为改善传统立体匹配视差图中目标边缘的毛刺现象,以及弱纹理和视差不连续区域的“阶梯效应”等,提出了一种结合局部二进制表示和超像素分割的立体匹配方法。首先融合二进制表示的窗口内像素的空间和颜色特征进行代价计算,并以此求得初始视差;然后将简单线性迭代聚类方法分割的结果作为像素的空间和颜色标记,为超像素内的目标边缘和其他像素点选择恰当的稳定点进行视差传播,以达到视差优化时边缘保持和空间平滑的目的。在Middlebury数据集上分别进行代价计算与优化方法的对比实验,结果表明,采用该算法获取的目标边缘的视差更为平滑,在左右视图中的遮挡区和不重叠区域获得的视差也比较准确,有效地降低了非遮挡区、全图、不连续区域的误匹配率。

In order to polish up the target edge burring and staircase effect in low-textured regions or discontinuous regions, a stereo matching method based on the local binary description and superpixel segmentation is proposed. Firstly, the initial disparity is obtained by space and color features binary cost computation and winner-takes-all method. Then the segmentation results by simple linear iterative clustering method are labeled for each pixel's space and color features. In disparity refinement procedure, the appropriate fixed points are chosen to propagate disparity for both edge and inner pixels of each superpixel. Experiments with Middlebury datasets are mainly carried out in the initial disparity considerations and disparity refinement. The result shows that the disparity maps are much smoother especially in target boundary. The proposed method can achieve more accurate disparity value in non-overlapping and occluded regions between reference image and matching image, which effectively reduces the mismatching rate in non-occluded, all, and discontinuity regions.