Abstract
We propose a fast vision-based pose tracking method, which is suitable for mobile augmented reality. Conventional pyramid-based approaches have been faced with the problem of complexity since their processing pyramid level is fixed regardless of camera motion state. Our method estimates the motion state initially from the movement of a subset of keypoints to be tracked, and switches processing pyramid levels according to the motion state. Experimental results demonstrate that our method speeds up the conventional pose tracking by about 37% without loss of robustness and precision.
