基于可移动窗口和扩散距离的立体匹配算法

Stereo Matching Algorithm Based on Shiftable Window and Diffusion Distance

自适应权重立体匹配算法需要评估像素对之间的相似度,不同于传统方法利用色彩差异、距离远近进行度量,引入"扩散距离"这种新的度量方式,能够顾及像素在特征空间中的全局分布。为了减少计算的复杂性和提高鲁棒性,代价聚合分为两个阶段:(1)针对每一个像素,从预先定义的9个窗口中选择最佳的;(2)如果像素对应的最佳窗口均方差低于某一阈值,则对其采用自适应权重的方法进行处理,其余像素则利用最佳窗口来计算。代价计算使用颜色和梯度信息,并通过加权中值滤波进行视差优化。运行结果在Middlebury网站上评估,测试集上平均误匹配像素百分比为5.63,低于传统自适应权重算法及其相关改进方法。

Adaptive support-weight stereo matching algorithm requires assessing the degree of similarity between pairs of pixels. Different from the traditional algorithm based on adaptive support weight which is measured by color difference and distance, a new measure method based on diffusion distance was introduced which can account for the global distribution of pixels in their feature space. In order to reduce the complexity of computing and improve the robustness, the cost aggregation was divided into two phases. Firstly, the best window was selected from nine predefined windows for each pixel. Secondly, the pixel was processed by means of adaptive support weight if the mean square error of the corresponding optimal window was below a certain threshold. And otherwise it was calculated through the best window. The matching cost computation used color and gradient information. The weighted median filtering was applied to disparity refinement. The average percent of bad pixels of the algorithm on Middlebury stereo dataset is 5.63, which is below to that of the traditional algorithm of adaptive support window and improved ones.