一种基于数据聚类的鲁棒SIFT特征匹配方法

A Data-Clustering Based Robust SIFT Feature Matching Method

针对噪声敏感造成的SIFT特征匹配鲁棒性低问题,提出一种基于数据聚类的两阶段特征匹配方法.在满足特征匹配几何距离最邻近本质要求下扩展了kd数据结构,使其不但能够完成算术平均化匹配特征离线聚类,而且能够实现第1阶段聚类特征在线匹配.在此基础上,给出一种概率最优投票策略选择关键图像进行第2阶段匹配,最后合并两阶段属于关键图像的所有匹配特征对.实验结果表明,对于大量存在重叠关系的图像集合,该方法能够有效减少重复特征数量,降低噪声信息对特征匹配的干扰,极大地提高特征匹配的鲁棒性.

We present a data clustering method for robust SIFT matching. Our matching process contains an offline module to cluster features from a group of reference images and an online module to match them to the live images in order to enhance matching robustness. The main contribution lies in constructing a composite k-d data structure which can be used not only to cluster features but also to implement features matching. Then an optimal keyframe selection method is proposed using our composite k-d tree, which can not only put the matching process forward but also give us a way to employ a cascading feature matching strategy to combine matching results of composite k-d tree and keyframe. Experimental results show that our method dramatically enhances matching robustness.