On the binding mechanism of synchronised visual events

We address the problem of interpreting visual surveillance data by learning appropriate spatio-temporal subspaces of active image regions caused by scene activities. We focus on identifying regions of sustained change for recognising key stages of a visual behaviour. Our behaviour representation is based on the asynchrony or delay patterns of occurrence among local events which need not be spatially connected. We use an automatic normalised cut structure discovery algorithm with a hybrid similarity criterion for simultaneously identifying relevant spatio-temporal subspaces and clustering similar behaviour patterns in those subspaces. We compare the automatically discovered classes with conceptual classes of behaviours in a semi-constrained "shopping" scenario.