2020 Volume 13 Pages 7-15
Advancements in technology have recently made it possible to obtain various types of biometric information from humans, enabling studies on estimation of human conditions in medicine, automobile safety, marketing, and other areas. These studies have particularly pointed to eye movement as an effective indicator of human conditions, and research on its applications is actively being pursued. The devices now widely used for measuring eye movements are based on the video-oculography (VOG) method, wherein the direction of gaze is estimated by processing eye images obtained through a camera. Applying convolutional neural networks (ConvNet) to the processing of eye images has been shown to enable accurate and robust gaze estimation. Conventional image processing, however, is premised on execution using a personal computer, making it difficult to carry out real-time gaze estimation using ConvNet, which involves the use of a large number of parameters, in a small arithmetic unit. Also, detecting eye movement events, such as blinking and saccadic movements, from the inferred gaze direction sequence for particular purposes requires the use of a separate algorithm. We therefore propose a new eye image processing method that batch-processes gaze estimation and event detection from end to end using an independently designed lightweight ConvNet. This paper discusses the structure of the proposed lightweight ConvNet, the methods for learning and evaluation used, and the proposed method's ability to simultaneously detect gaze direction and event occurrence using a smaller memory and at lower computational complexity than conventional methods.