An improved ORB feature matching algorithm

Abstract

An improved ORB feature matching algorithm is proposed to solve the problem of decreasing matching accuracy under the conditions of image blur, light change, image compression and noise. First, in the process of extracting feature points, the image is meshed and quad-tree structure is introduced to make the extracted feature points evenly distributed in the image, thus solving the problem of feature points concentration encountered by traditional feature extraction methods. Then, the brute-force matching is used for preliminary matching, and cross validation is adopted to eliminate some mismatches and improve the result of brute-force matching. Finally, Gaussian kernel is used to weight the results of grid-based motion statistics to optimize the statistical results and further eliminate the mismatches to obtain the matching set with higher accuracy. The experimental results show that this algorithm improves the average accuracy by 3.5%, 4.2%, 1.8% and 6% respectively under the conditions of image blur, light change, image compression and noise.