Motion-compensating prediction within spatiotemporal band-pass outputs based on a feature specific vector representation

摘要

A new approach for motion estimation using a local/momentary signal representation within a spatiotemporal multichannel coding scheme is presented. The coding scheme is based on a decomposition of the image sequence into three-dimensional even- and odd-symmetric band-pass filters. Each filter is frequency-, orientation-, and motion-selective and consists of a quadrature filter pair. Their outputs are interpreted as a complex, analytic signal which is transformed into a polar representation. The subsequent interband vector representation allows an exact description of local/momentary signal features including motion direction and velocity. These properties can be used to estimate the position of a certain local feature some frames later and to integrate a motion-compensating prediction algorithm within the three-dimensional subband decomposition. It is shown that efficient temporal decorrelation can be achieved with low coding delay. In addition, no frame-to-frame displacement measurements are necessary and no motion vectors have to be transmitted.