Real-time reasoning: the case of surveillance systems

The paper presents a layered architecture for real-time surveillance systems. Each layer includes objects that model the "real world" at a specific abstraction level, from raw data up to domain concepts. Each layer performs abstractions on perceptions coming from the lower layer and formulates timed hypotheses about domain objects. The failure of a hypothesis causes a perception to flow up-stream. In turn, hypotheses flow down-stream, so that their verification is delegated to the lower layers. The proposed architectural patterns have been reified in a Java framework, which has been used in an experimental multi-camera tracking system.