This paper presents a method of selecting landmarks for real time robot navigation. We first build a 3-D landmark map of an unknown environment with trinocular vision, and then the robot with monocular vision localizes the own position by matching SIFT keypoints extracted from the input image to the landmarks based on their visual appearances. It requires much computation to detect visual features corresponding to all the visual landmarks. There are many redundant landmarks among them and they can be excluded in the calculation without major degradation of localization. We propose to select useful landmarks considering their contributions to the localization accuracy derived from the analysis of geometric error propagation, and both the reliability of detection and matching of the corresponding visual features. We compare the proposed method with the one which uses all the landmarks by experiments in real indoor environments. The result proved that the localization accuracy of the proposed method is almost the same, and the computation time is greatly reduced.