Abstract: In order to achieve high-quality augmented reality,the problem of illumination consistence between virtual objects and real scenes should be solved.High dynamic range technique can be used to recover the environment map of the scene.However,for using this technique,we need to first solve the align problem between the recovered illumination environment and the real scene.This paper proposes a feature-based alignment algorithm,which can automatically align the illumination environment map with the captured real scene.First,Affine-SIFT is employed to extract and match the common features between the environment map and the captured video.RANSAC method is used to remove the outliers.Then the 3D positions of the matched feature points can be computed by structure-from-motion technique.Finally,according to these 3D matches,the correlation between the environment map and the real scene can be obtained and the automatic alignment is accomplished.Based on this technique,this paper presents a high-quality real-time augmented reality system.The proposed system employs a key-frame based real-time camera tracking technique to robustly register the virtual objects into the online video stream.The users are allowed to real-time edit the realistically inserted virtual objects.During the rendering pass,with the auto-aligned illumination environment,we use the importance sampling algorithm and shadow mapping technique to realistically render the inserted virtual objects.The experimental results demonstrate that the proposed augmented reality system can well address the geometry and illumination consistency problems in real-time composition.