A novel parallelized motion estimation algorithm for GPU based video encoding

Abstract

Cloud-based video encoding has become more and more popular in Internet, especially for mobile clients, considering their limited resources. Recently, GPUs (Graphics Processor Units) make the cloud-based video encoding more economic and efficient. However, the motion estimation in inter prediction, which usually occupies about 70% encoding time in H.264/AVC, is still a big headache because of its complexity. In this paper, a novel motion estimation algorithm is proposed, which is customized for GPU-based cloud encoding, considering motion tendency. A mean subsampling template is presented for a pre-search approach to get motion tendency, which can reduce computation cost obviously with less quality loss. To improve the efficiency of the CUDA (Compute Unified Device Architecture) thread organization for motion estimation, a section-division method is presented. Experimental results show that the proposed algorithm can reduce nearly 22% computation time with less video quality loss, compared with the state-of-the-art work.