Fitting parametric curve models to images using local self-adapting separation criteria [Elektronische Ressource] / Robert Hanek

Sujets

• Informatik

Informations

• Publié par: technische_universitat_munchen
• Publié le: 01 janvier 2004
• Nombre de lectures: 16
• Langue: English
• Poids de l'ouvrage: 5 Mo

Extrait

Institut fur¤ Informatik
der Technischen Universitat¤ Munchen¤
Fitting Parametric Curve Models to
Images Using Local Self-adapting
Separation Criteria
Dissertation
Robert Hanek

Institut fur¤ Informatik
der Technischen Universitat¤ Munchen¤
Fitting Parametric Curve Models to
Images Using Local Self-adapting
Separation Criteria
Robert Hanek
Vollstandiger¤ Abdruck der von der Fakultat¤ fur¤ Informatik der Technischen
Universitat¤ Munchen¤ zur Erlangung des akademischen Grades eines
Doktors der Naturwissenschaften (Dr. rer. nat.)
genehmigten Dissertation.
Vorsitzender: Univ.-Prof. Dr. Dr. h.c. Wilfried Brauer
Prufer¤ der Dissertation:
1. Univ.-Prof. Dr. Bernd Radig
2. Univ.-Prof. Nassir Navab, Ph.D.
Die Dissertation wurde am 28.11.2003 bei der Technischen Universitat¤ Munchen¤
¤ ¤eingereicht und durch die Fakultat fur Informatik am 02.07.2004 angenommen.Abstract
The task of tting parametric curve models to boundaries of perceptually meaningful image
regions is a key problem in computer vision with numerous applications, such as image seg-
mentation, pose estimation, 3-D reconstruction, and object tracking. In this thesis, we propose
the Contracting Curve Density (CCD) algorithm and the CCD tracker as solutions to this pro-
blem. The CCD algorithm solves the curve- tting problem for a single image whereas the CCD
tracker solves it for a sequence of images.
The CCD algorithm extends the state-of-the-art in two important ways. First, it applies a
novel likelihood function for the assessment of a t between the curve model and the image
data. This likelihood function can cope with highly inhomogeneous image regions because it
is formulated in terms of local image statistics that are learned on the y from the vicinity
of the expected curve. Second, the CCD algorithm employs blurred curve models as ef cient
means for iteratively optimizing the posterior density over possible model parameters. Blurred
curve models enable the algorithm to trade-off two con icting objectives, namely a large area
of convergence and a high accuracy.
The CCD tracker is a fast variant of the CCD algorithm. It achieves a low runtime, even
for high-resolution images, by focusing on a small set of carefully selected pixels. In each
iteration step, the tracker takes only such pixels into account that are likely to further reduce the
uncertainty of the curve. Moreover, the CCD tracker exploits statistical dependencies between
successive images, which also improves its robustness. We show how this can be achieved
without substantially increasing the runtime.
In extensive experimental investigations, we demonstrate that the CCD approach outper-
forms other state-of-the-art methods in terms of accuracy, robustness, and runtime. The CCD
algorithm and the CCD tracker achieve sub-pixel accuracy and robustness even in the presence
of strong texture, shading, clutter, partial occlusion, poor contrast, and substantial changes of il-
lumination. We present results for different curve- tting problems such as image segmentation,
3-D pose estimation, and object tracking.
iiiZusammenfassung
Das Anpassen parametrischer Kurvenmodelle an die Grenzen perzeptuell relevanter Bildregio-
nen ist ein Kernproblem der Bildverarbeitung. Es tritt in zahlreichen wichtigen Anwendungen
wie z.B. Bildsegmentierung, Lageschatzung,¤ 3-D Rekonstruktion und Objektverfolgung auf. In
dieser Dissertation werden der Contracting Curve Density (CCD) Algorithmus und der CCD
Tracker als Losungen¤ fur¤ dieses Problem vorgeschlagen. Der CCD lost¤ das An-
passungsproblem fur¤ ein einzelnes Bild, der CCD Tracker hingegen fur¤ eine Bildsequenz.
Der CCD Algorithmus erweitert den derzeitigen Stand der Forschung in zweierlei Hin-
¤sicht. Erstens verwendet er eine neuartige Likelihood-Funktion fur¤ die Bewertung der Uberein-
stimmung zwischen dem Kurvenmodell und den Bilddaten. Die Likelihood-Funktion ist selbst
fur¤ inhomogene Bildregionen geeignet, da sie auf lokalen Statistiken basiert, welche iterativ
von der Umgebung der Kurve gelernt werden. Zweitens verwendet der CCD Algorithmus un-
scharfe Kurvenmodelle als effektives Mittel zur iterativen Optimierung der a-posteriori-Dichte.
Unscharfe Kurvenmodelle erlauben einen Ausgleich zwischen zwei gegensatzlichen¤ Zielen,
namlich¤ einem gro en Konvergenzbereich und einer hohen Genauigkeit.
Der CCD Tracker ist eine schnelle Variante des CCD Algorithmus. Selbst fur¤ Bilder mit
hoher Au osung¤ erzielt er eine geringe Rechenzeit, indem er sich auf eine kleine Menge spe-
ziell ausgewahlter¤ Pixel fokussiert. In jedem Iterationsschritt verwendet er nur solche Pixel,
die hochstw¤ ahrscheinlich die Unsicherheit der Kurve weiter reduzieren. Daruber¤ hinaus nutzt
der Tracker statistische Abhangigk¤ eiten zwischen aufeinanderfolgenden Bildern. Dies fuhrt¤ zu
einer zusatzlichen¤ Steigerung der Robustheit, ohne die Rechenzeit substanziell zu erhohen.¤
Umfangreiche experimentelle Untersuchungen zeigen, dass der CCD Ansatz anderen An-
satzen¤ sowohl in der Genauigkeit, der Robustheit als auch der Rechenzeit uberle¤ gen ist. Der
CCD Algorithmus und der CCD Tracker erzielen Subpixel-Genauigkeit und Robustheit selbst
bei starken Texturen, Schattierungen, Teilverdeckungen, geringem Kontrast und erheblichen
Beleuchtungsanderungen.¤ In dieser Arbeit werden Ergebnisse fur¤ verschiedene Anpassungs-
probleme vorgestellt, z.B. Bildsegmentierung, 3-D Lageschatzung¤ und Objektverfolgung.
iiiivAcknowledgments
This dissertation would not have been possible without the assistance of several people. First
of all, I would like to thank my thesis advisor, Prof. Dr. Bernd Radig, for giving me the oppor-
tunity to work on this dissertation. I am particularly grateful for his enthusiasm in supporting
Thorsten’s and my idea of founding our own company.
I would like to express my thanks to Prof. Dr. Nassir Navab, the second reporter on this
dissertation. His passion for science was one of the sources of my motivation.
Special thanks go to Michael Beetz. His comments, encouragement, and constructive crit-
icism helped me many times, especially in writing publications and preparing presentations.
I would especially like to thank Thorsten Schmitt and Sebastian Buck for being great colleagues
and close friends. Working with you and the other members of the AGILO robot soccer team
was fun even during stressful periods such as the preparations for World Cups and other events.
I would also like to thank my other colleagues at the Image Understanding and Knowledge-
Based Systems Group at Munich Technical University. Thank you for your assistance and
many amazing lunch and coffee breaks. I am indebted to Fabian Schwarzer, Bea Horvath,
Andreas Hofhauser, and Gillian McCann for proofreading this thesis.
Thanks go to Michael Isard, Andrew Blake, and the other members of the Oxford Visual
Tracking Group for providing the code of their tracking library. I am grateful to Juri Platonov
for his assistance in conducting the experiments with the condensation and the Kalman tracker.
I would like to thank Carsten Steger for helpful discussions and for providing the DLL of the
color edge detector. Jianbo Shi and Jitendra Malik at the University of California at Berkeley
provided the executable of their Normalized Cuts algorithm. Thanks go to Lothar Hermes at the
University of Bonn for running the PDC algorithm on the test data, see Figure 2.1. Furthermore,
I am grateful to Christoph Hansen for providing the PDM depicted in Figure 5.13.
Last, but certainly not least, I would like to thank my girl-friend, Wiebke Bracht, who
supported me with her love, encouragement, and assistance in many ways.
vvi