Automated 3D Reconstruction of a Scene From Persistent Aerial Reconnaissance Video at High Zoom

3D scene model generation is a valuable tool for mission planning, precision targeting and sensor data exploitation. Our team of Systems & Technology Research (STR) and Vision Systems Inc. (VSI) proposes to develop, evaluate, demonstrate and deliver a 3D reconstruction application which provides a user-friendly interface for reconstructing accurate 3D mesh models from narrow field-of-view (NFOV) aerial video sequences with challenging collection conditions such as uncontrolled zoom, planar scene content, and metadata burn-in. STR and VSI have developed a voxel-based reconstruction framework called Probabilistic Volumetric Reconstruction (PVR) which excels at reconstructing detailed 3D models from aerial imagery. This technique fuses stereo appearance probabilities from all images using belief propagation to produce a single composite 3D model, and works particularly well in challenging textureless areas. VSI has also developed a structure-from-motion (SFM) technology for estimating camera model parameters for long video sequences called SIGMA (Speedy Imagery Geo-registration and Motion Analysis). In Phase II we plan to update PVR and SIGMA to handle viewpoint-limited data with frequent uncontrolled zooms and other challenging operating conditions. We also plan to develop a lightweight graphical user interface (GUI) to run the end-to-end reconstruction and provide feedback at all stages of the reconstruction process.