Date of Award


Level of Access Assigned by Author

Open-Access Dissertation

Degree Name

Doctor of Philosophy (PhD)


Spatial Information Science and Engineering


Peggy Agouris

Second Committee Member

M. Kate Beard-Tisdale

Third Committee Member

Max J. Egenhofer


In today's world of vast information availability users often confront large unorganized amounts of data with limited tools for managing them. Motion imagery datasets have become increasingly popular means for exposing and disseminating information. Commonly, moving objects are of primary interest in modeling such datasets. Users may require different levels of detail mainly for visualization and further processing purposes according to the application at hand. In this thesis we exploit the geometric attributes of objects for dataset summarization by using a series of image processing and neural network tools. In order to form data summaries we select representative time instances through the segmentation of an object's spatio-temporal trajectory lines. High movement variation instances are selected through a new hybrid self-organizing map (SOM) technique to describe a single spatio-temporal trajectory. Multiple objects move in diverse yet classifiable patterns. In order to group corresponding trajectories we utilize an abstraction mechanism that investigates a vague moving relevance between the data in space and time. Thus, we introduce the spatio-temporal neighborhood unit as a variable generalization surface. By altering the unit's dimensions, scaled generalization is accomplished. Common complications in tracking applications that include occlusion, noise, information gaps and unconnected segments of data sequences are addressed through the hybrid-SOM analysis. Nevertheless, entangled data sequences where no information on which data entry belongs to each corresponding trajectory are frequently evident. A multidimensional classification technique that combines geometric and backpropagation neural network implementation is used to distinguish between trajectory data. Further more, modeling and summarization of two-dimensional phenomena evolving in time brings forward the novel concept of spatio-temporal helixes as compact event representations. The phenomena models are comprised of SOM movement nodes (spines) and cardinality shape-change descriptors (prongs). While we focus on the analysis of MI datasets, the framework can be generalized to function with other types of spatio-temporal datasets. Multiple scale generalization is allowed in a dynamic significance-based scale rather than a constant one. The constructed summaries are not just a visualization product but they support further processing for metadata creation, indexing, and querying. Experimentation, comparisons and error estimations for each technique support the analyses discussed.