Accurate and robust feature-based homography estimation using HALF-SIFT and feature localization error weighting

Abstract

The goal of homography estimation is to find global transformation between two images of the same scene taken from different viewpoints. The feature-based homography estimation method uses a local feature extractor, a RANSAC-like method and the Levenberg-Marquardt method to estimate the homography matrix. However, in practical applications, the accuracy and robustness of homography estimation are significantly affected by feature localization error. In this paper, we first use the HALF-SIFT method to compensate for localization error caused by the feature extraction method and use the covariance matrix to represent localization error caused by image noise. Then, we proposed a new inliers selection method named CW MLESAC and a new homography matrix refinement method named CW L-M to improve accuracy and robustness. Experimental results show that the proposed method is more accurate and robust under different noise levels and inlier ratios than state-of-the-art methods such as LMedS, RANSAC, MSAC and MLESAC.