Architectural design issues for Bayesian contextual vision

Sensor fusion technology has been so far developed for the fusion of camera and different sensors, e.g. radar, sonar, etc. The same techniques apply to integrating several vision algorithms into a multi-modular system. In this paper, we abstract our attempt on the matter and propose a uniform paradigm to integrate both "vision modules" directly observing targets (e.g. intruders in video surveillance) and "accessory modules" observing scene features that may trigger system adaptation to the current context. To be concrete, we completely develop a real example in re-designing a previous context-dependent video surveillance system.