Alistair J. Bray
Submitted as thesis for the degree of Doctor of Philosophy. This thesis addresses the problem of recognising a known 3D object (modelled as a polyhedron) in a monochrome image, and tracking the object through a further sequence of images. The methods devised are efficient, and results obtained with natural sequence demonstrate that they are robust, even when the object is partly obscured. In order to find the object, and continue to track it, it is argued that any system must integrate "bottom-up" and "top-down" processing in order to exploit the benefits of both. In this system the object is located in the first frame using Model-Based Search. Image features are then tracked into the next frame using Optic Flow techniques, and their disparities are used to invert the Perspective Transform. The system can detect when object tracking is becoming inaccurate; it then recaptures its position through another model search that uses the reduced disparity information, and a "best guess" at position, to constrain the size of the search space. Int his manner the system is highly predictive when it has good information concerning object position, but resorts to non-redictive strategies when it does not. The system has been implemented as a computer program, and combines novel algorithms with established ones. A new method of matching 3D models agains 2D data is presented that would be suitable for implementation on a massively parallel architecture. An algorithm for tracking objects between closely spaced frames is also presented and consideration is given to modelling object motion in order to allow tighter constraints to e placed on object position. Finally, a detailed analysis is undertaken of the local constraints used when matching model against data; this results in the postulation of a "Shape Only" constraint set.
This paper is not available online