pixel value in a frame was less than 15 (out of 255) or changed by
more than 25 from frame to frame. Correlation peaks were rejected
if the next highest peak was within 15% of the highest peak value.
All subjects showed occasional tracking loss based on these criteria,
and these sections were not included in our statistical analysis.
Regions of tracking loss are indicated by the fine magenta line in
Figs. 4–6.
Ten seconds of video were recorded and saved for off line analysis.
Monocular (right eye) videos were taken with the splitting
mirror removed. The best quality of several videos from each subject
were analyzed and are presented here. Analysis proceeded
according to the method of Stevenson, Roorda, and Kumar (2010)
in which strips of each video frame were cross correlated to an
average frame constructed from the best frames of the video. Sections
of video with blinks or poor quality images were ignored for
the purposes of the statistical analysis. For these measures we
broke each video frame into 64 strips, resulting in an effective
eye tracking sample rate of 1920 Hz.
Conventional calibration of eye movement amplitude with the
TSLO is not required. The only requirement is that the angular subtense
of the scanned field is known. Eye motion is then simply calculated
from the pixel shifts in the retinal image. The exact field
size measurement in pixels per degree was made with a calibrated
model eye.
The accuracy of the resulting eye traces depends principally on
the quality of the image and on the selection of a reference frame
against which all others are compared. Artifacts due to the reference
frame show up as 30 Hz periodic motions and are removed by subtracting
the average motion across frames (Stevenson et al., 2010).
This has the drawback that any actual 30 Hz eye movements are
removed, but these have very small amplitude in normal fixation.
In order to estimate the noise level of the binocular analysis, we
also recorded monocular images of the right eye only by removing
the knife edge mirror. These videos were split in half and analyzed
with the same procedure as for binocular recordings. Our assumption
is that the motion in the left and right halves of a single eye
image are essentially identical and so any difference found can
be attributed to noise. This assumption is violated if the eye makes
significant amounts of torsion, because the rotation about the line
of sight produces vertical motion that is in opposite directions on
the left and right sides of the image. Torsion also produces an
apparent shear of the image relative to the reference frame, and
we analyzed the videos for this combined signature of torsion.