3-d Scene Reconstruction from Aerial Imagery Thesis

3-D scene reconstructions derived from Structure from Motion (SfM) and Multi-View Stereo (MVS) techniques were analyzed to determine the optimal reconnaissance flight characteristics suitable for target reconstruction. In support of this goal, a preliminary study of a simple 3-D geometric object facilitated the analysis of convergence angles and number of camera frames within a controlled environment. A series of 2-D images were acquired at convergence angles from 1◦ to 100◦ in 1◦ increments with the number of images varied from 2 to 20 at each angle. Reconstruction accuracy measurements revealed at least 3 camera frames and a 6◦ convergence angle were required to achieve results reminiscent of the original structure. Furthermore, improved results are realized with additional camera frames and expanded convergence angles enabling refinement of the focal length and camera motion estimation. The central investigative effort sought the applicability of certain airborne reconnaissance flight profiles to reconstructing ground targets. The data sets included images collected from a synthetic 3-D urban environment along circular, linear, and s-curve aerial flight profiles equipped with agile and non-agile sensors at look angles ranging from 0◦ (nadir) to 60◦ in 15◦ increments. S-curve profiles and dynamically controlled linear flight profiles produced the most diverse data sets resulting in superior reconstruction accuracy and density of points. Linear profiles equipped with non-agile cameras failed to reproduce identifiable results at near nadir look angles, but the results were dramatically increased when multiple orthogonal passes were combined and only overlapping images employed. Furthermore the effects of prominent images pivotal to the reconstruction processes were analyzed where a bimodal structure was observed relating the frequency of image use for each reconstructed 3-D vertex.