Development of a robust active infrared-based eye tracking system

Abstract

Eye tracking has a number of useful applications ranging from monitoring a vehicle driver for possible signs of fatigue, providing an interface to enable severely disabled people to communicate with others, to a number of medical applications. Most eye tracking applications require a non-intrusive way of tracking the eyes, making a camera-based approach a natural choice. However, although significant progress has been made in recent years, modern eye tracking systems still have not overcome a number of challenges including eye occlusions, variable ambient lighting conditions and inter-subject variability. This thesis describes the complete design and implementation of a real-time camera-based eye tracker, which was developed mainly for indoor applications. The developed eye tracker relies on the so-called bright/dark pupil effect for both the eye detection and eye tracking phases. The bright/dark pupil effect was realised by the development of specialised hardware and near-infrared illumination, which were interfaced with a machine vision camera. For the eye detection phase the performance of three different types of classifiers, namely neurals networks, SVMs and AdaBoost were directly compared with each other on a dataset consisting of 17 individual subjects from different ethnic backgrounds. For the actual tracking of the eyes, a Kalman filter was combined with the mean-shift tracking algorithm. A PC application with a graphical user interface (GUI) was also developed to integrate the various aspects of the eye tracking system, which allows the user to easily configure and use the system. Experimental results have shown the eye detection phase to be very robust, whereas the eye tracking phase was also able to accurately track the eyes from frame-to-frame in real-time, given a few constraints. AFRIKAANS : Oogvolging het ’n beduidende aantal toepassings wat wissel van die deteksie van bestuurderuitputting, die voorsiening van ’n rekenaarintervlak vir ernstige fisies gestremde mense, tot ’n groot aantal mediese toepassings. Die meeste toepassings van oogvolging vereis ’n nie-indringende manier om die oë te volg, wat ’n kamera-gebaseerde benadering ’n natuurlike keuse maak. Alhoewel daar alreeds aansienlike vordering gemaak is in die afgelope jare, het moderne oogvolgingstelsels egter nogsteeds verskeie uitdagings nie oorkom nie, insluitende oog okklusies, veranderlike beligtingsomstandighede en variansies tussen gebruikers. Die verhandeling beskryf die volledige ontwerp en implementering van ’n kamera-gebaseerde oogvolgingsstelsel wat in reële tyd werk. Die ontwikkeling van die oogvolgingsstelsel maak staat op die sogenaamde helder/donker pupil effek vir beide die oogdeteksie en oogvolging fases. Die helder/donker pupil effek was moontlik gemaak deur die ontwikkeling van gespesialiseerde hardeware en naby-infrarooi illuminasie. Vir die oogdeteksie fase was die akkuraatheid van drie verskillende tipes klassifiseerders getoets en direk vergelyk, insluitende neurale netwerke, SVMs en AdaBoost. Die datastel waarmee die klassifiseerders getoets was, het bestaan uit 17 individuele toetskandidate van verskillende etniese groepe. Vir die oogvolgings fase was ’n Kalman filter gekombineer met die gemiddelde-verskuiwings algoritme. ’n Rekenaar program met ’n grafiese gebruikersintervlak was ontwikkel vir ’n persoonlike rekenaar, sodat al die verskillende aspekte van die oogvolgingsstelsel met gemak opgestel kon word. Eksperimentele resultate het getoon dat die oogdeteksie fase uiters akkuraat en robuust was, terwyl die oogvolgings fase ook hoogs akuraat die oë gevolg het, binne sekere beperkinge. Copyright