The use of an avatar-enabled applications have been rapidly growing over the last decade as they promise more natural computer interaction with advanced technologies in various domains. Furthermore, recent research efforts towards natural and affective avatar capabilities have become more prevalent in the field. However, developing such an application still remains a very difficult and time-consuming task. This is mainly because a believable avatar model intuitively aims to mimic a real human, including realistic appearance and a wide spectrum of complex behaviors. Even though we have to approach this problem as a whole, most previous studies have focused on only a small part of the model due to its complexity.

Recently, state-of-the-art research in computer graphics has presented realistic renderings of the human face with a programmable commodity graphics processing unit. However, these results are mostly focused on a static model or non-interactive character animations, whereas avatar researchers tend to use very limited visualization capabilities. Merging these two research efforts will provide a better chance to surpass Mori’s Uncanny Valley and achieve a more natural experience interacting with an avatar. Another issue found in the literature is the use of two distinct methodologies in modeling human behavior: rule-based model and data-driven model.

The model of the rule-based system offers highly coherent behavior based on psychological theories, but it lacks in subconscious or unconscious behavior. The data-driven method relies on a large amount of rich data to extract the uncertain nature of human beings to mimic it on an avatar model, however it lacks the depth of knowledge required to understand progressive causalities in our face-to-face communications.

This thesis presents a high quality visualization method and a behavior-modeling framework that can enhance the user experience with an autonomous avatar and eventually achieve the goal of an avatar as a natural lifelike computer interface. A hybrid behavior modeling technique sets the middle ground to combine both rule-based and data-driven models. Highly realistic avatar visuals with an emotionally expressive behavior model offer better congruency and naturalness at the same time to increase avatar believability. It promotes simple design and development for an affective avatar-enabled application to overcome current limitations. This thesis contributes to the avatar research domain in computer science by promoting the synergy between widely available technologies to create a natural and believable avatar control framework. A user study is conducted to evaluate the perceived naturalness of the framework’s behavior model within an autonomous expressive storytelling context. As we establish a better tractable model for an avatar as a more natural alternative computer interface, it will broaden the possibilities of our computational needs where we suffer from limited resources.

PDF

Lee, Sangyoon, Supervised Hybrid Expression Control Framework for a Lifelike Affective Avatar, Thesis submitted as partial fulfillment of the requirements for the degree of Doctor of Philosophy in Computer Science, Graduate College of the University of Illinois at Chicago, August 15th, 2013.