Pyramidal motion estimation techniques exploiting intra-level motion correlation

The conventional pyramidal motion estimation technique, although capable of reducing the heavy computational load as required by the full search block matching algorithm, has the problem of propagating false motion vectors. In this work, pyramidal motion estimation techniques that exploit the intra-level motion correlation to overcome this serious drawback are presented. Instead of scaling the motion vectors from the corresponding positions at the adjacent lower pyramidal level as the prediction motion vectors for the current pyramidal level as is done in the conventional technique, the prediction motion vectors are generated through either median filtering or linear prediction of a set of neighboring motion vectors. Various pyramidal data structures are employed to test the proposed techniques. Simulation results show that the proposed techniques not only improve the prediction performance, but also result in a more consistent motion field. It is further shown that this improvement in the performance is achieved with negligible extra computational load.