Learning a Multi-camera Topology

We report an investigation to determine the topology of an arbitrary network of video cameras observing an environment. The topology is learnt in an unsupervised manner by temporal correlation of objects transiting between adjacent camera viewfields. We extract this information in two steps, firstly identifying the principal entry and exit zones associated with each camera view, and then establishing the correspondence or links between the exit zones of one camera to the entry zones of an adjacent one by accumulating evidence from many trajectory observations. A significant benefit of the method is that it doesn’t rely on establishing correspondence between trajectories. In addition to generating the topological structure, the method also results in a measure of inter-camera transition times, which can be used to support predictive tracking across the camera network.