Tesis y Tesistas 2020 - Postgrado - Fac. de Informática - UNLP
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
DOCTORADO EN
CIENCIAS INFORMÁTICAS
Dr. Nahuel Mangiarua
nmangiarua@unlam.edu.ar
Advisor
Dr. Jorge Ierache
Codirector
Dr. María José Abásolo
Thesis defense date
December 18, 2020
SEDICI
http://sedici.unlp.edu.ar/handle/10915/111301
Scalable integration of image and
face based augmented reality
Keywords: Augmented Reality; Face Recognition; Scalability; Integration Architecture; Biometric Inference;
Approximate Nearest Neighbor Search
Motivation
Both images based and face based AR are used by several
systems, applications or frameworks in various fields of
application. However, there are currently no frameworks
whose architectures integrate the ability to simultaneously
recognize images and faces in a scalable way, that is, with
a high number of augmentation objectives. Furthermore, no
framework integrates the ability to make biometric inference
of information from human faces.
The main objective of this thesis work is to design a scalable
AR architecture based on the monocular visual recognition
of images and human faces, including biometric inference
capabilities, without reliance on external services during its
exploitation stage.
The following are proposed as particular objectives:
• Establish the processes and their necessary steps to carry
out image augmentation, face detection and recognition,
and inference of biometric information.
• Comparatively analyze the theoretical computational
complexity and the empirical processing load of each step
of the AR processes, considering different variations of
algorithms available for the search and description of POIs.
• Design an architecture that integrates the processes
described, contemplating parallel and / or asynchronous
execution, identifying the step or steps that result in the
main bottleneck with respect to scalability.
• Comparatively analyze the speed and precision of the applicable
algorithms to alleviate or solve the detected bottlenecks.
• Design evaluation criteria and sets of test data for the
algorithms applicable to bottlenecks that are representative
of the proposed exploitation domain.
• Incorporate into the design an open integration mechanism
that facilitates the future addition of additional biometric
inference algorithms.
It is planned to design an architecture and develop a
prototype that integrates AR based on arbitrary images, facial
recognition and biometric data inference in a scalable way.
The computational complexity will be analyzed to identify
bottlenecks, analyzing and comparing existing algorithms in
order to solve scalability limitations, avoiding dependencies
on external systems during the exploitation phase.
While it does not seek to compete with existing systems in
terms of the refinement and quality that they have achieved
with years of continuous development, it is intended to
demonstrate that an integration of the proposed technologies
is possible, while surpassing their scalability limits.
The study of the particularities of each POI search, POI
description, face detection, face recognition and biometric
inference algorithms used will be outside the scope of this
thesis. Only their relative fitness will be considered in terms of
the minimum processing load required to achieve the desirable
RA quality under physical conditions (lighting, occlusion, etc.)
expected in a partially controlled operating environment.
Thesis contributions
This thesis work proposes a scalable architecture that
integrates AR based on arbitrary images with the detection
36