James Orwell, Paolo Remagnino and Graeme Jones
Computer Vision Group, CS&ES, Kingston University, UK.
We assume the target representation of the tracking operation to be a segmentation into separate physical entities, over space and time. A background model is used to isolate the foreground pixels, to be grouped into spatially connected regions. Splitting and merging of these regions may be necessary to establish a correspondence with the observed objects; they must also be tracked over the sequence. We show how a hierarchical ordering of regions allows a simple resolution of some possible ambiguities. We then suggest that the target representation must be synchronized with discrete updates of the background model, for effective tracking in surveillance applications.