This paper illustrates a novel registration method for augmented reality systems based on multiplanar structures and natural feature tracking. Our method distinguishes itself from other methods in the following ways: First, our method can use arbitrary multiple planes without any information on the physical relationship of the planes. Second, we put forward a new method to establish the world coordinate system. Compared with the traditional method, our approach needs only three noncollinear points of the real world and really enhances the usability of our system. Third, we propose novel classification based local and global tracking strategies to match features between reference images and the current frame directly. Our method casts off the requirement that the initial camera position should be close to the reference images while overcoming the problem of error accumulation. Some experiments have been carried out to demonstrate the validity of the proposed approach.