Homography Estimation Based On Error Elliptical Distribution

How to estimate accurately the homography is always a challenging problem in computer vision. In the reported literature, the measurement error of the image points is usually assumed to obey isotropic Gaussian distribution. However, real data very seldom follows this assumption. This paper proposes an estimation of homography under the assumption of image point errors following elliptical distribution, which is more coincident with real data. In the proposed method, the adaptive-scale elliptical residual kernel consensus (ASERKC) robust estimator is used to filter out inliers which are utilized to compute homography. Then, the elliptical weighted L-M (EW L-M) algorithm is optimized the homography. The experimental results show that the proposed method may present a more accurate homography. Especially when we applied it to incremental structure-from-motion (SFM), we find that the exact homography matrix is useful to select a better initial image pairs which can help obtain a more complete 3D points cloud.