This paper presents a vision-based registration method for Augmented Reality by merging multiple planar markers, which are placed at arbitrary positions and poses. In our approach, we use planar markers like AR-Toolkit markers. We need no information about geometrical relationship of the markers while most of marker-based approaches require the measurement of the markers' 3D positions and poses. Thus we can implement an AR system very easily without time-consuming calibration of markers and a camera.In the proposed registration method, we assign a 3D co-ordinate system for each marker independently, and compute a projection matrix between each marker's 3D coordinate system and the input image coordinate system. For merging all markers' 3D coordinate, we construct a projective 3D space defined by projective geometry between two reference images which are selected while the on-line process automatically. This merging process enables to use the multiple markers at arbitrary positions and poses, and also to reduce errors in registration. For demonstrating the effectiveness of the proposed method, we implement a real-time on-line AR system. Virtual objects are superimposed onto input image sequences that are captured with a handy camera. Since the multiple markers are placed in the wide area, the virtual objects can move around the real world widely. Moreover, the registration can be carried out stably, because the markers face to various directions and some of them can be always recognized from most view points.