Registration method of uncalibrated camera applied to SfM point cloud

As a research focus of Augmented Reality, functional technology with respect to markerless registration has few implementations. In order to enrich the development of markerless registration and broaden the scope of application, a novel method based on SfM point clouds is presented to have both calibration and registration done, by which taking a single query image from an uncalibrated camera. A set of SfM point clouds is to match with the features of query image, and directly returns homographic matrix with wrong matches removed by RANSAC, focal length is to be acknowledged by Kruppa equation afterwards. LevenbergMarquardt algorithm of PnP problem then is employed to calculate the rotation and translation of camera. Also we build a particle system based on GPU to observe our conclusion directly. Experimental results demonstrate the validity of estimation with this method.