A highly parallel motion estimation method based on temporal motion vector prediction for a many-core platform

In hybrid video coding such as H.264/AVC and H.265/HEVC using motion compensation, most coding processes are mainly used for motion estimation. Recently, highly parallel processing devices such as graphics processing units (GPUs) or many-core processors have been utilized to accelerate motion estimation. Although a straightforward way to parallelize motion estimation is block-based parallelization within a frame, motion information of the neighboring block is not available so coding efficiency loss is inevitable. A method using motion vectors of coded frames has been proposed to tackle this problem; however, it causes a decrease in the precision of motion vector predictions in a hierarchical reference structure. This paper proposes a temporal motion vector prediction-based, highly parallel motion estimation method that is applicable to a hierarchical reference structure. It utilizes motion vectors of non-encoded frames referring to the same reference frame of an encoding frame and a block size decision and a concatenation of motion vectors to improve the estimation accuracy. Experiments show that the proposed method achieves up to 9.2% rate-distortion improvement over the conventional method with the similar encoding speed improvement over HM.