3D image geo-registration using vision-based modeling

Image geo-registration is the process of relating a photograph and its pose to referenced world coordinates. The application is relevant, especially to the social networking, photo-sharing, and intelligence communities, where the locations of objects of interest are to be determined. This paper proposes an algorithm that identifies and geo-registers query photographs in the absence of any meta-data that are spatially co-located with training sets of user-collected images. Training images construct a 3D model using inherent structure from motion between images. Using SIFT features, the 3D model fuses co-located points and generates an averaged SIFT comparison feature. This paper also advises on training set augmentation to mitigate deleterious illumination, seasonal, and weather effects by introducing methods of merging separate point clouds. As such, probability of detection (and therefore, registration performance) is enhanced. After the 3D model is generated, each test photograph is matched to the 3D point cloud and registered using similar camera refinement and positional optimization techniques. The results provide accuracy to within a few meters of the absolute geo-coordinates.