Techniques for parallel execution of the particle filter

Abstract

Dynamic Bayesian networks are a promising approach to automatic video content analysis which allows statistical inference and learning to be combined with domain knowledge. When the particle filter (PF) is used for approximate inference, video data can often be classified in real-time (supporting e.g. on-line automatic video surveillance). Unfortunately, the limited processing resources available on a typical host restricts the complexity, accuracy, and frame rate of PF classification tasks. Here, we target this limitation by applying the traditional parallel pooled classifiers architecture to execute multiple PFs in parallel and to coordinate their output. We then identify a significant weakness of this approach in terms of loss of accuracy. To reduce the loss of accuracy, we propose a novel scheme for coordinating the pooled PFs based on the exchange of so-called particles. In an object tracking experiment, a significant loss of accuracy is observed for the naive application of the pooled classifiers architecture. No loss of accuracy is detected when our scheme for exchanging particles is used. © Springer-Verlag 2003.