中文摘要：

在传统的局部立体匹配算法中,代价聚合要对相关邻域内的点进行加权聚合,这种方式计算量大,非常耗时。文章提出一种基于最小生成树的立体匹配算法,该方法将图论中的最小生成树引入代价聚合和视差细化中,使得图像中所有点都对兴趣点进行聚合支持,弥补了局部算法在弱纹理区误匹配率高的局限性,提高了匹配的准确性,并且最小生成树能够对图像所有点进行层次性的划分,极大地简化了计算量。实验证明,该算法能够快速得到平滑且精度高的视差图。

英文摘要：

In the traditional local stereo matching algorithms, cost aggregation needs to accomplish the weighted point aggregation in the correlated neighborhood model, which causes large computation and extraordinary time consumption. In this paper, a new stereo matching algorithm based on the minimal spanning tree theory is proposed which is applied to the cost aggregation and disparity refinement. In this method, interesting points can get the aggregation support from all points in the image, whieh o- vercomes the deficiency of high false matching rate in weak texture region in local stereo matching al- gorithm, and improves the matching accuracy. At the same time, all points in the image can be divid- ed into different levels through the minimal spanning tree, thus reducing the computation time abun- dantly. The experimental results show that, through the algorithm proposed, smooth disparity maps with high accuracy can be achieved rapidly.