Nowadays one of the key problems that influence the performance of an AR (Augmented Reality) system is the registration error. It is common that in the current AR systems a virtual object appears to swim about as the user moves, and often does not appear to rest at the same location when viewed from different directions. In order to provide a stable tracking result for our AR application, a hybrid tracking scheme that combines the robustness of the magnetic tracking and the static accuracy of the vision based tracking is developed. The principle of the vision-based tracking is presented and the tracking accuracy of the rotation angle is studied. A magnetic tracker composed of magnetoresistive sensors and accelerometers is proposed to compensate the shortcomings of the vision-based tracking. The algorithm to calculate the position and orientation of the tracked object by combining the calculation result of the magnetic tracking and the vision-based tracking is analyzed. The setup and the experimental results of the proposed AR system are given. The results validate the feasibility of the proposed AR system.