Event Location: NSH 1305

Abstract: Complex visual tasks often require high resolution imagery of specific phenomena but run too slowly to process large fields of view. In this thesis, we explore a general technique for camera control designed to maximize the high resolution imagery captured that is useful to the given task while minimizing excess imagery that needs to be processed. Though this technique can be used for any visual task, this thesis explores those Visual tasks that operate in Unstructured environments and are computationally Expensive (VUE). In order to capture more relevant imagery, we introduce the concept of visual utility. Visual utility is a measure of the usefulness of each location for solving a given task. In our general control framework, we use estimates of future visual utility to direct th
e camera towards regions of the scene that are likely to contain useful information. In our analysis, we will measure the relationship between the accuracy of various visual utility estimates and the task performance in terms of both task accuracy and data bandwidth required. We will also explore three biologically inspired extensions to our general framework and their impact on task performance. These extensions are: foveated perception, microsaccades and visual attention. Finally, we will develop guidelines for identifying under what conditions using these techniques will be valuable in practice.

Committee:David Wettergreen, Chair

Martial Hebert

Yaser Sheikh

James J. Clark, McGill University