On analysing deformable (moving) objects

Abstract

Performing a high level vision is usually based on features extracted at low and intermediate levels of the process of perception of a visual scene.

Segmentation and matching are instrumental tasks in producing comparable features in applications such as medical imaging, mining and oil extraction, gaming consoles, face, ear and gait biometrics, and etc.

The ultimate goal of this study is to develop a fully functional prior aided segmentation framework to extract deformable shapes over a sequence of frames. This thesis acknowledges the demand by these applications for a robust and flexible approach which is particularly designed to extract deformable timely shape sequences. It is also recognised that existing methods are either too general, and thus inaccurate, or too specific, thereby limited in usability.

This thesis suggests a learning model for gait synthesis with the ability to extrapolate to novel data. It involves computing comparable features from multiple sources. We show that these features which we formulate as continuous functions can be modelled by linear PCA.

This thesis also proposes a new fast and robust shape registration algorithm to match shapes from different sources in the proposed framework. This algorithm is based on linear orthogonal transformations and shape moments. The registration parameters are computed directly by analysing the signed distance functions of the shapes. This is in-line with the level sets based prior shape segmentation framework adopted here.

The segmentation is performed in a balanced framework between the data in the given images on one hand and the prior induced by the shape model and the registration algorithm proposed here on the other hand. This configuration ensures more control for the shape force over the overall shape geometry. Thus, favouring shapes familiar to the learned knowledge.

More information

Identifiers

ORCID for Mark S. Nixon: orcid.org/0000-0002-9174-5934

Catalogue record

Export record