We propose a camera tracking method by on-line learning of keypoint arrange- ments in augmented reality applications. As target objects, we deal with intersection maps from GIS and text documents, which are not dealt with by the popular SIFT and SURF descriptors. For keypoint matching by keypoint arrangement, we use locally likely arrange- ment hashing (LLAH), in which the descriptors of the arrangement in a viewpoint are not invariant to wide viewpoints because the arrangement is changeable with respect to viewpoints. In order to solve this problem, we propose online learning of descriptors using new configurations of keypoints at new viewpoints. The proposed method allows keypoint matching to proceed under new viewpoints. We evaluate the performance and robustness of our tracking method using view changes.