A new corner detection algorithm based on the anisotropic Gaussian directional derivatives (ANDDs) autocorrelation matrix on edge contours is proposed to suppress noise and local variation, and to detect corners effectively. Firstly, the edge map of an image is extracted by the Canny edge detector. Secondly, the input image is smoothed by the ANDD filters; autocorrelation matrices are constructed for each edge pixel by the directional derivatives correlation of the pixel and its surrounding pixels. Finally, the contour pixels with local maxima of the sum of the normalized eigenvalues are labeled as corners. The proposed algorithm is different from the traditional contour-based detectors, and it uses the intensity variation auto-information on contours and their surrounding pixels rather than the curvatures of the planar curves, hence has better robustness to noise. Experimental results and comparisons with several state-of-art algorithms in both the noise-free and noise cases show that the average matched corner numbers of the proposed algorithm increase by about 7.4 and 9.3 percent, respectively; and the average positioning errors reduce by about 10 and 15.2 percent, respectively.