An improved algorithm for Harris rapid sub-pixel corner detection is proposed by considering the limitations of inaccurate localization and the low efficiency of the Harris algorithm. The improved algorithm is used to reject non-corners and false corners by screening the corners two times. After detecting the initial corners with the non-maximum suppression of the corners response function

the initial corners are refined to sub-pixel corners by the Euclidean distance between the corner cluster and an ideal corner using a weighted least squares method. The location of the initial corners are considered as the center of the searching corner cluster

and the value of the corners response function is considered as weight. Our tests show that the improved algorithm is effective and practical.