In this paper, a novel method for tracking without a marker was proposed. This method alleviated many constraints in augmented reality systems based on fiducial markers. To improve the tracking stability in multi-scale, the effective features and space geometry structure were detected, and the effective features were searched in a multi-scale model. Based on space geometry structure of features, the effective feature points were selected and fictitious features points were wiped off. Experiments show that the proposed algorithm works robustly in the video with scaling objects. The feature points reprojection error of every frame is smaller than one pixel, and the average of feature points reprojection error is kept in sub-pixel.