How the Video Eyetracker Works

Video eyetracking can be understood from simple geometry. When the eye moves relatively to the head it rotates within the orbit (eye socket). It rotates about three axes through the centre of the eye. Of most importance are the horizontal and vertical rotations, which are made to change the direction of gaze. These are large, up to ±50 degrees and can be exceedingly fast, up to 1000 degrees per second. The torsional (clockwise/counter-clockwise) rotation is small and for most applications of no interest.

The clear window at the front of the eye, the cornea, is smooth and kept constantly wet with tear fluid. When the eye is viewed in a dimly lit room on a bright sunny day, a bright image of a window can be seen on the cornea. This reflection is often referred to as the 'First Purkinje Image'. Although the eye is not a perfect sphere, the corneal surface is itself spherical and it is possible to determine its centre when illuminated by two known light sources. The pupil, being the dark black center of the eye, is easily distinguished from any reflections. With suitable optics, an infrared sensitive video camera can be used to observe the eye while remaining outside of the subject's field of view. By measuring the movement of the Purkinje reflections relative to the pupil, it is then possible to calculate head movement, eye rotation and consequently the direction of gaze. This is modelled by the equation:

The calibration procedure involves image measurements recorded at a set of known target positions presented on the stimulus display, which are then used to tune the parameters a - h, alpha - delta, Xoffset and Yoffset in the above equation. The eye tracker can then accurately monitor where the subject is looking from subsequent measures of pupil and Purkinje image centers while accommodating both eye and head movement.

Knowledge-Based, Adaptive Algorithm

Existing eye tracking systems, which are based conventional image segmentation algorithms, provide adequate performance most of the time in ideal conditions. We wanted a system that works all of the time and is extremely robust. The solution must operate in a wide range of illumination conditions and with any subject. To achieve this we employed knowledge-based image processing techniques developed for target identification in military applications. The recursive algorithm uses knowledge about the mechanics of the eye and previous history of eye position to give extremely robust tracking. This results in no 'dropped frames', i.e. occasions when the algorithm was unable to detect the features of the eye.

The algorithm is also adaptive. By incorporating prior knowledge of what the image should contain, the system is able to rapidly adjust itself to each individual. Conventional solutions all need to be tuned or adjusted for good tracking performance. This is often time consuming and in some cases the system is defeated by ambient illumination. We designed the software so that no adjustments are needed over a wide range of lighting conditions. The subject can be seated at the headrest, the camera aligned and focused and the calibration commenced in seconds.

The Toolbox incorporates an anatomically and physiologically plausible model of the head and eye. Rather than use a simple sphere to represent the eye, as most published devices do, we use a more complex model incorporating individual parameters for corneal and globe diameters. This enables us to increase the theoretical accuracy of measured eye rotation. By modelling possible sequences of eye positions, it is possible to resolve ambiguous measurements to give continuous measurement of eye rotation and gaze direction.

^ Back to top