Research Of Unknown Environment Reconstruction And Camera Pose Estimation

Augmented Reality is a high - technology with steady development and wide application in recent years, it comes from Virtual Reality technique and extend based on VR research. The nature of AR is to embed the virtual object generated by computer into real-world, and make the integration between virtual scene and reality in order to give people a sensitive impression that the information of real world have been added and enhanced. Registration is one of the most important problems that must be solved in AR system as well as a key factor that limits the application of AR. Generally speaking, the nature of registration is to compute exactly the real-time transformational matrix between camera and real world so that we can place virtual object at special position. As the development of computer technology, registration based on computer vision becomes more and more popular in AR area.For the augmented reality systems at present, most of them are applied for one scene, and make use of the known mask to register offline, so in this paper, together with pattern recognition, theory of computer vision and image processing techniques, we design a new and robust algorithm in unknown environment to solve the real-time registration of multi-scenes. At the same time, we have improved the accuracy and stability as well as extended the application area of AR.The thesis has accomplished some research work of registration based on scene recognition and natural feature tracking, including:(1)We propose a registration method using planar nature features, solve the problem of view-limiting in wide-area scene and finish registration successfully without the marker.(2)We present a key-frame recognition algorithm based on random classification trees, and make the system distinguish the object scene quickly so that the AR system can initialize or recover from the failure of registration.(3)We build a registration system in unknown environment based on scene recognition connected with local reconstruction technology. In the end, we realize the accurate registration of multi-scenes online.