邻域内坐标线性变换的光流迭代求解算法

New method to solve generalized optical flow with linear transformation

提出邻域内坐标线性变换的光流计算迭代解法。算法允许计算光流的图像间存在一定的灰度偏差,光流计算结果稳定。对于含较大位移的两幅图像,以往方法要用金字塔图像方法计算光流,而该算法则可以一步求解得到图像间最佳整体匹配,比金字塔方法更有优势。算法对图像整体的尺度、平移、旋转或其它线性变换能有效检测。用于人脸图像配准的实验显示,该算法对图像匹配空间搜索有效可行。

An iterative algorithm is proposed to solve optical flow problem,which defines a linear transformation of the pixels in the neighbourhood instead of translation.While the pixel intensity of the same object in the two frames is not equal,the algorithm can cope with it in a better way than other methods.It can tackle the problem of large translation flow by itself,rather than image pyramid method.The experiments show it can also be used for face registration.And it converges quickly in a certain variation of rotation,translation,scale or other linear transformation.