Comparative study of fixation identification algorithms
(Salvucci and Goldberg, 2002) suggests dispersion-threshold
method as a fast and robust mechanism for identification of
fixations. This method is also quite reliable in applications,
requiring real time data analysis, which is a critical aspect in
real-time photogrammetry applications.
Table 1 illustrates impact of camera resolution and sampling
rate on accuracy of fixation identification. The source data have
been acquired by eye-tracking systems with CCD size
1280x1024 pixels at 250 frames per second and spatially and
temporally down-sampled. Identification of fixations have been
implemented using Dispersion-Threshold Identification (I-DT)
with the fixation duration threshold of 250 ms and the
dispersion threshold of 25 pixels, constant for all resolutions
and sample rates in our experiments.
Sample Fixation RMISE : Neof
interval, | duration n Velocity : Fomples
E d 2 ms ; coord, pix/sec in
pix fixations
Sample rate: 250 frames per second
1280x 1024 4 324,6 2,85 198,7 81,1
960x768 4 328,0 2,86 215,6 82,0
728x582 4 326,5 2,99 227,7 81,6
500x576 4 327,4 3,00 222,2 81,8
360x288 4 340,2 3,32 233,1 85,1
CCD size: 1280x1024
250 fps 4 324,6 2,85 198,7 81,1
125 fps 8 334,7 3,00 166,8 41,8
050 fps 20 339,9 3,64 113,5 17,0
025 fps 40 350,5 5.51 75,9 8,8
010 fps 100 331.3 15,48 22,8 3,6
CCD size and sample rate
o 3206 | 2,85 | 1987 | s
NC 8 3308. [- 304-1752 | 414
m | 20 | 3348 83,21 | 1212 [16,7
EE 29951 338 [150004237 3,6
Table 1. Impact of camera resolution and sampling rate on
accuracy of fixation identification
Among many parameters which could be extracted and studied
from eye movement protocols, coordinates and duration of
fixations have major interests for accurate measurements of
objects on static images. Table 1 represents sample interval
(i.e. video frames acquisition interval), duration of identified
fixation (i.e. time of relatively stable position of an eye), errors
in determining of coordinates of fixation, mean velocity of eye
tremors and drifts during fixations and average number of
samples in fixations.
Table 1 illustrates variation of the above parameters depending
on camera resolution, sample rate (frequency) and their
combinations. As the coordinates of objects have the highest
priority in metric technologies, the corresponding parameter
(RMSE of fixations’ position) has drawn our major attention.
The first part of the table illustrates the idea that temporal
resolution has major priority over spatial resolution in eye-
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
tracking systems, aimed on measurements of static objects. The
results clearly show that change of spatial resolution from
1280x1024 pixels down to 360x288 pixels does not entail
coarsening of fixation coordinates on such a dramatic scale:
RMSE in coordinates change from 2.85 pixels to 3.32 pixels
only.
The second part of the table (opposite), illustrates significant
loss of accuracy in coordinates when changing sampling rate at
fixed spatial resolution (from 2.85 pixels at 250 frames per
second to 15.48 pixels at 10 frames per second). The third part
of the table demonstrates impact of combined changes — both
spatial and temporal resolution, outlining the balance between
spatial resolution and sampling rate of eye-tracker's CCD and
frame grabber.
The data, provided in Table I, are averaged from series of
experiments with eye-tracking protocols. Figures 4 and 5
illustrate particular data analysis — matching of fixations,
identified from different source data.
DP memset bettie EN
* |
4 i
900 - + |
¥ |
850 4 5 Sulit,
- : +
5
ann T T T i
450 S00 550 600 650
+1280 640 © mismatch 1250 « mismatch Q640
Figure 3. Fixation matching results; blue - fixations detected at
1280x1024/250fps, green - fixations detected at
640x480/50fps; red — fixation mismatches
500 A, AERO
500 + |
400 4 ne | | | |
W H A My . A 1 | À
an VÀ wx. LA |o
300 W Uu M { Li AM A | | \ i j | AN | X
200 + Lib Tl |
100] 4 LM HT
nd T LET TERT T
1 21 41 61
Duration Deviation Position Deviation
Figure 4. Fixation matching results 1280x1024/250fps vs.
640x480/50fps: blue - deviations in duration of
fixations (ms); purple — deviations in coordinates of
fixations (pixels)
6. CONCLUSION AND OUTLOOK: EYE-TRACKING
IN AUGMENTED PHOTOGRAMMETRY
Spatial and temporal data about eye movements, compiled
while observing geospatial imagery, bear the meaningful
Intern
——
inforn
augm
Fixati
interp
being
tracki
new t
abilit
highly
Regis
explo
throu
mind
whicl
image
about
of sc
surro
level
inter]
scene
devel
know
huma
imag:
and r
geosp
revol
know
Azun
Dyna
SIGC
Grap
Carp
Corn
Cricl
com
Nati
Deer
Proc
In Ce
Ditcl
Rese
Doer
Visu
Perf
Gold
proci
1955
Jaim
S.-F.
Auto