In this study, a method for landmark selection from image streams captured by a camera mounted on a mobile robot is described. To select stable visual landmarks for mobile robots, two measures regarding landmark “visibility” are considered: distinctiveness and repeatability. In the proposed method, several neighboring feature points form a visual landmark and their distinctiveness is evaluated in each image. Then, under the assumption that a robot can actively seek a feasible landmark, the repeatability of the landmark is evaluated. Weighting techniques using feature-position relations are proposed, and landmark selection criteria using a variation coefficient are employed. These allow us to select high-visibility landmarks. Experimental results obtained using a real mobile robot demonstrate the effectiveness of the proposed method.