ViewPointer: Lightweight CalibrationFree Eye Tracking for Ubiquitous Handsfree Deixis

1
ViewPointer Lightweight Calibration-FreeEye
Tracking for Ubiquitous Handsfree Deixis

• David Smith
• Roel Vertegaal, Changuk Sohn
• Human Media Lab
• Queens University

2
ViewPointer

• Eye tracking input device suitable for ubiquitous
  computing scenarios
• Wearable (lightweight)
• Calibration-free
• Little configuration
• Performs deixis towards real world objects rather
  than location

3
Eye Tracking

• The most common form of eye tracking today is the
  desktop corneal reflection eye tracker

Image courtesy http//www.polhemus.com
4
Desktop eye trackers

• Video-based
• Compares the position of the pupil to a
  reflection on the cornea from an IR light source
  near the camera

5
The Desktop Eye Tracker

• Problems with the desktop eye tracker
• Expensive
• Complicated setup
• Correlate eye position to screen coordinates via
  a calibration step

6
Wearable Eye Trackers

• Reports gaze coordinates within a scene from an
  onboard camera
• Camera must be stationary relative to the head
• Deixis towards objects requires object recognition

ASL Model H6 wearable eye trackerImage Courtesy
http//www.a-s-l.com
7
The Eye Contact Sensor

• Developed at the Human Media Lab at Queens
  University
• Reports deixis
• what the user is looking at rather than where
• Signals eye contact when a reflection is in the
  center of the users pupil
• Calibration-free

Image courtesy www.xuuk.com
8
The Eye Contact Sensor

• Problems with the eye contact sensor
• Does not identify the user
• Interfere with each other when placed close
  together
• Expensive with large deployments
• N attentive devices means N eye contact sensors
• Requires power
• Have to have a wire or battery

9
ViewPointer

• ViewPointer is a wearable eye contact sensor
• User identity is intrinsic
• Objects are tagged with an IR marker with an
  encoded ID
• Eye contact is determined when user looks at a
  marker
• Long range (currently up to 3 meters)
• Low-cost camera close to the eye
• One ViewPointer can be used to detect eye contact
  with N objects

10
ViewPointer

• User wears a small headset with a camera to
  monitor the eye

11
ViewPointer

• Headset requires little configuration
• The camera can be positioned anywhere it has a
  clear line of sight with the users eye
• Insensitive to camera movements
• Does not require any configuration of the
  pupil/corneal reflection detection algorithm
• Insensitive to changes in lighting

12
ViewPointer

• ViewPointer Markers
• Small and wireless
• Include onboard power from a small lithium battery

13
How it works

• The surface of the cornea is roughly spherical,
  with the pupil being recessed toward the center
  of that sphere
• Called the pseudophacic anterior chamber depth

14
How it works

• When looking at the eye from the headset camera,
  the cornea serves as a lens that refracts the
  image of the pupil
• Such that the pupil catches light from extreme
  angles
• The pupil still appears as a ellipse, even from
  extreme angles
• Marker reflections appear at the half angle
  between the visual axis and the camera
• The center of this ellipse continues to coincide
  with this reflection

15
How it works
The camera image when the user is looking at a
marker
16
ViewPointer

• Identifying markers
• The ID used to identify a marker is a binary
  signal transmitted by the LEDs.
• The headset uses the markers reflection from the
  cornea to acquire the ID.
• The data rate transmitted by the markers
  asymptotes at half the frame rate of the camera
  (Nyquist theorem)
• A trade-off exists between the time a user must
  look at a marker to identify it and the number of
  unique IDs
• Current implementation transmits at 14 bps
• This means 64 unique IDs with a 1 second fixation

17
Performance

• Current implementation is accurate to within 6
  degrees of visual angle
• 10 cm at 1 m
• When markers are close together they blend into a
  single reflection and the ID transmission fails
• Works at a range of up to 3 m
• Range can be increased by increasing the
  brightness of the LEDs

18
ViewPointer

• The headset also contains a microphone and
  speaker
• The eye tracker can provide context and focus
  selection for a speech interface through the
  headset.
• ViewPointer fits with Guiards kinematic chain
• Experimental evidence has shown that the eyes
  provide context for the hands
• In this sense, we can say the eyes provide a base
  link in the kinematic chain before the
  non-dominant hand

19
Design principles

• We derive a few design principles from this
  addition to the kinematic chain
• Eye contact sensing objects provide context to
  action
• Design for input output
• Avoid direct action upon eye contact
• Design for deixis

20
Scenario of use

• A poster in a store window with an IR marker
  behind it
• ViewPointer detects a URL when the user looks at
  the poster
• A PDA automatically locates interactive content
  about the poster

21
Future directions

• Increased frame rate
• An increase in frame rate could facilitate much
  larger bit lengths for marker IDs
• Current frame rate is 28 fps allowing for 64
  unique markers with a one second fixation
• An increase to a bit rate of 100 fps could allow
  for 17 million unique IDs with a one second
  fixation

22
Future directions

• Paper markers
• Markers could be reflective and lit by a
  high-powered LED on the headset
• Could use IR reflective inks to add eye contact
  sensing to paper such as magazine pages or
  instruction manuals
• However a new encoding scheme would be needed
• Spatial encoding similar to a barcode or AR glyph

23
Thank you

• Another UIST 2005 presentation from the Human
  Media Lab
• Connor Dickie eyeLook Using Attention to
  Facilitate Mobile Media Consumption
• Paper session 4 Tuesday, 830 945