Towards an universal person description framework for looking at people applications [Elektronische Ressource] / vorgelegt von Lutz Goldmann

Sujets

• Informatik

Informations

• Publié par: technische_universitat_berlin
• Publié le: 01 janvier 2010
• Nombre de lectures: 13
• Langue: English
• Poids de l'ouvrage: 29 Mo

Extrait

Towards an Universal
Person Description Framework
for Looking at People Applications
Von der Fakultat¨ IV - Elektrotechnik und Informatik
der Technischen Universitat¨ Berlin
zur Verleihung des akademischen Grades
Doktor der Ingenieurwissenschaften
- Dr.-Ing. -
genehmigte Dissertation
vorgelegt von
Dipl.-Ing. Lutz Goldmann
Promotionsausschuss
Vorsitzender: Prof. Dr.-Ing. Reinhold Orglmeister
1. Gutachter: Prof. Dr.-Ing. Thomas Sikora
2. Prof. Dr. Francesc Tarres´
Tag der wissenschaftlichen Aussprache: 07.09.2009
Berlin 2010
D 83To the light of my lifeAcknowledgement
This thesis emerged during my work as a research assistant at the Communication Systems
Group of the Technical University of Berlin. At this point I would like to thank all the people
that have supported me on this long way.
First of all, I would like to express my sincere gratitude to my supervisor Prof. Dr.-Ing.
Thomas Sikora for the chance to work on this challenging and interesting research topic. I
appreciate the freedom he gave me in my research and the fruitful advice I received from
him every time I was stuck in a problem. I also want to thank Prof. Dr. Francesc Tarres´ for
the detailed review of my thesis, the encouraging feedback and interesting discussions.
Special thanks go to my colleagues of the group for sharing time with me inside and
outside the office. In particular, I would like to thank my roommates Mustafa Karaman and
Amjad Samour for the helpful discussions, encouraging words and long evenings to meet a
particular deadline. Furthermore, I would like to appreciate all the organizational support
by Birgit Boldin who is the kind soul of the institute.
Within the scope of several European projects I had the chance to meet and exchange
ideas with numerous researchers working in the same research field. In particular, I would
like to express my gratitude to Toni Rama for the interesting discussions and the fruitful
collaboration that lead too several joint articles.
This work would not have been the same without the help and commitment of diploma
and master students under my supervision. Special thanks go to Ullrich Moenich and Lars
Thiele for their excellent work that is referenced within this thesis.
There are some moments in life when it is important to look back and remember where
we come from. This work would not have been possible without the constant support of
my dear parents and grandparents at home. At the same time I want to thank all my friends
that have continuously reminded me that there is more in life than work.
All the things we see and hence this whole work are based on the existence of light.
There is a light that has guided me along the way. This light is called Sylvia.
vviEidesstattliche Erklarung¨
Ich versichere an Eides statt, dass ich die von mir vorgelegte Dissertation selbststandig¨ ange-
fertigt und alle benutzten Quellen und Hilfsmittel vollstandig¨ angegeben habe.
Eine Anmeldung der Promotionsabsicht habe ich an keiner anderen Fakultat¨ oder Hoch-
schule beantragt.
viiviiiAbstract
During the last decade computers and the internet have become an important aspect in our
everyday life. We use this technology to communicate, study, work, shop, and entertain
ourselves. The vision of the future is to embed this computing technology into our home,
transportation and working environments. The ultimate goal is to develop intelligent ma-
chines that are aware of humans and can assist them if required. Therefore, visual data
needs to be analyzed with respect to humans which is often referred to as ”looking at peo-
ple”. So far the developments within this area have been largely influenced by the interests
and needs of specific applications (surveillance, biometrics, human computer interaction).
The objective of this dissertation is to move towards an universal framework for the vi-
sual analysis of humans, that describes humans at several levels including different body
parts (body, face, hands) and features (color, texture, shape and motion). Then, in analogy
to the human visual perception, an appropriate subset of the provided information can be
chosen, depending on environmental or application specific criteria. Within the scope of
such a framework, this work provides scientific contributions in several areas. For face de-
tection a novel component based face detection approach has been developed that combines
techniques from the statistical and structural pattern recognition domain for improved per-
formance especially in the presence of partial occlusions. It is not only able to detect faces
despite occlusions, but can also provide additional occlusion information to subsequent face
analysis steps. Based on that, existing appearance based face recognition approaches have
been extended through occlusion awareness by selecting the most reliable representation.
For appearance based body recognition both holistic and component based representations
and a large set of color and texture features have been considered to determine the optimal
description of a person’s clothes.
The developed framework has been used within several applications to prove its versa-
tility. The first original application, that has been developed, is an efficient system for the
audiovisual search of persons based on facial appearance and voice characteristics. A high
retrieval performance is achieved through the combination of multimodal fusion and rel-
evance feedback. For the second application, an original system for visual person search,
a different query paradigm was used. It provides an intuitive query interface through an
automatically derived human visual thesaurus that groups people based on their visual
similarity. Finally, the appearance based analysis was combined with motion based analysis
ixx
for a personalized human computer interface that detects, tracks and identifies humans and
interprets their gestures for the use in an intelligent cash machine scenario.
Although this dissertation focuses only on the appearance based description of face and
body, ideas and findings may also be applied to other channels (hands, limbs), features
(shape, motion) and tasks (tracking). Therefore it contributes to the gradual change from an
application specific view towards a universal framework for the visual analysis of humans,
which will enable machines to sense and react to humans in a more natural way.