基于结构纹理分解和多重网格的光流估计算法

分别针对光流计算对光照变化敏感以及运算复杂度高、迭代收敛缓慢的问题,使用一种基于ROF(Rudin-Osher-Fatemi)模型的结构纹理分解(STD)技术得到对光照变化不敏感的纹理图估计光流场,消除了光照变化产生的负面影响,并提出一种基于多重网格方法的分层处理策略.将光流计算的迭代过程分散在一系列粗细不同的网格上,在细网格上消除高频误差,在粗网格上消除低频误差,以达到加速收敛、提高光流计算速度的目的.实验结果表明,STD过程抑制了光照变化导致的负面影响,提高了光流估计精度.多重网格算法在保持优化精度的前提下,显著提高了光流计算的实时性.

An Improved Optical Flow Estimation Method Based on Structure-Texture Decomposition and Multiple Grid Method

The estimation of optical flow suffers from its sensitivity to illumination variation, its high computational complexity and slow convergence property which severely compromise its performance. In order to tackle the aforementioned problems, a structure-texture decomposition technique based on ROF(Rudin Osher Fatemi) model was introduced to deal with the variation in illumination, and a multi-grid based optical flow hierarchy strategy was presented for a fast implementation. The iterative procedure was proposed to be distributed on several grid layers with different resolutions in order to obtain a fast convergence and, in turn, accelerated the optical flow computation. The finer grid promised to eliminate higher frequency errors while the coarser level was employed to cope with lower frequency components. It is revealed from the experimental results that the structure texture decomposition is robust against the variation in illumination, and is beneficial for estimation accuracy. Additionally, the multi grid method offers an improved real time performance, without deteriorating the optimization accuracy.