Adaptive particle allocation for multifocal visual attention based on particle filtering

When confronting floods of visual inputs, it is usually impossible for computers to examine all possible interpretations based on given visual data. Despite these computational difficulties, humans robustly perform accurate visual processing. One of the most important keys in human visual processing is attention control. In this article, we first suggest that the particle filter (PF) is a major candidate for a model of multifocal visual attention. PF is a method which approximates intractable integrations in incremental Bayesian computation by means of stochastic sampling. One of the major drawbacks of PFs is a trade-off between computational costs and tracking performance; a large number of particles are required for accurate and robust estimations of state variables, which is time-consuming. This study proposes a computational model for multifocal visual attention which deals with the cost-performance trade-off with a restricted computing resource (the number of particles). Simulation experiments of tracking two targets with only tens of particles demonstrate the feasibility of the model.