Robust matching algorithm based on SURF

For matching visible image with many similar regions, the SURF matching algorithm based on Euclidean distance has disadvantages of limited matching constrains, higher false matching rate and difficult to remove error points effectively. To resolve these shortcomings above, a robust matching algorithm was proposed in which a combined measure of Cosine Similarity and geometry consistency were adopted to dispose multidimensional feature vectors. We compared with the original matching algorithm that only use Nearest Neighbor (NN) Searching and Random Sample Consensus (RANSAC) algorithm, our method performs well for eliminating error matching points, especially in terms of many similar regions of validity.