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

ABSTRACT: While the state of the art in both single-image reconstruction algorithms and multi-view structure from motion algorithms have advanced considerably in recent years, little work has been performed which leverages the constraints relied upon by both approaches. When an area of interest is imaged from a long distance with little angular diversity in viewpoint, multi-view constraints alone are often insufficient to produce an accurate 3-d reconstruction. It is proposed that the additional constraints provided by surface properties learned from image data will improve reconstruction performance significantly. The proposed Phase I effort is focused on the development of an aerial image-based 3-d reconstruction algorithm that combines the relative strengths of both single-image reconstruction and context algorithms and state of the art multi-view stereo. The result is an automatically generated 3-d model that is optimally constrained by all information contained in a set of collected aerial images. The proposed system is general enough to exploit high angular diversity datasets, but exhibits graceful degradation as the viewpoint diversity decreases. The decrease in information due to low view angle diversity is compensated by single-image constraints on surface orientation derived by machine learning algorithms. BENEFIT: Benefits of the proposed approach include improved sensor model estimation and high accuracy 3-D modeling capabilities. Applications include support of downstream processing (tracking, geo-positioning, geo-registration), augmented reality / situational awareness, and simulation/training.