适合硬件实现的运动估计方法及其VLSI实现

A new motion estimation algorithm based on hardware implementation and its VLSI design

根据视频编码器硬件实现的需求,给出了一种新的近似金字塔运动估计算法,通过对搜索窗中的匹配位置进行金字塔抽样,减少了每一层的搜索运算量.在此基础上,在最后一层的整像素搜索中,直接进行小块的匹配计算,去除了MPEG-4和H.264中宏块运动估计之后对8×8或4×4小块估计的冗余计算.在硬件实现中改进了数据流的输入方式,利用类全搜索估计中数据流的规则性提高了运算的并行度.位置抽样后的搜索中每一层的实现有很强的相似性,通过复用硬件结构大大节省了硬件资源.在MPEG-4 SP编码器上的测试说明,这种方法的PSNR值相对于全搜索平均只有0.02 dB的损失,但它的运算量只有全搜索的7%,占用的硬件资源仅为全搜索的1/4,并且可满足实时性的需求.

A new pyramid motion estimation algorithm and its VLSI architecture are proposed based on the demand for hardware implementation of a video eneoder. The computational load is reduced by the method of a pyramidal sample in searching positions. Redundant computation of 8 × 8 or 4 × 4 blocks is removed after macro blocks for the Sum of Absolute Difference（SAD） of sub-block are calculated directly in the last layer of integral search. During hardware implementation processing, data stream access mode is revised, and parallel process efficiency is impoved as well. After positions sampling, each layer has a similar architecture, and hareware resource is saved significantly by use of the common module. Simulation results on MPEG-4 SP show that although the new pyramid algorithm is 0.02 dB lower than full search in PSNR, the process load and hardware resource occupancy are 7% and 25%, respectively, that of full search, which will meet the requirements for real time annlications.