Gaze recognition: Current research and development of an AI based prototype

Abstract

Gaze tracking systems are being researched for more than a hundred years. Yet, there is still more to be learned and improved upon. They are at this time mostly used in the medical and scientific fields. There has been recent research in less confined methods of usage for these systems. The least confined method of gaze tracking, having a camera placed independently from the observed, is probably the least researched method. If this method would achieve high degrees of accuracy even people who would act unusually while wearing an eye tracker could be have their gaze tracked easily. Therefore, this method is suitable for analyzing the gaze of the severely psychologically impaired under natural circumstances. In this master thesis currently existent methods of gaze tracking are going to be compared against one another. There will be a focus on gaze tracking methods utilizing cameras placed independently from the observed. Further several machine learning-based prototypes designed for this situation will be presented. The development of gaze tracking methods utilizing cameras placed independently from the observed is a complex issue. None of the in this thesis developed prototypes give decent results in their analysis of images. There are however other systems presented in this thesis where the best has a mean angular error of 17,6 (degree angle) on the chosen dataset.