Speeding up motion estimation algorithms on CUDA technology

Motion estimation(ME) is one of the most important modules in digital video encoding/decoding and video post processing. High-quality and fast ME algorithms are desired by many applications. In this paper, motion estimation algorithms are explored in terms of execution efficiency on the CUDA (Compute Unified Device Architecture) technology which is a parallel computing architecture developed by nVIDIA. In this paper, we implemented the full search algorithm (FS), the diamond search algorithm, and the four step algorithms in both the CPU and the CUDA platforms. Experiment results show that the CUDA-based implementations of these ME algorithms can be more than 8 times as fast as the those CPU-based ME implementations.