Real-time Particle Filters Using Mixtures of Samples Sets

Recently, particle filters have been applied with great success to a variety of state estimation problems. In many real time applications, sensor information arrives significantly faster than the particle filter can process. The prevalent approach to this problem is to update the particle filter as often as possible and to discard sensor information that cannot be processed in time. In this paper we present an alternative approach. Our real time particle filter makes use of all sensor information by partitioning sample sets into mixtures of smaller sample sets. Each small sample set contains as many samples as can be processed between the arrival of two observations. These small sets are combined as soon as the total number of samples in the mixture is sufficiently high. Using efficient subsampling, we determine the mixture weights so as to minimize the KL-divergence between the mixture density and the true posterior. Thereby, our approach focuses computational resources on valuable sensor information. Experiments using data collected with a mobile robot show that our approach yields drastic improvements over alternative techniques.