Date of Award

Spring 1-1-2018

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

First Advisor

Xiao-Chuan Cai

Second Advisor

Elizabeth Jessup

Third Advisor

Jed Brown

Fourth Advisor

David Mulawa

Fifth Advisor

Brett Bader

Abstract

Bundle adjustment is the process of minimizing errors in camera and three-dimensional structure parameters. The bundle adjustment process is applicable to many areas of geospatial awareness, computer vision, robotics, and imaging, both terrestrial imaging and remote sensing. In the case of remote sensing and planetary imaging, current methods do not adequately address geographic areas consisting of both a large number of images and image observations. Other application domains focus on a single portion of the bundle adjustment process, the solution of a linear system, but ignore the computation of the coefficient matrix. In this thesis we propose a fully parallel approach to the bundle adjustment problem. This approach includes parallel computation of the required partial derivatives, which also addresses load-imbalance inherent in the problem, a parallel solution to the required linear system, and novel parallel preconditioning techniques for this system. Additionally we investigate the use of a relational database to enable fast recomputation due to image addition or removal.

As other research has shown, preconditioning the linear system present in the bundle adjustment problem is critical. We present two novel, parallel preconditioners, also based on the geographic information of the input data. These preconditioners are specific to the planetary imaging application domain and address the specific matrix structure that arises in this area.

We show that the parallel derivative methods achieve a high level of parallel efficiency and work well with the usage of a parallel, distributed memory, linear solver. The demonstrated preconditioners make a tangible reduction in the number of required solver iterations. Lastly, because these problems are solved many times for various applications, we present a database-backed method which stores derivative information, thereby easily allowing for projects to be re-run quickly, or modified slightly without a large recomputation cost. All of these elements result in a completely parallel bundle adjustment system capable of processing large geographic areas with millions of image observations.

Share

COinS