A high precision visual odometry is proposed.Small computational amount and high precision algorithm—SURF is used for the extraction of feature points and Hausdorff diatance is introduced for the matching.The transformation matrix of camera coordinates of adjacent frames are got by the nearest neighbor vector-matching algorithm and the precision is ensured.A method of odometry calibration is proposed to obtain the displacement information under the robot starting point coordinate and this method could make sure of the precision when the posture of robot changed.The feasibility of vision odometry is verified by the round trace experiments.