Three-Dimensional Registration Method of KLT Inter-Frame Tracking and Improved ORB Feature Detection in Augmented Reality

Due to the problems of poor real-time performance and unstable registration of visual-based augmented reality 3D registration technology, a 3D registration method based on KLT inter-frame tracking and improved ORB feature detection and matching is proposed in this paper. The registration process in our system is directly divided into offline stage and online stage. In the offline stage, the template image is trained with the improved ORB algorithm to generate a subset of feature description. In the online stage, I match firstly the global video frame with the template features, and then optimize the matching results through the RANSAC algorithm, ultimately through the feature matching threshold to determine whether there is a target in the video frames. If there is a target in the scence, the KLT tracking algorithm is used to track the target. By solving the homography matrix between the video frames, the homography matrix relative to the template is obtained at any time. The homography matrix obtained by solving uses the projection imaging theory and the PNP method to complete the calculation of the three-dimensional registration matrix. In addition, for KLT, it is easy to track the drift, and it is analyzed by establishing a threshold, and if the tracking fails, the target is detected again. Experiments show that the three-dimensional registration method proposed in this paper can work normally under the condition of target rotation and partial occlusion. The real-time FPS reaches more than 15fps, and the registration stability is higher than the KCF + ORB three-dimensional registration method.