Several tracking techniques for augmented reality have been proposed. In feature point tracking, a pose is computed by minimizing the error between the observed 2D feature points and the back-projected feature points from the 3D scene model. This minimization problem is usually solved by nonlinear optimization. The main advantage of this approach is its accuracy. However, it is difficult to compute the correct pose unless an appropriate initial value is used. In addition, when an observation contains some errors, this approach does not guarantee a correct pose even if it converges to the global minimum. Therefore, once an incorrect pose is computed in a frame, either the tracking in the next frame may fail or the result will deviate from the correct pose. In this paper, we propose a new tracking framework for augmented reality. The proposed method tracks features as multiple local hypotheses based on not just one pose but multiple poses that are computed from pose estimation in the previous frame. Since multiple poses are maintained as global hypotheses, as long as the correct pose is contained in the hypotheses, tracking can be continued even in difficult situations such as a simple iterative scene with high-speed movements.