Online learning of tracking and registration based on natural scenes

Registration is a fundamental technology for augmented reality. In this paper, a registration approach is proposed to accurately track the natural scenes. The matching method of SURF (speeded up robust features) descriptor is first improved to keep the initial registration matrix validity. Then, effective online learning of the scenes is used to improve the registration accuracy. Lastly, the registration matrix of the previous frame is utilized to rapidly restore the lost key points and accelerate the speed of registration. Experimental results show that the proposed method can keep smooth tracking for video frames and maintain high accuracy of registration.