In this paper we consider the problem of robust visual tracking of multiple targets using several, not necessarily registered, cameras. The key idea is to exploit the high spatial and temporal correlations between frames and across views by (i) associating to each viewpoint a set of intrinsic coordinates on a low dimensional manifold, and (ii) finding an operator that maps the dynamic evolution of points over manifolds corresponding to different viewpoints. Once this operator has been

more »

... , correspondences are found by simply running a sequence of frames observed from one view through the operator to predict the corresponding current frame in the other view. As we show in the paper, this approach substantially increases robustness not only against occlusion and clutter, but also against appearance changes. In addition, it provides a scalable mechanism for sensors to share information under bandwidth constraints. These results are illustrated with several examples.