Abstract: Homography plays an important role in visual metrology, camera calibration and 3D reconstruction. However, how to robustly and accurately estimate a homography from images is still a difficult problem in practice. In this work, we found that in the line-correspondence based homography estimation by a DLT(Direct linear transformation)-like method, some gross estimation errors could arise under some camera configurations, especially when there exists image line(s) passing through (or close to) the origin of the image coordinate system. The underlying reason is that under such configurations, the magnitude of the elements in the measurement matrix could vary significantly, which in turn results in a large condition number of the measurement matrix, and non-robustness of estimation. To alleviate this problem, a new normalized estimation method is proposed in this work, which consists of a new data pre-normalizing step, followed by a standard DLT estimation, and in which the robustness and accuracy of the estimated homography are substantially enhanced. Our new method retains the traditional DLT's conceptual simplicity and computational efficiency. Extensive experiments with both simulated data and real images validated our method.