LAUSR

This paper proposes an adaptive scale-invariant feature matching method based on data clustering, to solve the problem of low robustness of the KD tree matching method caused by SIFT feature noise sensitivity, and our method can be used to AR applications. The method has two stages: offline data re-clustering and online two-stage feature matching. This paper is the first to present a Vocabulary-KD data structure which achieves SIFT using KD tree by tuning the number of features of the Vocabulary nodes. Moreover, based on the Vocabulary-KD data structure, an adaptive feature matching method is proposed, which is consist of two clustering, one on the feature sets and the other on the feature sets contained by the leaf nodes of the Vocabulary-KD tree, along with adaptive adjustment of the relevant parameters of the Vocabulary-KD tree. At last, key images are selected in real-time for the second stage feature matching. The different results show that the proposed method can effectively resist noise, improve the adaptivity of the SIFT feature matching method, so as to achieve the trade-off between efficiency and robustness.