Probabilistic learning of salient patterns across spatially separated, uncalibrated views

We present a solution to the problem of tracking intermittent targets that can overcome long-term occlusions as well as movement between camera views. Unlike other approaches, our system does not require topological knowledge of the site or labelled training patterns during the learning period. The approach uses the statistical consistency of data obtained automatically over an extended period of time rather than explicit geometric calibration to automatically learn the salient reappearance periods for objects. This allows us to predict where objects may reappear and within how long. We demonstrate how these salient reappearance periods can be used with a model of physical appearance to track objects between spatially separate regions in single and separated views.