3D city scale reconstruction using wide area motion imagery
3D reconstruction is one of the most challenging but also most necessary part of computer vision. It is generally applied everywhere, from remote sensing to medical imaging and multimedia. Wide Area Motion Imagery is a field that has gained traction over the recent years. It consists in using an airborne large field of view sensor to cover a typically over a square kilometer area for each captured image. This is particularly valuable data for analysis but the amount of information is overwhelming for any human analyst. Algorithms to efficiently and automatically extract information are therefore needed and 3D reconstruction plays a critical part in it, along with detection and tracking. This dissertation work presents novel reconstruction algorithms to compute a 3D probabilistic space, a set of experiments to efficiently extract photo realistic 3D point clouds and a range of transformations for possible applications of the generated 3D data to filtering, data compression and mapping. The algorithms have been successfully tested on our own datasets provided by Transparent Sky and this thesis work also proposes methods to evaluate accuracy, completeness and photo-consistency. The generated data has been successfully used to improve detection and tracking performances, and allows data compression and extrapolation by generating synthetic images from new point of view, and data augmentation with the inferred occlusion areas.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 License