Abstract: Method for obtaining homography matrix precisely and eliminating false matching from initial matching points has been one of the hottest research topics as well as key problems in the field of computer vision. By applying the concept of similarity between matching pairs of local features to the sampling procedure of LMeds, a new adaptive method for estimating homography matrix is proposed. Different to the un-ordered random sampling procedure of the traditional LMeds method, it first sorts the similarity of initially matching pairs with a descending order, and then samples them in order. Experimental results demonstrate that this method shows improvements in precision, robustness and efficiency (iterating times needed to obtain the most optimized model is about 1/5 of that of LMeds), it can also compensate the shortage of setting a threshold on distance deviation in RANSAC and its related methods.

Key words: homography matrix, similarity, adaptive, estimation, LMedS

摘要： 如何从初始匹配点集中估计出精确的单应性矩阵，有效地剔除误匹配，一直以来都是视觉领域研究的重点和难点，也是实际相关技术应用中最为关键的一步。通过将特征点对相似度概念应用于LMedS的样本选取过程，提出了一种新的单应性矩阵自适应的估计方法。区别于传统LMeds方法从无序匹配点集中随机选取样本的过程，该方法首先以点对间的相似度对整个初始匹配点进行降序排列，然后从前往后依次选取样本。实验结果表明，与LMedS相比，该方法估计出的单应性矩阵更精确、鲁棒，效率更高（得到最佳模型所需的迭代次数仅约为LMedS的1/5），同时弥补了RANSAC及其改进方法需预先设置距离偏差阈值的不足。

关键词: 单应性矩阵, 相似度, 自适应, 估计, LMedS