Robust Corner Detection Using Linear Fitting Error Estimation

Abstract

Image corners have been widely used in image processing and computer vision applications. For a variety of contour-based corner detection methods the key step is curvature estimation, which reflects the bending strength of a curve. Corner detection aims to achieve the lowest possible computational cost and maximal detection accuracy. Instead of estimating the discrete curve directly like most contour-based corner detectors, we propose to utilize the linear fitting error to measure the corner response strength of the contour points. The proposed method consists of three aspects: First, a small curve segment is parameterized to two curves; then the minimum linear fitting errors with respect to the above two curves are estimated by employing the least-squares fitting technique; Finally, the estimated minimum fitting errors are considered as the local bending strength of the small curve segment for corner finding. The experimental results show that the proposed corner detection algorithm outperforms the five state-of-the-art corner detection methods in terms of localization error and average repeatability.