Enhanced Reality Visual-based High-precision Target Tracking Technology

Augmented Reality (AR) is a new technology, which has a good application prospect in area of medicine, military, entertainment, education, manufacturing and maintenance, robotic navigation etc. Augmented Reality aimed at enhancing the user's visual perception of the physical world with computer-generated virtual information. Registration technology has an important role in AR systems, and registration based on computer vision is hot investigated in AR. Virtual information, such as rendered 3D models, 2D textures and labels, must appear correctly projected onto live video captured by a mobile camera. To achieve such a synthesis, it is necessary to recognize what is viewed by the camera, and to accurately localize the camera and the virtual objects in the real world. Currently, a mature video registration technology which is accomplished by placing artificial marker in the scene is widely used in AR systems. Because of requiring pre-placed artificial markers, this method limited its application in outdoor environments. Therefore, developing an AR system, which is fit for complex environment, is extensively researched.In this thesis, by analyzing the research of AR, a registration technology which is based on natural features of image was proposed, and an AR system fit to complex environment was schemed out. The algorithm of this thesis was divided into two steps:(1) Off-line processing: Firstly, a set of reference images of objective scene was captured by a mobile camera. Afterwards, SURF algorithm was applied to extract the image features and matched image pairs were obtained by using feature matching technology. Lastly, 3D structure model of objective scene and camera parameters of reference images were produced by bundle adjustment.(2) On-line processing: SURF algorithm was firstly applied to extract image features of current frame. And then the image features of current frame were used to match with the reference images. And according to the 3D structure of objective scene, a set of correspondences of 2D to 3D was produced. Finally, the camera parameter of current frame was obtained by the correspondences, and the position of virtual objects was determined eventually. After the above steps, an AR system based on natural features was schemed out. At the end of this thesis, simulations were carried out of the algorithm. It is proved that this method which is robust to viewpoint varying, partly sheltering and model partly missing can achieve the goal of tracking the camera accurately.