Object tracking using motion-adaptive modeling of scene content

We describe a system which detects and tracks objects of interest specified a priori, such as human faces and bodies, in video sequences. Face location tracking algorithms described previously by Eleftheriadis and Jacquin (see Signal Processing: Image Communication, vol.7, no.3, p.231-48, 1995 and vol.7, no.4-6, p.435-57, 1995) are extended to enable robust and accurate tracking of faces and bodies in video scenes with complex spatio-temporal backgrounds, i.e. cluttered static backgrounds as well as moving backgrounds (for example due to camera motion and/or zoom). The new tracking algorithm includes background removal obtained from global motion estimation (GME), as well as the use of combined motion and edge data and knowledge-based temporal adaptation, which jointly add significant robustness to the tracking. Face/body tracking results show the accuracy of this new system for challenging video material