An improved SIFT algorithm for binocular vision

In this paper we proposes an improved SIFT algorithm of feature points matching, in order to improve the matching efficiency and `real-time´. First, we use the classical SIFT algorithm to extract image features, when formatting the feature descriptor, we use the gradient normalized; Second, since the SIFT feature points detected a large number of feature points, and each is 128-dimension vector, but for the matching algorithm based on the Euclidean minimum distance measure, we have to calculate the Euclidean distance of a point and all the feature points from the right image, then ascending sort, which leads to low efficiency of algorithm. Based on optical imaging theory and the theory of binocular vision, by the coordinate of each feature point of the first image, we narrow the search range of the second image from two directions of row and column, on the basis of maintaining the matching accuracy, we improve the efficiency of the algorithm.