Abstract: Based on affine reprojection and natural feature tracking techniques,a novel registration method for markerless augmented reality systems is proposed.This method represents the virtual objects in Euclidean space instead of in affine space,which is consistent with the popular graphics rendering tools such as OpenGL and is more common in practice.A new classification based local and global tracking strategy to match features between the reference image and the currently frame directly is presented,which enables the augmented reality system working effectively even under large changes of viewing angles.A novel affine coordinate system definition method based on a collection of feature points is also proposed,which makes the tracking method effective even when not all the reference points are detected during the whole process,as long as there are at least four pairs of reference points correspondences.The experimental results showed that the proposed method achieved 18fps which is fast enough for augmented reality applications,and the registration errors are within 5mm demonstrating the high precision of the proposed registration method.