Deep neural network ensemble architecture for eye movements classification

Abstract

Up to now, eye tracking technologies have been used for different purposes in various industries, from medical to gaming. Eye tracking methods could include predicting fixations, gaze mapping or movement classification. Recent advances in deep learning techniques provide possibilities for solving many computer vision tasks with high accuracy. Authors of this paper propose a novel deep learning based architecture for eye movement classification task. Proposed architecture is an ensemble approach which employs deep convolutional neural networks that run in parallel, for both eyes separately, for visual feature extractions along with recurrent layers for temporal information gathering. Dataset images for training and validation were gathered from standard web camera and pre-processed automatically using dedicated tools. Overall accuracy of developed classifier on the validation set was 92%. Proposed architecture uses relatively small networks which brings the possibility of real time usage (successfully tested on 15-20fps) on regular CPU. Classifier achieved overall accuracy of 88% on the real-time test, using standard laptop and web camera.