Probabilistic Integration of Cues From Multiple Cameras

Cue integration from multiple cameras is an important aspect for machine vision systems operating in complex, natural environments. One successful approach for self–organized cue integration is Democratic Integration. The hallmark of Democratic Integration is that different cues can autonomously determine whether and in how far they are useful for the current task, giving the system flexibilty to engage in different tasks and robustness in the face of sudden failures of cues. In this paper we embed Democratic Integration in a probabilistic framework and extend it hierachically in order to model adaptive cue integration for the general case of n calibrated cameras. Our experiments show that the method is capable of robust cue integration and adaptation during object tracking using three cameras placed arbitrarily in the scene.