支撑点扩展快速立体匹配方法的设计与应用

为精确构建计算机立体视觉中的视差图,提出了一种快速全局优化匹配算法。该算法采用吉布斯随机场模型描述空间点与其邻域之间的关系,由改进的Graph Cuts方法对空间点的邻域进行匹配来获取场景的致密视差图。首先,计算出一组具有明确匹配关系的稀疏匹配点,将这些匹配点命名为“支撑点”;然后,对每一个支撑点的邻域进行扩展,采用改进的Graph Cuts全局优化算法计算扩展后的邻域空间的匹配关系,并将满足一定匹配度的邻域点设置为新的支撑点。最后,重复上述步骤并逐级扩展,直至扩展出的匹配空间覆盖整个视图,进而获取待匹配图对的致密视差图。实验结果表明,该方法不仅对不同场景视差图的质量具有良好的一致性,而且匹配速度较快(匹配时间约为0.8~1.2 s),大大高于其他传统的全局匹配算法。为体现本文算法的实际应用价值,以Smart Eye Ⅱ立体视觉试验台为测试平台,对真实场景进行了视差图构建,取得了良好的试验效果。

Design and application of fast matching method based on support point expansion

To construct a high qualitative disparity space image in stereo vision, a fast global optimal matching algorithm based on Gibbs Random Field(GRF) model was proposed. In this algorithm, the relationship between a space point and its neighborhood was described by using the GRF, and an improved Graph Cut method was used to calculate the matching relationship of the neighborhoods and to obtain the density disparity space image of a scene. Firstly, a set of matching points with distinct matching relationship was calculated, and named them as support points. Then, these support points were taken for the center and their neighborhood spaces were expanded. The improved Graph Cuts algorithm was used to match the expanded neighborhood spaces, and then set the neighborhood points that meet matching degree as new support points. Repeating the above steps and extending progressively, until the expansion of the neighborhood covered the entire the scene image and the density disparity map of the image pair was obtain finally. Experimental results show that this method has good speed consistency on the disparity map of the different scenes, and the matching time is about 0.8-1.2 s. For reflecting the practical value of the algorithm, the proposed algorithm was adopted to construct the disparity map of a real scene on the binocular vision test bed Smart Eye Ⅱ, and good reconstruction results were obtained.