DEEP-LEARNING LOCAL FEATURES FOR PHOTOGRAMMETRIC APPLICATIONS AND VISUAL POSITIONING

Image matching is a crucial step in various photogrammetric and computer vision applications. Traditional handcrafted matching algorithms automatically identify correspondences between image pairs, correspondences that can be triangulated to create a sparse map of the environment, while estimating the pose of the cameras and, in some cases, the internal camera parameters and distorsions. However, these algorithms often suffer from performance degradation under challenging conditions, such as significant changes in illumination, time or viewing angle. To overcome these limitations, several deep-learning (DL) based matching algorithms have been developed. Trained on large datasets, they have demonstrated considerable potential across a range of computer vision tasks. However, their application in photogrammetric and positioning applications remains underexplored. With image datasets captured within a short time frame and with high redundancy and gradual changes in baselines and viewing angles, DL-based approaches typically provide an accuracy that is similar to traditional handcrafted approaches, such SIFT, which is normally considered the reference approach. However, the radiometric invariance of DL-based methods can be leveraged to address situations where SIFT is prone to failure, thus paving the way for new applications in both photogrammetric and positioning domains. The aim of this work is to explore these new DL-based methods and propose new workflows validated under a metrological approach for four application scenarios: 1) multi-modal matching between infrared and RGB images; 2) multi-modal matching between RGB images and LiDAR point clouds; 3) matching off-track satellite images for 3D reconstruction and change detection; 4) precise positioning in GNSS-denied environments. In the PhD work, new workflows and processing methods have been implemented and validated taking into consideration the accuracy and metric aspects of the Geomatic community.

Grantor dc:publisher

Università degli studi di Trento

Year dc:date

2025

Author dc:creator

• Morelli, Luca

Contributors dc:contributor

• Vitti, Alfonso

• Image matching, Deep learning, Photogrammetry, Visual Odometry, Simultaneous Localization and Mapping

Statement dc:rights

• info:eu-repo/semantics/openAccess
• license:Creative commons
• license uri:http://creativecommons.org/licenses/by-nc/4.0/

Language dc:language

eng

Identifier

http://dx.doi.org/10.15168/11572_447731
10.15168/11572_447731

OAI identifier oai:identifier

oai:iris.unitn.it:11572/447731