Tracking as recognition: A stable 3D tracking framework

This paper proposes a novel framework for monocular tracking of object undergoing rigid transformation in 3D real world. In this method, instead of computing object motion only with respect to one previous frame as many previous 3D tracking algorithms did, we combine the idea of 'tracking as recognition' in our approach, using the global information of the object in the image to obtain an approximate absolute object pose, which can prevent tracking drift accumulation. Two-frame registration methods are also employed to refine this approximation and approach optimal. Our approach provide robustness for tracking as it can avoid infinite error accumulation and large estimate drift during processing.