A Low Complexity Motion Segmentation Based On Semantic Representation Of Encoded Video Streams

Video streaming is characterized by a deep heterogeneity clue to the availability of many different video standards such as H.262, H.263, MPEG-4/H.264, H.261 and others. In this situation two approaches to motion segmentation are possible: the first needs to decode each stream before processing it, with a high computational complexity, while the second is based on video processing in the coded domain, with the disadvantage of coupling between implementation and the coded stream. In this paper a motion segmentation based on a "generic encoded video model" is proposed. It aims at building applications in the encoded domain independently by target codec. This can be done by a video stream representation based on a semantic abstraction of the video syntax. Tins model joins the advantages of the two previous approaches by making it possible working in real time, with low complexity, and with small latency. The effectiveness of the proposed representation is evaluated on a low complexity video segmentation of moving objects.