Attentional control for visual surveillance

The paper introduces a general framework for attentional control related to current work and looks at the HIVIS-WATCHER implementation. This system uses a separation of simple and complex operators that are formally defined within a deictic frame of reference and coordinated by the "official-observer". The purpose of the separation is to avoid complex attentive processing except for objects selected as likely to be interacting in a task-related way by preattentive cues, e.g. mutual-proximity. The deictic representation allows typical behaviour to be modelled from a selected agent's point of view. The behavioral interaction used for illustration here is "overtaking" which involves aspects of the heading, speed and relative position. A tasknet implemented as a static Bayesian belief network integrates this evidence to infer the likely episode as it evolves in the dynamic scene. Tables for the combination of deictic viewpoints and results from looking for likely overtaking behaviour are presented. Conclusions and future work centre on the need for full integration of early visual processing and object recognition with the high-level behavioural interpretation system.