Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)

Graduate Group

Computer and Information Science

First Advisor

Konstantinos Daniilidis


During the last decade, visual sensors have become ubiquitous. One or more cameras

can be found in devices ranging from smartphones to unmanned aerial vehicles and

autonomous cars. During the same time, we have witnessed the emergence of large

scale networks ranging from sensor networks to robotic swarms.

Assume multiple visual sensors perceive the same scene from different viewpoints. In

order to achieve consistent perception, the problem of correspondences between ob-

served features must be first solved. Then, it is often necessary to perform distributed

localization, i.e. to estimate the pose of each agent with respect to a global reference

frame. Having everything set in the same coordinate system and everything having

the same meaning for all agents, operation of the agents and interpretation of the

jointly observed scene become possible.

The questions we address in this thesis are the following: first, can a group of visual

sensors agree on what they see, in a decentralized fashion? This is the problem of

collaborative data association. Then, based on what they see, can the visual sensors

agree on where they are, in a decentralized fashion as well? This is the problem of

cooperative localization.

The contributions of this work are five-fold. We are the first to address the problem

of consistent multiway matching in a decentralized setting. Secondly, we propose

an efficient decentralized dynamical systems approach for computing any number of

smallest eigenvalues and the associated eigenvectors of a weighted graph with global

convergence guarantees with direct applications in group synchronization problems,

e.g. permutations or rotations synchronization. Thirdly, we propose a state-of-the

art framework for decentralized collaborative localization for mobile agents under

the presence of unknown cross-correlations by solving a minimax optimization prob-

lem to account for the missing information. Fourthly, we are the first to present an

approach to the 3-D rotation localization of a camera sensor network from relative

bearing measurements. Lastly, we focus on the case of a group of three visual sensors.

We propose a novel Riemannian geometric representation of the trifocal tensor which

relates projections of points and lines in three overlapping views. The aforemen-

tioned representation enables the use of the state-of-the-art optimization methods on

Riemannian manifolds and the use of robust averaging techniques for estimating the

trifocal tensor.

Files over 3MB may be slow to open. For best results, right-click and select "save as..."