Adaptive-size hierarchical block matching for efficient motion compensation of video sequences

The paper addresses the problem of reducing the computational load in block-based motion estimation (BBME) by exploiting a preliminary motion field, which is obtained from the coarsest level of a multiresolution pyramid, and its successive refinements in the finer ones. The first estimation is achieved by a full search, even on greatly reduced scales, whereas at the other levels just a small correction is computed, thereby achieving a speed increase. However, this results also in difficulty to improve an inaccurate initial estimation. The algorithm proposed in the paper overcomes such a drawback and even outperforms BBME in terms of rate vs. distortion through the computing of a variable-resolution motion field which allows the same reconstruction quality with fewer vectors. A good correction of wrong initial estimates is achieved by use of an appropriate propagation of displacement vectors from one level to the next one and of an adaptive search range. The vectors propagation is carried out by choosing the best vector among a set of neighbors taken from the previous level and among the already refined ones in the current level. The search range is chosen on the basis of the error in reconstructing the block related to the propagated vector. Finally, the adaptive resolution of the field is achieved by ending the propagation at an intermediate level of resolution, if a vector allows a good reconstruction quality of its block.