Author ORCID Identifier

https://orcid.org/0009-0009-2137-0995

Semester

Summer

Date of Graduation

2023

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Gianfranco Doretto

Committee Co-Chair

Thomas Devine

Committee Member

Donald Adjeroh

Committee Member

Yu Gu

Committee Member

Katerina Goseva-Popstojanova

Abstract

Scene representation and matching are crucial steps in a variety of tasks ranging from 3D reconstruction to virtual/augmented/mixed reality applications, to robotics, and others. While approaches exist that tackle these tasks, they mostly overlook the issue of efficiency in the scene representation, which is fundamental in resource-constrained systems and for increasing computing speed. Also, they normally assume the use of projective cameras, while performance on systems based on other camera geometries remains suboptimal. This dissertation contributes with a new efficient scene representation method that dramatically reduces the number of 3D points. The approach sets up an optimization problem for the automated selection of the most relevant points to retain. This leads to a constrained quadratic program, which is solved optimally with a newly introduced variant of the sequential minimal optimization method. In addition, a new initialization approach is introduced for the fast convergence of the method. Extensive experimentation on public benchmark datasets demonstrates that the approach produces a compressed scene representation quickly while delivering accurate pose estimates.

The dissertation also contributes with new methods for scene matching that go beyond the use of projective cameras. Alternative camera geometries, like fisheye cameras, produce images with very high distortion, making current image feature point detectors and descriptors less efficient, since designed for projective cameras. New methods based on deep learning are introduced to address this problem, where feature detectors and descriptors can overcome distortion effects and more effectively perform feature matching between pairs of fisheye images, and also between hybrid pairs of fisheye and perspective images. Due to the limited availability of fisheye-perspective image datasets, three datasets were collected for training and testing the methods. The results demonstrate an increase of the detection and matching rates which outperform the current state-of-the-art methods.

Download

Share

COinS