Fast and Accurate Half Pixel Motion Estimation Using the Property of Motion Vector

Abstract

To estimate an accuracy motion vector (MV), a two step search is generally used. In the first step, integer pixel points within a search area are examined to find the integer pixel MV. Then, in the second step, 8 half pixel points around the selected integer pixel MV are examined and the best matching point is chosen as the final MV. Many fast integer pixel motion estimation (ME) algorithms can be found by examining less than about 10 search points. However, the half pixel ME requires huge computational complexity. In this paper, We propose a new fast algorithm for half pixel ME that reduces the computational overhead by limiting the number of interpolations of the candidate half pixel points. The proposed method based on the property of MVs and the correlations between integer pixel MVs and half pixel MVs. Experimental results show that the speedup improvement of the proposed algorithm over a full half pixel search (FHPS), horizontal and vertical directions as references (HVDR), chen's half pixel search algorithm (CHPS-1) and a parabolic prediction-based fast half-pixel search (PPHPS) can be up to 1.4 - 3.9 times on average. Also the image quality improvement can be better up to 0.05(dB) - 0.1(dB) compare with CHPS and PPHPS. © Springer-Verlag 004.