PPT - Andrew T. Duchowski

```Xiaoyong Ye
Franz Alexander Van Horenbeke
David Abbott
Index




Introduction
Background
Hardware
Software
 System Design
 Algorithm
 Pupil Localization
 Ellipse Fitting
 Calibration
 Homographic Mapping
 Experimental Results
 Future Work
Introduction
 A complete system able to track the user’s eye and map
the position of their pupil with the area at which they
are looking at in the scene in front of them
Background
 Wearable Eye-Tracking information
 Who has done previous work
 What they have used
 Recent Methods used with eye tracker
Objectives
 Hardware
 Wearable
 Low-Cost
 Light and Confortable
 Moveable eye-camera
 Software
 Real-Time
 Accurate
Hardware
 Two Cameras:
 Scene Camera
 Eye Camera
Hardware
Scene Camera
Eye Camera
 Captures the scene in
 Captures the eye
front of the user
 With 5 DOF with respect
System Design
Eye Image
Pupil
Localization
Scene Image
Calibration
Done?
No
Ellipse Fitting
Ellipse Center
Marker
Detection
Calculate
Homography
Mapping
Yes
Pupil Localization
 Automatic Threshold (Modified Otsu’s Method)
 Image Morphology(Dilation, Erosion)
 Connected Components Analysis(Find Pupil)
 Pupil Center Estimation
Histogram of an Eye Image
Background
Pupil
Graylevel
Threshold
Pupil Localization
Threshold
Erosion
Connect
Components
Pupil
Detection
Dilation
Fill holes
Ellipse Fitting
 1. Updating the pupil Center
 2. Need 5 points for Fitting Ellipse model
 3. RANSAC to deal with noisy points
Ellipse Fitting
Edge Image
 RANSAC method
Starburst
Algorithm
Feature
Points
RANSAC
Ellipse
Fitting
Calibration
 Relationship between Ellipse center to Scene Image
*
=
Scene Position
Homography
Pupil Center
Solving for homographies
X’ = Hx
 wx' 
a



wy'  d



 w 
 g
b
e
h
c  x
 
f
y
 
i   1 
 8 degrees of freedom in 3 x 3 matrix H, so at least n
points are sufficient to determine it
 Set up a system of linear equations:
Ah = 0
 where vector of unknowns h = [a,b,c,d,e,f,g,h]T
 Need at least 8 eqs, but the more the better…

Solve for h. solve using least-squares
= 8 pairs of
calibration method
1. Look at Scene Marker and Press corresponding number on keyboard,
2. Each marker press 2 to 3 times.
3. Randomly select 8 pairs of points to calculate Homography.(Repeatly)
3. Choose the best Homography matrix.
Mapping
(x1, y1)
(x2,
y2)
Experimental Results
 Frame rate 25/second
 Accurate Pupil Ellipse
 Mapping error is low( 13 pixels in 640*480 image)
Demo