The technology for determining eye gaze direction on a monitor screen is considered by analyzing the images received from a video camera directed at a user without using additional equipment. We propose a combined convolutional neural network extracting high-level features of images with a neural network of long short-term memory that takes into account the temporal dynamics of oculomotor activity. To train the model, a representative database of video sequences with reference information concerning eye direction was collected. Experiments confirmed that accounting for temporal information increased the accuracy of recording the eye direction.
© 2018 Optical Society of AmericaPDF Article