Real-Time Rough Extraction of Foreground Objects in MPEG1,2 Compressed Video

This paper describes a new approach to extract foreground objects in MPEG1,2 video streams, in the framework of “rough indexing paradigm”, that is starting from rough data obtained by only partially decoding the compressed stream. In this approach we use both P-frame motion information and I-frame colour information to identify and extract foreground objects. The particularity of our approach with regards to the state of the art methods consists in a robust estimation of camera motion and its use for localisation of real objects and filtering of parasite zones.
Secondly, a spatio-temporal filtering of roughly segmented objects at DC resolution is fulfilled using motion trajectory and gaussian-like shape characteristic function. This paradigm results in content description in real time, maintaining a good level of details.