Interpreting a dynamic and uncertain world: task-based control

摘要

In this paper we show that it can be beneficial to have a high-level vision component that guides the reasoning of the whole vision system when interpreting a dynamic and uncertain world. This guidance is provided by an attentional mechanism that exploits knowledge of the specific problem being solved. Here we develop a general framework for such an attentional mechanism and its application to understanding dynamic scenes. This attentional mechanism can enable a vision system to perform a given domain task while expending minimal resources. We have developed a component that uses Bayesian networks combined with a deictic representation to select what, when and how to use processed data from a fixed camera. We apply two forms of Bayesian network, which (1) create a dynamic structure to reflect the spatial organisation of the data and (2) measure task relatedness. Together these give attentional focus making the reasoning performed relevant to the task.