This paper presents a new method for defining and extracting visual landmarks for indoor navigation using a single camera. The approach considers that navigating from point A to point B amounts to navigating to intermediate positions, which are signified by recognition of local landmarks. To avoid the pose problem we seek scene representations that rely on clustered corners of physical objects on corridor walls. These representations are scale and translation independent and allow for the construction of a metric that can match pre-detected landmarks of a learning phase with landmarks extracted from images captured at run-time. The validity of our approach has been verified experimentally.