Abstract: As traditional stereo matching algorithms cannot provide an appropriate size aggregate window for image edges and low texture regions at the same time, an improved stereo matching algorithm is proposed, which is based on Gaussian mixture model and minimum spanning tree structure. The image is divided into the several initial regions and the candidate pixels to be segmented as the first step, which is obtained by the pixel color and the distance information, together with the initial disparity. The Gaussian mixture model, which is run in parallel, is secondly leveraged to get the final segmentation by updating the parameters of each region iteratively. Then, a minimum spanning tree is built on each segment to obtain the disparity. Finally, a high precision dense disparity map is obtained by correcting the mismatch of valid disparity from the neighbors. Compared to other algorithms proposed in the literature, this algorithm provides substantial precision improvement, especially in the depth discontinuity region.

Key words: image processing, stereo matching, Gaussian mixture model, minimum spanning tree, disparity refinement

摘要： 针对传统立体匹配算法无法同时为图像边缘和低纹理区域提供一个合适大小的聚合窗口而导致匹配精度较低的难题，提出一种结合高斯混合模型及最小生成树结构的立体匹配算法。通过图像初始视差、像素颜色及距离信息将图像分为初始若干区域及待分割候选像素；基于高斯混合模型并行迭代更新各区域参数，得到最终的分割；在各分割上建立最小生成树计算聚合值求取视差；通过邻域内的有效视差修正误匹配点，获取精度较高的稠密视差图。与其他算法相比，该算法能有效降低误匹配率，尤其在深度不连续区域的匹配效果显著改善。

关键词: 图像处理, 立体匹配, 高斯混合模型, 最小生成树, 视差求精