A new docking method for mobile robot is proposed. The desired docking state (desired position and orientation) of the robot is defined by a pre-captured reference image. The scale invariant feature transform (SIFT) algorithm and a double direction best bin first (BBF) based feature matching algorithm are used to extract the visual feedback information by matching the current view of the docking station and the reference image. The robot is aligned against the reference image based on a epipole servoing strategy. A centroid tracking is used to avoid the object escaping from the field of view. A random sample consensus (RANSAC) algorithm is used to solve the affine transform between current and reference images, and a control rule at last stage is presented to realize precise docking. No more knowledge of the scene geometry or artificial marks are needed. The experiment results of a realistic indoor environment demonstrate the effectiveness of the proposed method.