EYE GAZE FOR COMPUTER CONTROL

Poole E, Kumar Dinesh K

School of Electrical and Computer Engineering RMIT University GPO Box 2476 Melbourne,VIC 3001,Australia

Keywords: Electro-oculargram, Classification, and Computer Interface.

Abstract: This paper reports the results of experiments that were conducted with five subjects to determine the

reliability of the use of Electro-ocular gram (EOG) for controlling computers. Experiments included

vertical and horizontal eye motion. Consideration was given to identify a relationship between the angle of

the gaze and signal that could be applied to cover all test subjects and generate the required spatial control

signals.

The results obtained are encouraging. An assessment of the data has concluded that the

EOG can be successfully utilised for spatial control applications. The study recommends the

choice of bandwidth for the recording, inter and intra subject reliability and difference between

the vertical and horizontal movement control.

1 INTRODUCTION

There has been extensive research in the field of

developing control systems for improved man

machine interface to assist the entry of computers in

all gadgets and equipments. This has resulted in the

development of a variety of systems that have

applications in fields such as robotics, defense,

computer games and medicine. An important part of

these systems is the front-end module that accepts

the command from the human operator. Improved

and natural ‘Man Machine Interface’ (MMI) systems

can open the world of virtual reality, control of

machinery without physical contact and increased

independence from disabilities.

Human have an extremely well developed

vision control system that can track an object with

the help of dynamic corrections and feedback. This

is a result of movement and shape change of the eye.

Electro-oculargram (EOG) is the electrical

recording from the outside of the eye and

corresponds to the direction of the eye. There have

been attempts by researchers to use EOG as a means

for the user to control the machine. While the basic

system is extremely simple, the difficulty is in the

lack of reliability of such a system where there often

is a large variation with time.

It is important to realize the limitations,

potential errors and possible techniques to improve

the system for identifying applications of EOG

based computer control systems. There are several

possible sources of error that affect the accuracy of

the HCI using EOG signals. The angular

displacement between head and torso, physiological

disorders, an individual perception of gaze point,

and movement of the individual relative to a known

reference point are some of these. While a system

level approach to identify the system reliability

provides the users with simple answers, it is not

possible to identify individual sources of error and

determine ways of improvement. It is important to

determine the effects of error for static positions and

dynamic movement of an individual’s gaze with and

without visual feed back. With suitable controlled

experiments with pre-defined gaze targets for test

subjects to fix their gaze on, the errors can be

investigated.

This paper reports the work conducted to

experimentally determine the sources of error and

effect of these in EOG recordings for determining

the suitability of the use of EOG for MMI. The aim

was to determine the limitations of the use of EOG

for developing a system that would enable the

disabled to control machines and computers with the

help of their eye movement.

479

E P. and Dinesh K K. (2004).

EYE GAZE FOR COMPUTER CONTROL.

In Proceedings of the First International Conference on Informatics in Control, Automation and Robotics, pages 479-483

DOI: 10.5220/0001147004790483

 SciTePress

2 METHODOLOGY

Experiments were conducted where EOG data was

acquired from various test subjects while they

fixated their gaze on specific target points. Specific

features of the recorded EOG were extracted during

analysis to assist with the selection of a method to

classify the angular movement based on EOG and to

generate an output that is representative of the

required movement.

The test environment was modeled as a

series of target points in 3D space. Each target point

was on the surface of a sphere with the human

subject effectively at the centre of the sphere. The

spatial orientation of the subject’s eyes, as they

fixate on a specific target point, could then be

determined using the detected EOG relative to the

origin of the sphere.

2.1 Apparatus

Nine (9) target points were positioned on the

circumference of a circle of known radius, as shown

in Figure 2 using firm vertical supports. A reference

target point was positioned directly forward of the

test subject’s eyes, at a specified eye height,

allowing a known starting point for all EOG

measurements in the horizontal and vertical planes.

Horizontal Vertical

Right

Test

Subject

Floor

Reference

Point

Left

Eye Height

Target Points

Figure 1: Configurations of Targets

For monitoring the horizontal movement, the targets

were located 15 degrees apart relative to the center

of the circle and located on the horizontal plane at a

specified eye height. The target points at the

extremes were located at 60 degrees either side of

the reference center target point. Vertical test points

were positioned on the reference pole to give five (5)

equally spaced sections between 45 degree above

eye level and the target point at floor level, below

the reference pole as shown schematically in Figure

2. To maintain commonality of test target angles,

test subjects were seated so that their eye height

from the floor was approximately the same.

2.2 Signal Acquisition

EOG signals were recorded from five subjects,

consisting of four males and one female. The eye

physiology of each subject was not considered for

this investigation and therefore any effects can only

be surmised.

Five disposable electrodes (Nessler Med –

Technin, Austria, universal Ag/AgCl Ref 1066)

were attached as shown in Figure 3 were put on each

of the subjects. Two electrodes were attached to the

outer canthi of each eye and formed a differential

electrode pair for horizontal movement. Two

electrodes were placed above and below the right

eye to form a differential electrode pair for vertical

movement. A fifth electrode was place at the centre

of the forehead as a common. Prior to attachment of

the electrodes, the skin was cleaned to remove

contaminates and improve signal reception.

The integrity of the acquisition system was

checked prior to recording using, sinusoidal signals

of known amplitude and frequency from a signal

generator. A short circuit was used to determine

offset levels. The signal generator and the short

circuit were interfaced to a, MR01B, analog to

digital acquisition card that was installed as part off

an AMLAB system (a Computer Based Instrument

Emulator).

Figure 2: Placement of Electrodes on Test Subjects

Amplifier gain and offset settings were determined

for each subjects after consideration of expected

full-scale levels recorded during preliminary testing.

During recording sessions, test subjects were

instructed to minimise the relative movement of the

torso and head, as well as other facial movements.

To determine the relationship of EOG with the

ICINCO 2004 - ROBOTICS AND AUTOMATION

480

angular displacement of the eyes, eye movements

relative to the reference point were conducted. To

ensure stable data, a minimum of 5 seconds of EOG

signal was recorded while fixating on the reference

target. At other target points, the EOG was recorded

for approximately 2.5 seconds.

The system was programmed to sample at

1000Hz with the anti-aliasing filter less than 500Hz.

The AC coupled amplifiers were programmed with a

time constant, T

of 15 seconds - an effective

0.011Hz high pass filter. The EOG data along with

the eye movement data were recorded onto a file.

3 ANALYSIS

The data files were read into MATLAB version 6 for

analysis. To determine the bandwidth of EOG

useful for HCI, the power spectral densities of

several EOG traces from each subject were analysed

to determine frequency content. The EOG data was

initially decimated and passed through a 12

order

5Hz Butterworth LP filter and histograms of the

unfiltered and filtered data were plotted. The

filtered data was differentiated to determine the

accuracy of the initiation and end of eye movement

The filtered EOGs for each subject were

analysed to extract statistical data as an indicator of

the reliability of the EOG. The change of voltage

for each step of eye movement was determined by

calculating the mean and median of the initial level

and final level for each EOG transition. The absolute

difference of the means and the medians were

tabulated. The respective means derived from the

EOG signals were divided by the expected angular

displacement of the eyes producing a pseudo system

resolution, in Volts/Degree.

The cumulative amplitude {histograms} sets

were statistically analysised (mean, median and

standard deviation), to assist with the study of EOG

reliability as eye an gaze indicator while determining

the reliability of EOG for HCI. Linear regression

was applied to determine a suitable mathematical

expression. The deviation of the EOG from the

mean were plotted against the angle of the eye gaze.

This was repeated for vertical and horizontal

movements.

The analysis was repeated after filtering the

recorded data using a 2

order 1H z Butterworth LP

filter and then using no filter.

Table 1. Linear Regression Analysis – Horizontal Eye

Movement

Eye Gaze Range

±60º ±45º

squared

0.9788 0.9921

adjusted

0.9782 0.9918

standard

error (σ

model

)

0.5423 0.2902

computed

1751.02

3512.485

Intercept 0.0111 -0.023

Angle/ V

of EOG

0.0873 0.0969

Data

Points

40 30

Table 2. Linear Regression Analysis – Vertical Eye

Movement

4 RESULTS

From the data (Figure 3 and 4) it is observed that

there is a region where the eye gaze angle and the

value of EOG are directly related. In this region, the

angle of the gaze can be estimated from the EOG

with good reliability. A linear region was indicated

between ±45 degrees, for horizontal eye movement,

and between −38.7 and +30.7 degree for Vertical eye

movement. The data for eye gaze angles versus

measured EOG are plotted in Figures 4 and 5. The

figures show the standard deviation bounds on

specific EOG data points and show increasing

variance as eye gaze angle approaches its limits.

Eye Gaze Range

40.9≤θ≤30.7

38.7≤θ≤30.7

squared

0.9796 0.9884

adjusted

0.9787 0.9878

standard

error (σ

model

)

0.3672 0.2507

computed

1105.35 1613.72

Intercept 0.0424 -0.0221

Angle 0.1002 0.1060

Data

Points

25 21

Eye Gaze for Computer Control

481

Results from linear regression analysis, support the

reliability of EOG to estimate eye gaze angle for

horizontal and vertical eye movement. Tables 1 and

2 tabulate specific results of the linear regression

analysis of the EOG amplitude, after filtering by a

1Hz 2

order Butterworth low pass filter. Similar

data was obtained, not tabulated, after linear

regression analysis was conducted on EOG data

filtered by a 5Hz 12

order Butterworth low pass

filter and using no filter. The 95% confident interval

for the predicted eye gaze angles were observed to

be improved when analysis of the vertical and

horizontal eye movement EOG data was restricted to

the apparent linear region.

5 OBSERVATIONS AND

DISCUSSION

Linear regression analysis of the EOG data resulted

in good correlation supported by F-test confirmation,

that the mathematical expression adequately

modelled the relationship between eye gaze angle

and EOG. As this occurred with out the use of

visual feedback by the test subjects to adjust their

point of eye gaze, improved reliability can be

expected when feed back is implemented.

A finite intercept from the linear analysis has

indicated that either experimental or physiological

artifacts were present during the acquisition phase.

These artifacts can be manifested in various ways.

From the results of analysis it was observed that:

o There was a small DC drift when the

subject is at reference point.

o There was a degree of hysteris in the

EOG recording when the eye gaze

shifts from right to left, or top to

bottom extremes to the reference point.

o The variations of the EOG increased at

large angles - where the vision boarded

the peripheral vision.

o Differentiation of EOG signals shows

distinct pulses where voltage

transitions occur on the EOG trace. It

was however observed, that if

transitions were to close they tended to

merge.

o Independence of distance from a target

was assumed for this study, this may

not be the case. A target may not be

precisely at the center of an

individual’s eye gaze, center of FOV,

when fixating on a target.

o A high order LP Butterworth filter

reduces the confidence intervals. A low

cut off frequency reduces signal

information.

o The existence of a linear region in the

relationship between eye gaze angle

and EOG, confirmed during linear

regression analysis, ensures that a

simple relationship can be

implemented. The size of the linear

region also ensures that a usable range

of eye gaze angles can be utilized in a

HCI. The simple linear relationship

also ensures that confidence can be

assured in the reliability when

implemented in a HCI.

-8.0

-6.0

-4.0

-2.0

0.0

2.0

4.0

6.0

8.0

-80 -60 -40 -20 0 20 40 60 80

Eye Gaze (Degree)

EOG (Voltage)

Figure 3: EOG Horizontal Movement versus Eye Gaze

Angle

-6

-4

-2

-60 -40 -20 0 20 40 60

Eye Gaze Angle (Degree)

Measured EOG (Volts)

Figure 4: EOG Vertical Movement versus Eye Gaze Angle

6 CONCLUSIONS

Based on this study, it The current study has

confirmed that it is possible to derive signals for the

generation of spatial controls from electro-

oculography recordings and to have confidence in

the reliability when used in a HCI. It has been found

that for restricted ranges of eye gaze angles that

ICINCO 2004 - ROBOTICS AND AUTOMATION

482

EOG and eye gaze are related by a simple linear

mathematical expression. It was also found that

95% confidence intervals and the correlation

between EOG and Eye Gaze angle were improved.

Greater confidence in the reliability of an

HCI utilizing EOG data is assured with further

reduction of system artifacts that result in errors and

appropriate methods of calibrating the HCI system

for individuals. Improved reliability of the HCI

system would also been obtained by implementing

visual feedback allowing fine adjustments to target

position.

REFERENCES

Barea R, Boquete L, Mazo M, Lopez E, Bergasa L

M, (2000), EOG guidance of a wheelchair using

neural networks, IEEE workshop, Electronics

Department,

Gips J, Olivieri, (1996), EagleEyes: An Eye Control

System for Persons with Disabilities. 11

International Conference on Technology and

Persons with Disabilities, Los Angeles

http://www.cs.bc.edu/~eagleeye/University of

Alcala, Spain.

Kuno Y, Yagi T, Uchikawa Y,( 1997), ‘Biological

Interaction between Man and Machine’,

Intelligent robots and Systems, IEEE- IROS 97

Kuno Y, Yagi T, Uchikawa Y, (1998), Development

of eye pointer with free head-motion,

Proceedings of the 20

Annual International

Conference of the IEEE Engineering in Medicine

and biology Society, Vol 20, no4 1998

Kwon S H, Kim H C, (1999), EOG-Based Glasses-

Type Wireless Mouse for the Disabled,

Proceedings of the First Joint BMES/ EMBS

Conference Serving Humanity, Advancing

Technology, Oct 13-16, 99, Atlanta, GA, USA

(IEEE).

Kaufman A E, Bandopadhay A, Shaviv B, (1993),

An Eye tracking Computer user Interface, US

patent 5360971, 1993.

www.cs.sunysb.edu/`vislab/projects/eye/Papers/shor

tpaper.pdf

Norris G, Wilson E, (1997), The eye mouse, An Eye

Communication Device, IEEE Trans Comp,

1997.

Eye Gaze for Computer Control

483