Efficient and fast predictive block motion estimation for low bit rate video coding

Abstract

An efficient block based motion estimation algorithm is proposed and evaluated for performance improvement. Spatio-temporal correlation characteristics and histograms of motion vector magnitude and direction are investigated through computer simulation with test video sequences. Its initial search pixel location is predicted spatially from the neighboring macro block in the current picture frame. In order to utilize center oriented motion vector characteristics from the histogram analysis, fast estimation algorithms such as three step search, diamond search, 2D logarithmic search and 1-D gradient search are evaluated and compared in terms of its search performance with respect to Euclidean search distance. A new predictive 1-D gradient search algorithm is proposed which shows the best results for motion video having moderate motion inside and is applicable for low bit rate video coding such as video telephony. Computer simulation using test video sequences in the H.263 framework shows that the proposed algorithm reduces computation cost measured in the number of SAD calculations. Its compression quality represented in terms of PSNR is improved further than conventional methods by controlling the search termination condition. It is proved that the proposed algorithm improves the search speed as well as PSNR performance of 1-D gradient motion estimation and outperformed other fast algorithms with respect to search speed. Furthermore, its search performance can be traded off with computation cost according to application requirements.