Updating Globally-Referenced Sub-Decimeter-Accurate Visual 3D Reconstructions
Tucker C. Haydon and Todd E. Humphreys, University of Texas at Austin
Alternate Number 3
A technique is presented that assesses and incrementally updates a visual 3D reconstruction using a near-minimal set of images such that the final reconstruction is both globally referenced and centimeter-accurate. Starting with an initial 3D point cloud, a surplus bank of images, and the corresponding cm-accurate Earth-Centered-Earth-Fixed (ECEF) image poses, the technique evaluates the covariance of each point and determines a subset of images that, when incorporated into the reconstruction, produce a new reconstruction that is survey-accurate in the sense that each of the feature points is registered to a global frame with centimeter accuracy.
The technique operates iteratively with images taken from a camera- and Carrier-Phase-Differential-GNSS-equipped Micro Aerial Vehicle (MAV), scoring each image in the bank according to how it may reduce the covariance of feature points, and then greedily incorporating the images into the reconstruction until the accuracy constraints are satisfied. In testing, the technique iteratively updates a reconstruction to an accuracy of 1 cm, 1 sigma.
A MAV then references the reconstruction as it flies through the modeled scene,
demonstrating that the reconstruction is sufficiently accurate to be used for