle filter is equivalent to a conventional particle filter that has between 10 and 60% more particles, depending on their "sparsity" in the state-space. We also present results on inferring 3D human motion from moving light displays.