Experimental Results

The current prototype was implemented on a dual Pentium II 400MHz machine running Windows NT4, using a commercial PCI capture card compatible with Video for Windows. The eye tracker runs at frame-rate (30 frames per second), processing interlaced frames of resolution 640$\times$480$\times$8 bits. We have also achieved this frame rate with an IBM 300 MHz Pentium II Thinkpad 770X machine running Windows NT4, using a PCI capture card installed in an IBM dock III.

Eyeglasses and contact lens do not change the retro-reflectivity of the eye and unless the glasses or contacts are tinted with an IR blocking coating, they do not inhibit detection. Figure 7 shows the bright, dark, and difference images for a person with glasses. Observe in Figure 7c that spurious false pupil candidates are generated by the specular reflections from the eyeglasses, and these reflections can also block the dark pupil response under very particular head orientations. In such cases, if the head motion must be restricted, a slight change in the orientation of the glasses is enough to reestablish detection and gaze estimation. Pupil detection using only the dark or only the bright pupil images, as it is done by most commercial eye trackers for gaze estimation, would have a lot more spurious responses, which can be expected from images of the same kind shown in Figure 7.
 
 

Figure: (a) Bright and (b) dark pupil images with glasses. (c) Difference image after thresholding. The strong glints on the glasses can be avoided with a slight change in its orientation. 

 

The retro-reflectivity property of eyes is uncommon in man-made and natural objects resulting in pupils generally being the only objects appearing with high contrast between the two pupil images. Pupil detection is greatly facilitated by the enhanced signal-to-noise ratio, and the simple process of thresholding the difference between bright and dark pupil images is generally sufficient, as shown in Figure 2c. Our experience shows that most retro-reflectors we tested, used for example in running shoes, reflect light in a reasonable wide angle, so that they appear bright in both images and do not cause artifacts. Certain lamps, like table lamps and ceiling lamps, have reflectors that when pointed to the camera can cause artifacts. The one degree accuracy mentioned in Section 3 and the small head motion allowed by the system is comparable with commercial systems. The limitations on head motion is due to the simple motion model adopted, because the calibration changes with different head positions. We are currently working on more complex models to allow free head motion. Other applications such as real-time face tracking [3] and enhanced human-computer interaction [4,9] using the frame-rate pupil detector are described in other publications.

1999-07-13