This paper proposes a new relative map method for the efficiency of the monocular EKF-SLAM. The landmarks in the traditional EKF-SLAM are fully covariant with each other and cause a quadric computational complexity according to their number in the map. Our approach maintains the map with the relative locations of the landmarks and the corresponding believes, and only a small number of the landmarks are active in each EKF step. So the computational cost could be restricted. The absolute locations of the features can be retrieved with a fast covariance intersection filter. The experiment shows that this method could maintain the computational complexity with only a little cost of the global map consistence.