Research On The 3D Object Registration Method In Augmented Reality And Its Application

Augmented Reality (AR) technology is a hotspot which has a wide range of applications and broad prospects. Being different from the traditional virtual reality (VR) technology, AR augments the real environment with the virtual contents generated by the computer to enhance the user's perception. The virtual contents and the real environment are seamlessly integrated, and the communication interface between the user and real world is enhanced.In augmented reality, comparing with the registration method for planar target, the 3D object registration technology has the problems such as huge searching space and long computational time. Aiming at these problems, the research works of the thesis are as follows:Firstly, aiming at the mark-less 3D object recognition and registration problem, a real-time object recognition algorithm based on discrete gradient feature is presented. The aspects of the object are recorded into a series of views, and each view is divided into several information rich sub-regions. The recognition is performed by using the discrete gradient feature. The computational efficiency of the view-based 3D object recognition task is significantly enhanced by the binary encoding and the parallel computing technology, which can meet the requirement of the AR frame rate (≥15fps). And then, a machine learning based FAST algorithm is used to extract the feature points, and the BRIEF operator is used to descript and to match the feature points.Secondly, theories and methods are proposed for the stability and efficiency of the camera pose estimation process in the AR registration. (1) A direct solution of Perspective-3-Point problem based on a new geometric constraint named Perspective Similar Triangular (PST) is presented. By constructing the new geometric constraint, the number of the unknown variables of the equation system was reduced, which significantly enhanced the numerical stability and the accuracy of the solution. (2) The configuration of the 2D/3D corresponding point pairs in Perspective-n-Point problem is studied, and a new middle state, we called "quasi-singular", is proposed. A high robust algorithm RPnP is designed to resolve the stability degeneration problem in the quasi-singular case. RPnP is the first non-iterative solution for the PnP problem and can achieve more accurate results than the iterative algorithms when no redundant reference points can be used (n≤5), and large-size point set can be handled efficiently due to the O(n) computational complexity.Thirdly, an efficient lookup table (LUT)-based camera pose estimation method is presented. It can achieve high stability in the presence of noise with very little computational time, and it is very suitable for the AR applications on mobile equipments with limited computational resources.Finally, aiming at the product assembly application in the virtual-real environment, an AR assembly prototype system is developed. The user is guided by the augmented information to perform the correct assembly operations. The effectiveness of the 3D object registration method in AR is validated by this prototype system.