基于色彩权值和树形动态规划的立体匹配算

Stereo MatchingAlgorithm Based on Color Weights and Tree Dynamic Programming

针对立体匹配算法在图像非遮挡区域,特别是弱纹理区域误匹配率较高的问题,提出一种基于十字交叉窗口下自适应色彩权值和树形动态规划的立体匹配算法。首先结合颜色、梯度信息及Census变换作为相似性测度函数构建代价计算函数;然后以图像的距离和色彩信息构建自适应十字交叉窗口,并提出基于色彩权值的代价聚集方式;将树形结构动态规划算法的思想引入到视差计算,代替单独采用赢者通吃策略的方法,对视差进行全局优化;最后通过视差求精得到稠密视差图。实验结果表明,本文算法在Middlebury测试平台4幅标准图像上非遮挡区域的平均误匹配率为2.45%,同时对其他10组图像进行了对比评估,本文算法有效地提高了图像非遮挡区域匹配的准确率。

Aiming at the problem that the stereo matching algorithms have high mismatching rates in non-occluded regions of the images, especially the weak texture regions, a stereo matching algorithm based on adaptive color weights over cross window and tree dynamic programming is proposed. Firstly, we combine color, gradient information and census transform as similarity measure function to propose the cost calculation function. Then the adaptive cross window is constructed with distance and color information of the image, and the cost aggregation based on color weights is proposed. Instead of using winner-take-all strategy solely for global optimization of disparity, the dynamic programming algorithm based on tree structure is introduced to calculate disparity. Finally, the dense disparity maps are obtained by the process of disparity refinement. The experimental results demonstrate that on the Middlebury test platform, the average mismatching rate evaluated with proposed algorithm in non- occluded regions of four standard images is 2.45 %. Meanwhile, the other ten images are compared and evaluated. The proposed algorithm effectively improves the accuracy of stereo matching in non-occluded regions.