The doctoral dissertations of the former Helsinki University of Technology (TKK) and Aalto University Schools of Technology (CHEM, ELEC, ENG, SCI) published in electronic format are available in the electronic publications archive of Aalto University - Aaltodoc.
|
|
|
Models and Methods for Bayesian Object Matching
Toni Tamminen
Dissertation for the degree of Doctor of Science in Technology to be presented with due permission of the Department of Electrical and Communications Engineering for public examination and debate in Auditorium S4 at Helsinki University of Technology (Espoo, Finland) on the 11th of November, 2005, at 12 noon.
Dissertation in PDF format (ISBN 951-22-7907-X) [3071 KB]
Dissertation is also available in print (ISBN 951-22-7906-1)
Abstract
This thesis is concerned with a central aspect of computer vision, the object matching problem. In object matching the aim is to detect and precisely localize instances of a known object class in a novel image. Factors complicating the problem include the internal variability of object classes and external factors such as rotation, occlusion, and scale changes. In this thesis, the problem is approached from the feature-based point of view, in which objects are considered to consist of certain pertinent features, which are then located in the perceived image.
The methodological framework applied in this thesis is probabilistic Bayesian inference. Bayesian inference is a branch of statistics which assigns a great role to the mathematical modeling of uncertainty. After describing the basics of Bayesian statistics the object matching problem problem is formulated as a Bayesian probability model and it is shown how certain necessary sampling algorithms can be applied to analyze the resulting probability distributions.
The Bayesian approach to the problem partitions it naturally into two submodels; a feature appearance model and an object shape model. In this thesis, feature appearance is modeled statistically via a type of bandpass filters known as Gabor filters, whereas two different shape models are presented: a simpler hierarchical model with uncorrelated feature location variations, and a full covariance model containing the interdependeces of the features. Furthermore, a novel model for the dynamics of object shape changes is introduced.
The most important contributions of this thesis are the proposed extensions to the basic matching model. It is demonstrated how it is very straightforward to adjust the Bayesian probability model when difficulties such as scale changes, occlusions and multiple object instances arise. The changes required to the sampling algorithms and their applicability to the changed conditions are also discussed.
The matching performance of the proposed system is tested with different datasets, and capabilities of the extended model in adverse conditions are demonstrated. The results indicate that the proposed model is a viable alternative to object matching, with performance equal or superior to existing approaches.
Keywords: statistical image analysis, object recognition, Monte Carlo simulation, Bayesian inference
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
© 2005 Helsinki University of Technology
Last update 2011-05-26