垂直视差的存在是影响立体视频观视舒适度的主要因素。为了在不影响水平视差的条件下实现对垂直视差的消减, 本文引入Levenberg-Marquardt(L-M)非线性算法实现变换矩阵的精确求解。首先用抗缩放、旋转及仿射变换的SIFT(Scale-invariant feature transform)特征匹配算法检测出双目图像对的特征匹配点, 然后根据匹配点的坐标位置运用L-M算法计算可消减垂直视差的变换矩阵, 将变换矩阵作用于目标图像, 计算出该视图每个像素点的新坐标位置。实验结果表明: 与利用线性算法求解二维射影变换矩阵的垂直视差消减方法相比, 本文提出的求解方法在垂直视差消减上比该算法提高了约0029 1～0323 2个像素, 对水平视差的影响比该算法降低了约0118 7～1139 1个像素。因此本文提出的方法对垂直视差的消减起到了优化作用。

The existence of vertical parallax is the main factor of affecting the viewing comfort of stereo video. In order to reduce the vertical parallax without affecting the horizontal parallax, Levenberg-Marquardt(L-M) algorithm which is the nonlinear algorithm, is introduced in this paper to achieve the accuracy of the transformation matrix. Firstly, the SIFT algorithm, which is invariant to scaling, rotation and affine transformation, is used to detect the feature matching points from the binocular images. Then according to the coordinate position of matching points, the transformation matrix, which can reduce the vertical parallax, is calculated using Levenberg-Marquardt algorithm. Finally, the transformation matrix is applied to target image to calculate the new coordinate position of each pixel from the view images. The experimental results show that compared with the method that can reduce the vertical parallax using linear algorithm to calculate two-dimensional projective transformation, the proposed method using nonlinear algorithm improves the vertical parallax reduction from about 0029 1 to 0323 2 pixel and the effect of horizontal parallax is reduced from about 0118 7 to 1139 1 pixel. Therefore, the proposed method can optimize the vertical parallax reduction.