基于加权中值滤波的MST立体匹配算法

传统的最小生成树立体匹配算法对低纹理区域和遮挡区域不敏感，虽然最小生成树立体匹配算法后处理的中值滤波能够消除噪点，但是不能够消除边缘模糊。本文提出一种改进算法来克服这些局限性。首先，由于最小生成树匹配成本区分度不够高，研究并提出新最小生成树的匹配成本，使其可以减小不敏感区域的误匹配。其次，在后处理中使用加权中值滤波，以改善深度图像边缘。实验结果表明，在最小生成树立体匹配算法中使用改进匹配成本算法和加权中值滤波算法，在Middlebury数据集中平均误匹配率达到6.9%，本文算法在Middlebury和KITTI场景中都优于最小生成树立体匹配算法。

MST Stereo Matching Algorithm Based on Weighted Median Filtering

The traditional Minimum Spanning Tree(MST) stereo matching algorithm is insensitive to exture-less regions and occlusion areas. Moreover, the median filter of the minimum spanning tree stereo matching algorithm can eliminate the noise, but it cannot eliminate the edge blur. This paper presents an improved algorithm to overcome these limitations. Firstly, because the cost discrimination of the minimum spanning tree matching is not enough high, the matching cost of the new minimum spanning tree is studied and it can reduce the mismatching of the insensitive region. Secondly, weighted median filtering is used in post-processing to improve image boundaries. The experimental results show that the improved matching algorithm and the weighted median filtering algorithm are used in the minimum spanning tree stereo matching algorithm, and the average false matching rate is 6.9% in the Middlebury data set. The algorithm is superior to minimum spanning tree in Middlebury and KITTI.