In the conventional Augmented Reality (AR) systems, three main tasks are sequentially performed: image acquisition, camera pose tracking and rendering. Among these tasks, the camera pose tracking occupies the largest proportion of total processing time. Thus, frame rate can be reduced in small mobile devices when the tracking processing time exceeds the image acquisition time interval needed for desired frame rate. Moreover, the temporal jitter can be introduced due to the irregular tracking processing time. In this paper, we present a new framework for AR that can achieve the high frame rate and reduce the temporal jitter. In the proposed method, the three main tasks of AR are performed in parallel. This enables the image acquisition and rendering thread to work independently with the tracking thread so to prevent frame loss and temporal jitter. The tracking thread calculates the camera pose of the coarsely selected frames. For the unselected frames, the parabolic motion model is applied. Experimental results indicate that the proposed system outperforms conventional methods in terms of frame rate and temporal jitter.