Dwell Selection with ML-based Intent Prediction Using Only Gaze Data
Toshiya Isomoto, Shota Yamanaka, Buntarou Shizuki
Abstract
We developed a dwell selection system with ML-based prediction of a user's intent to select. Because a user perceives visual information through the eyes, precise prediction of a user's intent will be essential to the establishment of gaze-based interaction. Our system first detects a dwell to roughly screen the user's intent to select and then predicts the intent by using an ML-based prediction model. We created the intent prediction model from the results of an experiment with five different gaze-only tasks representing everyday situations. The intent prediction model resulted in an overall area under the curve (AUC) of the receiver operator characteristic curve of 0.903. Moreover, it could perform independently of the user (AUC=0.898) and the eye-tracker (AUC=0.880). In a performance evaluation experiment with real interactive situations, our dwell selection method had both higher qualitative and quantitative performance than previously proposed dwell selection methods.