Long-distance 3-D Reconstruction from EO/IR Imagery

ABSTRACT: The production of 3D models from aerial imagery has been an active research area for many years. Given sufficient image texture and a variety of divergent camera viewpoints, camera pose and scene structure are fully recoverable. Further, software for fully reconstructing a scene from an image set is now freely available. As the range of viewpoints decreases, however, traditional methods of 3D structure-from-motion (SFM) become mathematically ill- conditioned, resulting in noisy and/or erroneous results. Since real-world mission flight patterns often do not permit imagery of target terrain from the ideal range of viewpoints, current SFM methods are often unable to supply accurate 3D terrain models even during flights where extensive video imagery is acquired. Neya systems will exploit two types of learned heuristic scene reconstruction from individual images can help address this shortcoming. The first approach learns global, persistent 3D shape cues from image features culled from a library of images labeled with ground-truth structure .The second approach learns a temporally-evolving model of transient cues specific to target weather, lighting, and scene conditions. Our proposed reconstruction approach exploits both types of monocular processing along with traditional multiview algorithms to create a 3D terrain estimation system that operates without angular view constraints. BENEFIT: The economic and military of accurate aerial modeling from video is already established, and there are a number of existing companies that profitably market such services today (see, for example, http://www.geoinfra-engineering.com/en/services/aerial-photography-and-3d- terrain-modelling). In many current defense-related missions, gathering a wide range of target views is either impossible (i.e. cross-border surveillance into denied airspaces) or undesirable (to reduce flight time and/or fuel cost).The proposed work, providing an enhanced modeling-from-video capability without requiring specific flight patterns, will be of interest to many users inside and outside the military: aerial terrain modeling has applications in agricultural, construction, mining, and forestry applications, for example. Optimizing the product for real-time use will open up additional markets in defense and border security.