Object identification -- the task of deciding that two observed objects are in fact one and the same object -- is a fundamental requirement for any intelligent system or situated agent that reasons about individuals. An approach to this problem has been developed in recent years based on a Bayesian framework that determines the probability for an object's expected appearance at subsequent observations, given its current appearance. The theory has been successfully applied to the task of recognizing cars observed by cameras at widely separated sites in a freeway network. This research will build upon and extend theoretical results involving object identification and data association, and will show how to apply these results to real-world problems.
This research has three primary thrusts. One aim is to integrate and generalize work from the data association community, which has focused on tracking independently moving objects whose state can be modeled by multivariate Gaussians, with work from the uncertainty in artificial intelligence community, which has produced a variety of compact state representations and associated inference algorithms. The goal is to present a unified framework with broader applicability than current theory. A second aim is to develop improved heuristic algorithms for approximating solutions to intractable problems that arise when the theory is applied to real-world domains. The goal is to develop new algorithms and demonstrate improved performance over existing ones. The third aim is to apply the theory to other real-world tasks, such as maintaining consistency and eliminating duplicate entries in databases. The goal is to illustrate the patterns of reasoning involved in moving from domain-independent theory to domain-specific problems. These research activities form part of a more general effort within the AI community to develop models and algorithms for reasoning in uncertain environments with noisy and imprecise sensors.
The PI's education plans will strengthen the computer science program at Middlebury College in particular, and at liberal arts institutions in general, in two ways: through innovative, inter-disciplinary approaches to presenting introductory computer science concepts so as to increase the interest among first and second year college students in this field at a time when many of them might otherwise stop taking courses in mathematics or science; and through early involvement of students in undergraduate research projects. The project's ultimate goal is to provide a model for others to follow in ways to integrate research and education activities at undergraduate liberal arts institutions.
