A locally adaptive subsampling algorithm for software based motion estimation

Motion estimation is an important technology in video coding systems. Although many algorithms have been proposed to reduce its computational complexity, they are still insufficient for implementation of a software based system. The paper proposes several techniques for low-complexity motion estimation, such as a locally adaptive subsampling. Since they reduce the number of SAD (sum of absolute differences) operations, the computational complexity can be dramatically reduced while maintaining high image quality. Simulation results show that the proposed algorithm has about 20 times the speedup as TSS (three step search). This means that the proposed algorithm can attain enough processing performance for real-time implementation on embedded processors. Furthermore, PSNR of our method is about 0.2 dB superior to that of TSS.