Eye tracker uncertainty analysis and modelling in real time

Techniques for tracking the eyes took place since several decades for different applications that range from military, to education, entertainment and clinics. The existing systems are in general of two categories: precise but intrusive or comfortable but less accurate. The idea of this work is to calibrate an eye tracker of the second category . In particular we have estimated the uncertainty both in nominal and in case of variable operating conditions. We took into consideration different influencing factors such as: head movement and rotation, eyes detected, target position on the screen, illumination and objects in front of the eyes. Results proved that the 2D uncertainty can be modelled as a circular confidence interval as far as there is no stable principal directions in both the systematic and the repeatability effects. This confidence region was also modelled as a function of the current working conditions. In this way we can obtain a value of the uncertainty that is a function of the operating condition estimated in real time opening the field to new applications that reconfigure the human machine interface as a function of the operating conditions. Examples can range from option buttons reshape, local zoom dynamically adjusted, speed optimization to regulate interface responsiveness, the possibility to take into ac count the uncertainty associated to a particular interaction. Furthermore, in the analysis of visual scanning patterns, the resulting Point of Regard maps would be associated with proper confidence levels thus allowing to draw accurate conclusions. We conducted an experimental campaign to estimate and validate the overall modelling procedure obtaining valid results in 86 % of the cases.