Dynamic programming is one of the oldest but still popular methods for stereo correspondence. Because the traditional stereo matching based on the dynamic programming has the well-known horizontal "streaking" effect for skiping the vertical edges,a stereo correspondence using the multi-stage dynamic programming is proposed to eliminate this effect in this paper. Firstly, the initial disparity space is constructed, and the down and up dynamic programming processes are operated across the epipolar lines. Then, the initial disparity space is optimized by the combined results above. Finally, based on the new disparity space, the forward and backward dynamic programming processes are implemented along the epipolar lines, and the disparities are obtained by minimizing the combined results of the dynamic programming along and across the epipolar lines. This algorithm is evaluated on the benchmark Middlebury database. The experimental results show that the "streaking" effect is suppressed and the general error matching rates have decreased by 28.60% and 40.42% respectively compared with that of the traditional dynamic programming algorithm and the scan-line optimization algorithm, which can offer a good trade off in terms of the accuracy. It is demonstrated that the proposed algorithm is more efficient.%动态规划算法是立体匹配中比较流行的一种全局优化方法,然而传统的动态规划立体匹配算法因忽略了核线间像素对视差的影响,产生了比较明显的横向"条纹"效应.为了消除此效应,本文提出了一种采用多级动态规划的立体匹配算法.该算法通过建立初始视差空间,对核线间进行双向的动态规划运算,然后将合并得到的结果用于优化初始视差空间.基于新的视差空间在核线上进行双向的动态规划运算,最后求取使核线上和核线间动态规划合并结果最小的值作为所求视差.利用标准数据库Middlebury平台对提出的算法进行了测试.实验结果表明,该算法与传统的动态规划算法和基于扫描线优化算法相比,横向"条纹"效应有所改善,且总体误匹配率分别降低了28.60%和40.42%,提高了匹配结果的准确性.
展开▼
