Is Human Visual Activity in Simple Human-Computer Interaction

Search Tasks a Lévy Flight?

Jerzy Grobelny, Rafał Michalski and Rafał Weron

Computer Science and Management Faculty, Wrocław University of Technology, Wrocław, Poland

Keywords: Eye Tracking Data, Heavy Tailed Distribution, Visual Search, Search Tasks, Human-Computer Interaction.

Abstract: The paper tries to answer the question regarding the nature of the statistical distribution of data gathered by

eye tracking software. The experimental data regarding typical search tasks performed while using web sites

were formally analysed and discussed. Results show some resemblance of the obtained experimental

distributions of distance travelled to heavy tailed power-law type distributions characteristic of Lévy flights.

However, the similarity is not as strong as it has been suggested by previous studies. The results of this

paper may be used in further attempts of modelling human visual processing in the context of simple

human-computer interfaces.

1 INTRODUCTION

Measuring and modelling human behaviour is very

interesting for researchers from various fields and

has been subject to intense scientific investigations.

Since most data that the human being gets from the

environment comes from our visual system, it is not

surprising that many studies are focused on this

aspect.

There has been plenty of research taking

advantage of various modern technologies that allow

for gathering information about how people see and

how visual information is processed by our brains.

Among techniques that are relatively simple on

one hand side and provides significant amount of

interesting objective data are the eye tracking

systems. The gathered data may be used for various

types of analyses both qualitative as well as

quantitative. In this paper we take advantage of the

eye tracking data to examine some statistical

properties of the human visual behaviour registered

during search tasks performed in the context of

simple human-computer interactions.

Some of the previous general studies show that

the human visual activity seems to be similar to the

so-called Lévy Flights. Recall that a Lévy flight is a

random walk in which the step-lengths have a

probability distribution that is heavy-tailed. A

heavy-tailed distribution is a probability law that has

tails at least heavier than exponential. Some authors

require more – that the tails are at least of power-law

type (see e.g. Janczura and Weron, 2012).

This idea was put forward by Brockmann and

Geisel (1999) in their conference paper. Their model

was experimentally verified for free scans of natural

scenes presented on a computer screen.

A different model automatically generating

human scanpaths in a nondeterministic way deriving

from that work was proposed by Boccignone and

Ferraro (2004). They combined the algorithms of

providing the saliency maps with gaze shifts

determined by a stochastic process with non-local

transition probabilities. A random walk steps were

generated according to a Lévy process distribution.

The artificially obtained scanpaths were qualitatively

very similar to the real eye tracking data. Also here

the validation involved free visual screening tasks.

The suggestion that the human visual activity

may be successfully modelled by Lévy flights or

some combinations of Lévy flights and other

processes appeared also in a number of other papers

e.g. Stephen et al. 2009, Shinde et al. (2011),

Boccignone and Ferraro (2013), Liu et al. (2013) or

lately Clavelli et al. (2014).

However, it seems that the presented model

might not be applicable in all situations. Stephen and

Mirman (2010) focused their study on the nature of

gaze shifts statistical distribution both in single-

feature search and visual world paradigm tasks.

They analyzed the data individually for six

participants and came to the conclusion that the

Grobelny J., Michalski R. and Weron R..

Is Human Visual Activity in Simple Human-Computer Interaction Search Tasks a Lévy Flight?.

DOI: 10.5220/0005329500670071

In Proceedings of the 2nd International Conference on Physiological Computing Systems (PhyCS-2015), pages 67-71

ISBN: 978-989-758-085-7

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

visual behaviour is not purely the result of generic

oculomotor dynamics but is also strongly influenced

by the task being performed.

In light of these studies, the main objective of

this research is to examine the visual search

properties in the form of statistical empirical

distributions for two types of target objects in the

context of the human-computer interaction.

Two experiments were conducted for this

purpose. In both, the visual activity was registered

by the eye tracking system. In the first experiment

participants searched for a target object in a toolbar

like graphical panel. In the second study the subjects

were to find a hyperlink in the web page or specified

information in tabulated data.

Next sections of this work describe and discuss

in detail the aforementioned eye tracking

experiments.

2 METHOD

2.1 Participants

Thirty students aged between 23 and 25 years took

part in the studies. They were volunteers and did not

receive any gratification. None of the subjects wore

eye glasses or contact lenses.

2.2 Apparatus

Examinations were conducted in a dim-lit laboratory

using the Pan/Tilt version of the ASL 6000 eye

tracker (Applied Science Laboratories, 2005). The

system records eye position with the frequency of

60 Hz and the precision of one visual angle. Visual

stimuli were presented by GazeTracker™ software

as still images. The same application gathered eye

tracking data sent by the ASL 6000 control unit.

2.3 Independent Variables

The study consisted of two visual search

experiments concerning toolbar like panels and

typical web pages.

2.3.1 Toolbar Search

In the first part of the research users searched for a

target object within a set of distractors arranged in a

panel similar to computer toolbars. Toolbars

included 36 identical objects with letters and

numbers randomly situated within the panel. We

also employed a variety of target background

colours to make the experimental results more

diverse.

2.3.2 Web Site Search

The second part of the study was focused on

searching various targets in web sites. We used

popular in our country web sites in one domain

namely V, X, Y and Z. These web sites had different

design and layout. Similarly as in the first part of the

examination also here to make the gathered data be

more ecologically valid two types of typical search

tasks were involved. The first one was a simple

search for a given link on the typical multimedia

page, and the second one dealt with finding specific

information presented in a tabular way.

2.4 Dependent Measures

Temporal and spatial parameters of the visual

activity gathered by the GazeTracker™ computer

program, integrated with the eye tracking equipment

were used as dependent measures. In this work we

specifically analysed saccade lengths gaze points’

shifts.

2.5 Experimental Design and

Procedure

Before the experiment, participants were informed

about the goal and the scope of the examination.

Next, eye tracking system calibration took place.

Then, subjects performed first the toolbar search

tasks and after a short break, the web sites searches

took place.

2.5.1 Toolbar Search

Six various colourful toolbars were examined. The

within subjects design was applied, thus, the given

participant tested all the graphical variants.

Figure 1: Toolbar visual search task.

PhyCS2015-2ndInternationalConferenceonPhysiologicalComputingSystems

There were two trials per condition so the

participants searched the graphical panels 12 times.

The toolbars were displayed in the left upper corner

of the screen, and they were visible only during the

visual search. The subjects were to click on the

target item displayed in the instructive slide, then the

appropriate panel appeared and their task was to find

and select the desired object. This procedure is

illustrated in Figure 1.

2.5.2 Web Site Search

The web site search was repeated ten times for each

participant. There were six simple visual searches

and four trials of complex table searches. The tasks

were presented randomly. Each participant

performed all ten search tasks. Every time the

instruction appeared in the middle of a screen and

then, after mouse click, the tested screen was

exposed. The subject was to find the target object as

fast as possible and click it. Then, the next

instruction was presented. The exemplary task is

presented in Figure 2.

Figure 2: Web site visual search task.

3 RESULTS AND DISCUSSION

Since the main goal of this paper was to verify if eye

tracking data gathered for the typical human-

computer interaction search tasks come from the

process akin to Lévy flights, we prepared empirical

probability distributions (PDF) and cumulative

probability distributions (CDF). The resulting data

are illustrated in Figures 3-8.

Apart from Euclidean distances between

gazepoints the graphs also contain standardized

distances computed by dividing the gaze shift by the

time between two gazepoints and multiplied by the

average shift time for all saccades. For the

comparison purposes, the figures present also the

theoretical power-law and exponential curve fits.

The CDFs/PDFs were elaborated separately for the

graphical panel and web site search tasks and are

given in Figures 3 and 4 respectively.

Figure 3: Empirical probability distribution of Euclidean

distances between gazepoints (in pixels) for the

experiment with toolbar-like panels.

Figure 4: Empirical probability distribution of Euclidean

distances between gazepoints (in pixels) for the

experiment with web pages.

While analysing an empirical distribution that

could exhibit heavy tails, it is convenient to plot the

(right) tail of the distribution, i.e. 1 – CDF, on a

double logarithmic or a semi-logarithmic scale. If

the data points constitute a straight line on the

double logarithmic scale then the tail of the

distribution is approximately of power-law type. On

the other hand, if the data points constitute a straight

line on the semi-logarithmic scale then the tail of the

distribution is approximately exponential. Therefore,

Figures 5 and 6 demonstrate 1 – CDFs for both

experiments and in Figures 7 and 8 the Euclidean

distances between gazepoints are additionally

presented on a logarithmic scale.

IsHumanVisualActivityinSimpleHuman-ComputerInteractionSearchTasksaLévyFlight?

Figure 5: Right tail of the empirical cumulative

distribution of Euclidean distances between gazepoints (in

pixels) for the experiment with toolbar-like panels on a

semi-logarithmic scale.

Figure 6: Right tail of the empirical cumulative

distribution of Euclidean distances between gazepoints (in

pixels) for the experiment with web pages on a semi-

logarithmic scale.

Figure 7: Right tail of the empirical cumulative

distribution of Euclidean distances between gazepoints (in

pixels) for the experiment with toolbar-like panels on a

double logarithmic scale.

Figure 8: Right tail of the empirical cumulative

distribution of Euclidean distances between gazepoints (in

pixels) for the experiment with web pages on a double

logarithmic scale.

The obtained plots show that the tails are almost

linear on a semi-logarithmic scale for the web pages

search tasks and only slightly heavier than linear for

the toolbar-like panel searches. In general, the

obtained empirical distributions have tails that are

not heavy enough to come from a power-law type

distribution. Thus, these findings rather do not allow

for comparing the registered human visual activity to

Lévy flights.

4 CONCLUSIONS

In general, the obtained empirical distributions do

not fully reflect the distributions characteristic of

Lévy flights. On the other hand, results and

simulation studies of previous papers (Brockmann

and Geisel, 1999; Boccignone and Ferraro, 2004)

quite convincingly support the hypothesis that the

human visual activity is to some degree similar to

Lévy flights. Thus, one may suggest that the people

may flexibly adapt their visual activity -

intentionally or unconsciously - to the existing

needs. Presumably, there does not exist any single

process that would be appropriate for modelling eye

balls behaviour in all situations.

The obtained results seem to be in concordance

with the findings of Stephen and Mirman (2010)

where the distributions gathered for various viewing

conditions for six individuals were considerably

different.

The presented data give some more insight into

the nature of human visual behaviour in a specific

context. Answering the question whether and in

what circumstances the human visual behaviour can

PhyCS2015-2ndInternationalConferenceonPhysiologicalComputingSystems

be modelled by Lévy flights may be helpful in the

further development of models from the Human-

Computer Interactions field trying to imitate the

human visual activity such as ACT-R (Anderson et

al. 1997) or EMMA (Salvucci, 2001).

ACKNOWLEDGEMENTS

The work was partially financially supported by

Polish National Science Centre Grant No.

2011/03/B/ST8/06238.

REFERENCES

Anderson J.R., Matessa M., Lebiere C., (1997), ACT-R:

a theory of higher level cognition and its relation to

visual attention, Human Computer Interaction, 12,

439–462, doi:10.1207/s15327051hci1204_5.

Boccignone, G., and Ferraro, M. (2004). Modelling gaze

shift as a constrained random walk. Physica A:

Statistical Mechanics and its Applications, 331(1–2),

207–218. doi:10.1016/j.physa.2003.09.011.

Boccignone, G., and Ferraro, M. (2013). Feed and fly

control of visual scanpaths for foveation image

processing. Annals of Telecommunications - Annales

Des Télécommunications, 68(3-4), 201–217.

doi:10.1007/s12243-012-0316-9.

Brockmann, D., and Geisel, T. (1999). Are human

scanpaths Lévy flights? In: Artificial Neural

Networks, 1999. ICANN 99. Ninth International

Conference on (Conf. Publ. No. 470) (T. 1, pp. 263–

268 vol.1). doi:10.1049/cp:19991119.

Clavelli, A., Karatzas, D., Lladós, J., Ferraro, M.,

and Boccignone, G. (2014). Modelling Task-

Dependent Eye Guidance to Objects in Pictures.

Cognitive Computation, 6(3), 558–584.

doi:10.1007/s12559-014-9262-3.

Janczura, J., and Weron, R. (2012). Black swans or dragon

kings? A simple test for deviations from the power

law, European Physical Journal - Special Topics (EPJ

ST) 205, 79-93.

Liu, H., Xu, D., Huang, Q., Li, W., Xu, M., and Lin, S.

2013. Semantically-Based Human Scanpath

Estimation with HMMs. In: 2013 IEEE International

Conference on Computer Vision (ICCV) (ss. 3232–

3239). doi:10.1109/ICCV.2013.401.

Salvucci D.D., 2001, An integrated model of eye

movements and visual encoding, Cognitive Systems

Research, 1 (4), 201–220. doi: 10.1016/S1389-

0417(00)00015-2.

Shinde, D. P., Mehta, A., and Mishra, R. K. (2011).

Searching and fixating: Scale-invariance vs.

characteristic timescales in attentional processes. EPL

(Europhysics Letters), 94(6), 68001. doi:10.1209/

0295-5075/94/68001.

Stephen, D. G., and Mirman, D. (2010). Interactions

dominate the dynamics of visual cognition. Cognition,

115(1), 154–165. doi:10.1016/j.cognition.2009.12.

010.

Stephen, D. G., Mirman, D., Magnuson, J. S., and Dixon,

J. A. (2009). Lévy-like diffusion in eye movements

during spoken-language comprehension. Physical

Review E, 79(5), 056114. doi:10.1103/PhysRevE.

79.056114.

IsHumanVisualActivityinSimpleHuman-ComputerInteractionSearchTasksaLévyFlight?