Data-adaptive motion estimation algorithm and VLSI architecture design for low-power video systems

Abstract

A data-adaptive video motion estimation algorithm and its low-power VLSI implementation are presented. The proposed technique exploits the input data variations to dynamically reconfigure the search-window size of an exhaustive block-matching search thus avoiding unnecessary computations and memory accesses. Both spatial and temporal correlations of the motion vector field are taken into account. A quality analysis based on several test conditions confirms the efficiency of the motion estimator when compared with the conventional full-search as well as other low-complexity motion-estimation techniques. Realised in 0.25 μm CMOS technology, it allows for the same high performance of the full-search approach while considerably reducing power consumption from 70 up to 90% for typical QCIF and CIF video sequences.