Adaptive multi-resolution motion estimation using texture-based search strategies

Motion estimation is the most complex module which contributes nearly 70% of computation resources in a hardware-based video encoder. This huge computational complexity limits the performance of HD video encoders in terms of encoding speed and power consumption. This paper presents a hardware oriented multi-resolution motion estimation algorithm using adaptive search strategies to reduce computational complexity. The spatial homogeneity and the temporal stationarity characteristics of video sequences are adaptively detected to determine search range and down-sampling rate. Homogeneous regions are detected by using Sobel edge operators and stationary regions are detected by using temporal information. These texture-based search strategies make motion estimation more concise under fixed computational complexity constraint. Additional computational cost originated by determining the search strategies can be neglected due to the simple addition and shift operations. Experimental results show that the proposed algorithm achieves better performance and reduces computation cost by 40% compared with previous works.