Points and lines for vision-based navigation using 3D structure and localization
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
Reliable vision-based navigation is a critical challenge for unmanned aerial ve-hicles (UAVs) operating beyond line-of-sight (BLOS). Traditional methods relying on GPS and inertial measurement units (IMUs) may not work reliably in such con-ditions, necessitating new vision-based techniques for autonomous navigation. A novel framework is presented that combines points and lines for 3D reconstruction and localization in challenging environments, alongside tools supporting navigation, including visualization and verification of 3D-2D matching and ground truth data. This work addresses limitations in perspective-n-point (PnP), 3D-2D correspon-dences, and bundle adjustment, particularly where point features fail, by advancing PnP with novel perspective-point-line (PPL) techniques. Geometric constraints were derived from minimal point-point, point-line, and line-line correspondences, yielding al-gebraic surfaces resembling a pumpkin, neck pillow, and double umbrella, respectively. These constraints enable direct camera center estimation and allow more accurate localization with noisy feature detections and spurious matches than standard PnP solvers. The PILLO framework integrates these different geometric constraints in a flexible hybrid least squares solver, complemented by Procrustes-based orientation estimation adapted for lines. Synthetic, simulated, and real-world experiments using wide-area motion imagery (WAMI) demonstrate the efficacy of these methods, advanc-ing vision-based navigation by improving localization accuracy, reducing point-only reliance, and introducing novel PPL geometric frameworks, line-based orientation techniques, and validation tools for absolute camera pose estimation.
Table of Contents
DOI
PubMed ID
Degree
Ph. D.