A User Interface via Agent-Based Interaction Using Gaze Input

摘要

This paper proposes a gaze input interface based on the interaction between an agent and the user. The device is evaluated in real tasks. In the proposed system, multiple agents are defined as objects of selection and serve as the unit of interaction with the user. Each agent determines the degree of attention toward it from the gaze input from the user, in the sense of a probability estimated by a Bayesian network. After learning, the agent infers the probability that it is the subject of attention from the input to the system and performs the corresponding action. The action serves as feedback to the user, and by its activation gives the user the sensation that the user has selected the agent. Eventually, by activation of the agent, a preregistered shortcut command is started, and the wish of the user is carried out. In this study, the functions of event registration and performance are linked to each agent in Windows, and a mechanism is realized in which command selection, scrolling, and IME mode switching can be performed when the user selects the agent. The system is compared qualitatively and quantitatively with the conventional interface and the efficiency is found to be improved by approximately 10% for command selection. A nearly comparable efficiency is obtained for other tasks.