Controlling the Focus of Perceptual Attention in Embodied Conversational Agents
Abstract
In this paper, the authors present a computational model of dynamic perceptual attention for virtual humans. The computational models of perceptual attention that they surveyed fell into one of two camps: top-down and bottom-up. Biologically inspired computational models typically focus on the bottom-up aspects of attention, while most virtual humans implement a top-down form of attention. Bottom-up attention models only consider the image information without taking into consideration saliency based on tasks or goals. As a result, the outcome of a purely bottom-up model will not consistently match the behavior of real humans in certain situations. Modeling perceptual attention as a purely top-down process, however, also is not sufficient for implementing a virtual human. A purely top-down model does not take into account the fact that virtual humans need to react to perceptual stimuli vying for attention. Top-down systems typically handle this in an ad hoc manner by encoding special rules to catch certain conditions in the environment. The problem with this approach is that it does not provide a principled way of integrating the ever-present bottom-up perceptual stimuli with top-down control of attention. The model presented here extends the prior model introduced by D. Traum and J. Rickel with perceptual resolution based on psychological theories of human perception. This model allows virtual humans to dynamically interact with objects and other individuals, balancing the demands of goal-directed behavior with those of attending to novel stimuli. The model has been implemented and tested with the MRE Project.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 01, 2005
- Accession Number
- ADA459162
Entities
People
- David R Traum
- Randall W. Hill Jr.
- Youngjun Kim
Organizations
- University of Southern California