Abstract: This paper presents a supervised machine learning method to detect and track complex man-made logos in real-time. The key-frame based registration method is applied to estimating the camera pose and the randomized tree method is used to matching key-points which are extracted from the input image and from key-frames. In order to overcome the problem of false feature matching caused by the repetitive texture in the real environment, a false feature matching recovery mechanism is also proposed to effectively improve the feature matching performance. The presented algorithm has been applied to the mobile augmented reality system.