A fast block motion estimation algorithm using dynamic pattern search

Abstract

In block-based motion estimation algorithms, it has always been desired to reduce search point computation with quality as good as full-search algorithm. A number of such algorithms like diamond search (DS) and hexagon search (HS) have been proposed in literature, which use fixed-size search patterns for finding motion vectors. The drawback with these fixed-size search pattern–based algorithms is that they may suffer from oversearch/undersearch problem depending on the magnitude of the motion vector. In this manuscript, a dynamic pattern search–based algorithm (DPS), which uses spatial and temporal coherence among blocks and dynamically adapts its search pattern for a candidate block, has been proposed. The proposed algorithm has been compared with various motion estimation algorithms like DS, HS, adaptive rood pattern search (ARPS) and full search in terms of various performance parameters. Experimental results show that proposed DPS has a speed gain of 1.18 over ARPS, whereas it is nearly 1.94 and 1.33 over DS and HS algorithms in terms of average search points/block. Further, in terms of peak signal-to-noise ratio (PSNR) (dB)/frame, DPS produces almost same average value than ARPS and HS, whereas it is only 1% inferior to DS. A modified version of DPS has also been proposed, which increases its speed gain by 1.39 times with negligible decrease in PSNR. In terms of another time parameter—average execution time per frame (s)—for DPS, it is 0.66 s, whereas this time is 0.71, 0.77 and 1.06 for ARPS, HS and DS algorithms, respectively.