| Summary: | Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2007. === Includes bibliographical references (p. 137-142). === The tracking of visual phenomena is a problem of fundamental importance in computer vision. Tracks are used in many contexts, including object recognition, classification, camera calibration, and scene understanding. However, the use of such data is limited by the types of objects we are able to track and the environments in which we can track them. Objects whose shape or appearance can change in complex ways are difficult to track as it is difficult to represent or predict the appearance of such objects. Furthermore, other elements of the scene may interact with the tracked object, changing its appearance, or hiding part or all of it from view. In this thesis, we address the problem of tracking deformable, dynamically textured regions under challenging conditions involving visual clutter, distractions, and multiple and prolonged occlusion. We introduce a model of appearance capable of compactly representing regions undergoing nonuniform, nonrepeating changes to both its textured appearance and shape. We describe methods of maintaining such a model and show how it enables efficient and effective occlusion reasoning. By treating the visual appearance as a dynamically changing textured region, we show how such a model enables the tracking of groups of people. By tracking groups of people instead of each individual independently, we are able to track in environments where it would otherwise be difficult, or impossible. We demonstrate the utility of the model by tracking many regions under diverse conditions, including indoor and outdoor scenes, near-field and far-field camera positions, through occlusion and through complex interactions with other visual elements, and by tracking such varied phenomena as meteorological data, seismic imagery, and groups of people. === by Joshua Migdal. === Ph.D.
|
|---|
