Interaction-based Implicit Calibration of Eye-Tracking in an

Aircraft Cockpit

Simon Schwerd

and Axel Schulte

Institute of Flight Systems, Universität der Bundeswehr München, Neubiberg, Germany

Keywords: Eye-Tracking, Implicit Calibration, Aviation, Flight Deck.

Abstract: We present a method to calibrate an eye-tracking system based on cockpit interactions of a pilot. Many studies

show the feasibility of implicit calibration with specific interactions such as mouse clicks or smooth pursuit

eye movements. In real-world applications, different types of interactions often co-exist in the “natural”

operation of a system. Therefore, we developed a method that combines different types of interaction to enable

implicit calibration in operational work environments. Based on a preselection of calibration candidates, we

use an algorithm to select suitable samples and targets to perform implicit calibration. We evaluated our

approach in an aircraft cockpit simulator with seven pilot candidates. Our approach reached a median accuracy

between 2° to 4° on different cockpit displays dependent on the number of interactions. Differences between

participants indicated that the correlation between gaze and interaction position is influenced by individual

factors such as experience.

1 INTRODUCTION

Eye-tracking is an important tool to study human

factors in aircraft cockpits. In many simulator studies,

gaze measurement is used to analyse pilot attention

(van de Merwe et al., 2012; Ziv, 2016) or conduct

research about pilot cognition (Dahlstrom &

Nahlinder, 2009; Schwerd & Schulte, 2020; van de

Merwe et al., 2012). With ongoing development, it

might be used in real cockpits to complement "black

box" flight recorder information (Peysakhovich et al.,

2018) or to inform adaptive support systems (Brand

& Schulte, 2021; Honecker & Schulte, 2017). Robust

and accurate measurement of pilot gaze is

fundamental to these undertakings. Therefore, proper

calibration of the eye-tracking system plays an

important role (Nyström et al., 2013).

2 BACKGROUND

Calibration of eye-tracking systems requires multiple

pairs of calibration targets 𝑡



(

𝑥, 𝑦

)

and corresponding

measured gaze samples 𝑠



(

𝑥, 𝑦

)

. For explicit

calibration, a user is required to fixate predefined

https://orcid.org/0000-0001-6950-2226

calibration targets while gaze is measured. This time-

consuming process might decrease acceptance of eye-

tracking applications in aircraft cockpits. Further,

measurement accuracy can deteriorate during

operation due to head movements, blinking or a

change in the user’s relative position to the screen (M.

X. Huang et al., 2016; Sugano & Bulling, 2015),

which could go unnoticed without re-calibration.

With implicit calibration, no explicit user

cooperation is required, and calibration targets are

determined based on assumptions about probable or

required fixation. This idea was originally proposed

by Hornof and Halverson (2002), where participants

in an experiment had to fixate a specific display

position for task-relevant information which was used

to monitor gaze accuracy. Following this idea,

subsequent studies investigated implicit calibration

based on the correlation of gaze and interaction such

as mouse clicks (M. X. Huang et al., 2016; Sugano et

al., 2008), scene properties like saliency (Kasprowski

et al., 2019; Sugano & Bulling, 2015), moving objects

(Blignaut, 2017; Drewes et al., 2019; Pfeuffer et al.,

2013) or saccade properties (M. X. Huang & Bulling,

2019). Most of this research was conducted in the

context of eye-tracking applications for untrained

users, such as for webcams of internet users

Schwerd, S. and Schulte, A.

Interaction-based Implicit Calibration of Eye-Tracking in an Aircraft Cockpit.

DOI: 10.5220/0011657200003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 2: HUCAPP, pages

55-62

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

(Papoutsaki et al., 2017) or for displays in public

spaces (Khamis et al., 2016).

One of the most studied implicit calibration

methods is the correlation of gaze and mouse clicks.

Sugano et al. (2008) and Sugano et al. assumed, that

gaze and mouse click position align when the click

occurs and presented an algorithm that updated their

gaze estimation with every click. According to J.

Huang et al. (2012) the proximity of mouse cursor

and gaze depends on the current mode of interaction

(e.g. read or click) with a median distance of 70-80px

for mouse clicks. In a subsequent study, the authors

found, that the correlation of gaze and interaction is

stronger before than exactly at the moment of

interaction peaking at around 0.4s before interaction

(M. X. Huang et al., 2016). In a more recent study,

Zhang et al. (2018) implemented implicit calibration

combining clicks, touches and keypresses over

multiple devices. These studies showed the feasibility

of interaction-based implicit calibration but pointed

out two relevant problems: First, gaze data is usually

noisy, which challenges the basic assumption, that

gaze always aligns with an interaction. Second, task

context and individual differences cause inconsistent

alignment of gaze and interaction. Both these

observations suggest that the interaction-based

approach must be able to filter for correct gaze-

interaction alignments.

The mentioned research showed the feasibility of

implicit calibration. In most studies, the calibration is

usually based on one specific type of interaction or

scene property. From our review, we identified two

open questions: First, how different interactions can

be used in the same method for calibration. Second,

if the implicit calibration assumptions hold in the use

of human-machine-interfaces not specifically

designed for this use case. We are convinced that

implicit calibration will prove to be useful in

operational workplaces to avoid explicit calibration

and continuously verify eye-tracking quality. This

paper makes the following contributions: First, we

propose a method to select calibration target

candidates that integrates different types of

interactions. Second, we describe an algorithm, which

can process these different types in a unified manner.

Third, we evaluate the implementation of this system

in our aircraft cockpit simulator.

3 METHOD

We propose a method with three stages displayed in

Figure 1. The first step “Calibration Candidate

Selection” is a classification to identify different

types of interaction in natural human-system-

interaction that are suitable for implicit calibration.

After a calibration target and associated samples has

been identified, the second module “Sample

Selection” identifies a subset of samples that are

likely to be a fixation upon the assumed target. After

that, a suitable calibration procedure uses the new and

past selected samples to calibrate the system.

Figure 1: Calibration pipeline. Two stages of selection

result in a selection of samples and calibration targets for

calibration.

3.1 Selection of Calibration Candidates

To identify suitable calibration targets in a natural

human-machine interaction, we propose the

following approach: We think about interaction and

gaze measurements in terms of two parallel data

streams (see

Figure 2

): First, a gaze data stream, which

arrives at high frequency but possibly interrupted

(e.g., when the eye-tracker loses the track) and

second, interactions, which either happen at distinct

moments (e.g., click) or are continuous over a time

period (e.g., hold gesture on a touch screen).

Given this perspective, we propose three

categories to compute calibration targets and gaze

sample candidates: an interaction is either a

successor, companion, or precursor of a probable

fixation target. We define an interaction to be a

HUCAPP 2023 - 7th International Conference on Human Computer Interaction Theory and Applications

Figure 2: Gaze samples and interactions as parallel data

streams with three types of calibration target candidates.

successor when the interaction follows a probable

fixation target, for example mouse clicks, which

succeed a fixation on the click position. After

occurrence of a successor, the preceding gaze

samples are selected as calibration candidates. Next,

we define an interaction to be a companion when it

likely occurs together with its associated fixation

target. As an example, consider a slider widget in a

desktop application that controls a numerical value

displayed next to it. A user needs to fixate the

changing number when he uses his mouse to slide,

thus, it can be considered a fixation target. In this

case, all gaze samples measured during an active

companion interaction are calibration candidates.

Finally, we define an interaction to be a precursor

when the interaction occurs before a probable

calibration target. Many interactions with the

interface cause changes that are verified by the user

just after interaction. To give a flight deck example,

when a pilot moves the gear of their plane by control

input, he confirms the position of the gear a few

seconds after his input. The action and confirmation

of the expected result could be used for implicit

calibration, which is especially useful when the

interaction is spatially or temporarily dislocated from

the location where the change is displayed (e.g.,

control input and delayed indication of gear).

3.2 Sample Selection

The calibration candidate selection returns an

assumed fixation position and associated gaze sample

candidates for calibration that were measured either

before, during or after the interaction. Since this data

set is often noisy, we need to identify the most

probable subset of samples representing the fixation

on the assumed target. For this, we propose a simple

algorithm inspired by RANSAC (Fischler & Bolles,

1981), which selects samples that are in both spatial

and temporal proximity to the interaction.

Pseudocode is depicted in Algorithm 1.

INTERACTION -TIME -CONSENSUS

Input: gaze-samples, calibration-target, interaction-time, threshold-dist, threshold-time,

minimum-samples

Result: Best samples for calibration

all-consensus-groups

← empty list

for

sample-candidate

gaze-samples

do:

consensus-group

←

gaze-samples

as selected by

candidate selection

for

consensus-candidate

consensus-group

do:

spatial-dist

← SPATIAL_DISTANCE (

consensus-candidate

sample-candidate

)

time-dist

←TIME_DISTANCE_ABSOLUTE(

consensus-candidate

interaction-time

)

if spatial-dist

threshold-dist

time-dist

threshold-time

do:

REMOVE

consensus-candidate

from

consensus-group

SIZE(consensus-group) < minimum-samples

do:

continue

ADD

consensus-group

all-consensus-groups

return consensus-group with lowest time to interaction

Algorithm 1: Algorithm to select most probable subset of

fixation samples for calibration target.

3.3 Calibration

Optimally, the sample selection algorithm returns a

set of suitable pairs of calibration targets 𝑡



(

𝑥, 𝑦

)

and

measured gaze samples 𝑠



(

𝑥, 𝑦

)

while omitting

targets with noisy measurements or insufficient

sample size. Then, gaze can be calibrated when

enough of these pairs have been identified. The

method of calibration is flexible and depends on the

use-case and desired fitting attributes. In the

following, we present an experiment where we

collected data in an experiment and applied a linear

calibration function.

4 EXPERIMENT

To evaluate our approach, we conducted an

experiment in a research cockpit simulator

resembling a generic fast jet cockpit (see Figure 3),

which is normally used to evaluate human-autonomy

teaming applications (e.g., (Lindner & Schulte,

2020)). The cockpit contains three touchscreens and

a Head-Up-Display (HUD) projected into the visuals

of the outside world. The participant’s gaze was

measured with a commercially available four-camera

system connected to the simulation software

(SmartEye Pro 0.3MP) and gaze was tracked over the

three touchscreen displays and the HUD.

In the experiment, we processed the uncalibrated

screen positions provided by the eye-tracking system

consisting of x- and y-coordinate and the associated

screen. The basic idea of the experiment was to

Interaction-based Implicit Calibration of Eye-Tracking in an Aircraft Cockpit

collect natural interaction data and uncalibrated gaze

measurements in a flight mission, which was then

used to perform implicit calibration. Results of this

process were compared to the results of a baseline

calibration based on a conventional 9-point

calibration procedure.

4.1 Calibration Targets

There are a great number of different controls and

displays in the cockpit simulator. The participants

controlled their aircraft via throttle, stick, and touch

interactions. Possible touch interactions were tap,

hold, drag, and pinch dependent on the control

element. From these different touch interactions, we

used all taps as successors, and all holds as

companions. Drag and pinch were not used. In the

following, we want to give a brief overview of the

most important elements on the touchscreen, which

are also displayed in Figure 3.

The tactical map in the central touchscreen can

be used for navigation and creation of mission tasks.

It displays the aircraft, all unmanned systems as well

as tactical objects and mission information. The map

can be dragged and zoomed. It is possible to interact

with tactical elements (e.g., building or waypoint)

by tapping on the respective element, which is

followed by a pop-up context menu used to plan

mission tasks

(e.g., reconnaissance). Taps were used as

predecessors.

On the left side display, the main interaction

element is the mission plan timeline. After the

creation of a task in the map, it can be inserted into

the timeline by tapping a position within the timeline

after which a task box moves to its position from the

bottom of the screen. Taps were used as successors.

The autopilot control can be opened by tapping

on the associated button in the side bar of the central

display. The control contains buttons to increase and

decrease autopilot speed, altitude and heading as well

as buttons to engage autopilot modes. The numeric

values can be adjusted by either tap or hold

interactions on the increase and decrease buttons. The

calibration target of each interaction lies on the

position of the value display. Taps and holds were

used as successor and companions, respectively.

The HUD shows several different flight

parameters of the aircraft. For implicit calibration, we

used the air brake indicator in the centre of the

display. The participants controlled the air brake via

a throttle button and the indicator in the shows its

current position as illustrated by Figure 3. The

interaction to control the air brake is considered a

predecessor.

Figure 3: Simulator (top center) with relevant interaction elements: Mission timeline (top left), HUD air brake indicator (top

right), autopilot control (bottom left), tactical map (bottom center) and context menu (bottom right)

HUCAPP 2023 - 7th International Conference on Human Computer Interaction Theory and Applications

4.2 Experimental Task

The experiment contained different tasks for the

participants to trigger interactions on all displays.

Before take-off, participants were asked to plan a

mission, which required the creation of several tasks

for their aircraft and a UAV. This included an

interaction with the mission timeline. With their own

aircraft, they had to take-off and fly over several way

points or tactical objects, while their UAV

investigated four different locations on the map. After

the take-off, participants received messages with

instructions to program the autopilot with given

values. During the flight, the UAV sent pictures

which had to be classified in two categories. The

participant landed again at the airport after

completion of all waypoints.

4.3 Participants & Procedure

We conducted the experiment with seven participants

(all male, 𝜇



=24.9𝑦). Four participants were

students of aeronautical engineering and three were

research assistants. All participants had prior

experience in handling the simulator. In the beginning

the participants received an introduction to the

simulator and the specific tasks necessary to complete

the experimental requirements. Then, each

participant had to complete three training missions to

get familiar with the mission tasks. Before each

mission there was a briefing in which the tasks and

circumstances were demonstrated. Prior to the

experimental mission, we collected validation data

based on a 9-point validation on every screen. After

this data collection, the participant received a briefing

for the experimental mission, which followed on the

briefing.

4.4 Data Processing

During the 9-point validation procedure, we collected

100 samples for 9 points on each display. The error

was computed by average Euclidian distance of each

sample to the calibration target. In the experiment, we

logged the following data: (1) position and timestamp

of uncalibrated gaze measurements and (2) position,

timestamp, type, and associated display element (e.g.,

autopilot value, map) of interaction. For calibration,

we computed the constants that optimized error given

the application of the following function to the

samples for both x- and y-coordinates:

𝑓

(𝑥) = 𝑎



+ 𝑎



𝑥

(1)

For the HUD, we used the following function

since there was only one possible calibration point in

the centre of the display:

𝑓

(𝑥) = 𝑎



+ 𝑥

(2)

Note, that we computed the calibration

parameters for each display independently.

5 RESULTS

First, we aimed to identify the optimal set of

parameters for the sample selection algorithm. After

selecting a set of parameters, we evaluated how many

calibration targets and samples were selected for

calibration. Finally, we compared calibration

performance of implicit calibration with 9-point

calibration.

5.1 Optimal Parametrization of the

Selection Algorithm

The algorithm presented in section 3.2 expects three

parameters: distance threshold, time threshold and

minimum number of samples. We performed a grid

search to identify optimal parameters. We

tested the combinations of the following

parameters: distance threshold

[

1.0, 1.5, 2.0

]

° ,

time threshold

[

0.5, 1.0, 1.5, 2.0

]

𝑠 and minimum

samples [10, 15, 20, 25] . We found varying

parameters for the participants suggesting individual

differences in the gaze-interaction correlation. Also,

an individual random component in gaze

measurement could influence optimal distance

threshold because a low threshold filters fixations

with high variance. For the following analysis, we

selected a set of parameters returning the best median

calibration performance: distance threshold of 2.0°,

time threshold of 2.0𝑠 and minimum samples of 15.

These parameters were used for all participants to

generate the following results.

5.2 Interactions

Interaction with each display differed in frequency as

shown by Table 1. Participants mainly interacted by

taps because this was the primary way to operate the

system. In addition, some participants did not use a

hold interaction to control the autopilot but preferred

to tap the buttons multiple times, which is one reason

for the high variations of successors and companions

in the centre display.

Interaction-based Implicit Calibration of Eye-Tracking in an Aircraft Cockpit

Table 1: Number of occurrences of the different forms of interactions, formatted as “Mean (Standard Deviation)”.

There were very few samples for predecessor

interaction on the HUD because was only controlled

to land the aircraft. One participant did not use the air

brake at all. In average, the sample selection removed

15.8% of successors and 81.5% of companions

because no valid sample group was found.

Surprisingly, not many successor interactions on the

left and right display were removed.

5.3 Calibration Performance

We computed calibration constants with the selected

calibration targets and samples from the experiment.

Using these constants, we calibrated the data

collected in the 9-point validation data of each

display. Then, we compared the error of this

implicitly calibrated data with uncalibrated and point-

calibrated data. Results are displayed in Figure 4.

There are improvements against no calibration for

most participants on most displays but there are also

instances where our method deteriorated the gaze

measurement. Table 2 reports absolute and relative

values for implicit calibration. Median accuracy is

between 2.1° on the left display to 3.8° on the HUD.

For very few cases, improvement by implicit

calibration was comparable to point calibration (e.g.,

see Center, P7 or Side Left P6). Standard deviation is

high on the HUD, because of a great implicit

calibration error for participant 2 and 4.

Table 2: Median accuracy and relative improvement of

implicit calibration.

Display Median Accuracy

(

Std

)

Median relative

rovement

(

Std

)

Cente

2.2° (0.96°

)

44%

(

65%

)

Side Left

2.1°

(

1.03°

)

22%

(

41%

)

Side Ri

2.8° (0.96°

)

36%

(

36%

)

HUD

3.8° (3.67°

)

(

136%

)

5.4 Discussion

The results showed that the implicit calibration

method can improve average accuracy on the 9-point

validation but did not reach point calibration quality.

Therefore, we conclude that the basic assumptions of

implicit calibration do hold for interactions in work

environments like an aircraft cockpit, but there are

limitations to the presented approach. First, the

parameter search revealed that the algorithm had

different optimal parameters for different participants

which indicates an individual component in implicit

calibration. We assume, that two factors are

responsible for these individual components. First,

participants had different levels of training on the

system e.g., the research assistants developed parts of

this prototype and students had only experience from

prior experiments. When a user is very experienced,

he anticipates system reactions and might be very

quick in his interaction. Second, individual factors

such as computer game experience might also be a

confounding factor influencing the correlation

between gaze and interaction. Apart from individual

difference, another problem is the varying frequency

of different interaction types. In our use-case,

successor interactions (taps on the central

touchscreen) occurred far more often than other forms

and therefore had the largest influence on the results.

Contrary, the HUD had very few interactions on one

possible calibration target, which was problematic in

two ways: First, when our method falsely selects

samples where the implicit calibration assumption

does not apply, there is a great accuracy degradation

due to false calibration. The risk of false calibration

is especially high when sample size is low. Second,

when there are very few targets on the screen (e.g., on

the HUD), calibration procedure might overfit areas

where most interactions happen.

Our experiment had two main limitations. First,

although all participants had prior experience on the

simulator, the degree of proficiency differed, which

could also explain individual calibration differences.

Second, the experimental design did not allow to

analyse the problem of continuous deterioration

during measurement. We collected validation data

before the experiment, so gaze tracking quality

change during the experiment could influence

implicit calibration performance.

HUCAPP 2023 - 7th International Conference on Human Computer Interaction Theory and Applications

Figure 4: Comparison of no, implicit and point calibration on each display with 9-point validation data (bar, blue/gray).

Relative improvement is displayed only for implicit calibration (point, orange). Note: There were no samples for Participant

7 on the HUD.

6 CONCLUSION

Our approach demonstrated the feasibility and the

challenges of implicit calibration in work

environments, specifically on flight deck. We are

planning to improve the current approach by

integrating smooth-pursuit calibration, which in our

case, could be implemented for moving objects on the

map or in the external view. Another possible

improvement is to extent the predecessor interactions

based on informed use cases of a professional pilot.

Since pilot interaction is standardized to a high

degree, there are many tasks in the cockpit that could

be leveraged for implicit calibration like for example

take-off procedures or checklists.

REFERENCES

Blignaut, P. (2017). Using smooth pursuit calibration for

difficult-to-calibrate participants. Journal of Eye

Movement Research, 10, 1–13. https://doi.org/

10.16910/jemr.10.4.1

Brand, Y., & Schulte, A. (2021). Workload-adaptive and

task-specific support for cockpit crews: design and

evaluation of an adaptive associate system. Human-

Intelligent Systems Integration, 3(2), 187–199.

https://doi.org/10.1007/s42454-020-00018-8

Dahlstrom, N., & Nahlinder, S. (2009). Mental Workload

in Aircraft and Simulator During Basic Civil Aviation

Training. The International Journal of Aviation

Psychology, 19(4), 309–325. https://doi.org/10.1080/

10508410903187547

Drewes, H., Pfeuffer, K., & Alt, F. (2019). Time- and

Space-Efficient Eye Tracker Calibration. In ETRA ’19,

Proceedings of the 11th ACM Symposium on Eye

Tracking Research & Applications. Association for

Computing Machinery. https://doi.org/10.1145/

3314111.3319818

Fischler, M. A., & Bolles, R. C. (1981). Random Sample

Consensus: A Paradigm for Model Fitting with

Applications to Image Analysis and Automated

Cartography. Commun. ACM, 24(6), 381–395.

https://doi.org/10.1145/358669.358692

Honecker, F., & Schulte, A. (2017). Automated Online

Determination of Pilot Activity Under Uncertainty by

Using Evidential Reasoning. In D. Harris (Ed.),

Engineering Psychology and Cognitive Ergonomics:

Cognition and Design (pp. 231–250). Springer

International Publishing.

Hornof, A. J., & Halverson, T. (2002). Cleaning up

systematic error in eye-tracking data by using required

fixation locations. Behavior Research Methods,

Instruments, & Computers : A Journal of the

Psychonomic Society, Inc, 34(4), 592–604.

https://doi.org/10.3758/bf03195487

Huang, J., White, R., & Buscher, G. (2012). User see, user

point: Gaze and cursor alignment in web search. In

Proceedings of the SIGCHI Conference on Human

Factors in Computing Systems (pp. 1341–1350).

https://doi.org/10.1145/2207676.2208591

Huang, M. X., & Bulling, A. (2019). SacCalib. In K. Krejtz

& B. Sharif (Eds.), Proceedings of the 11th ACM

Symposium on Eye Tracking Research & Applications

Interaction-based Implicit Calibration of Eye-Tracking in an Aircraft Cockpit

(pp. 1–10). ACM. https://doi.org/10.1145/3317956.

3321553

Huang, M. X., Kwok, T. C., Ngai, G., Chan, S. C., &

Leong, H. V. (2016). Building a Personalized, Auto-

Calibrating Eye Tracker from User Interactions. In

CHI ’16, Proceedings of the 2016 CHI Conference on

Human Factors in Computing Systems (pp. 5169–5179).

Association for Computing Machinery.

https://doi.org/10.1145/2858036.2858404

Kasprowski, P., Harezlak, K., & Skurowski, P. (2019).

Implicit Calibration Using Probable Fixation Targets.

Sensors (Basel, Switzerland), 19(1). https://

doi.org/10.3390/s19010216

Khamis, M., Saltuk, O., Hang, A., Stolz, K., Bulling, A., &

Alt, F. (2016). TextPursuits: Using Text for Pursuits-

Based Interaction and Calibration on Public Displays.

In UbiComp ’16, Proceedings of the 2016 ACM

International Joint Conference on Pervasive and

Ubiquitous Computing (pp. 274–285). Association for

Computing Machinery. https://doi.org/10.1145/

2971648.2971679

Lindner, S., & Schulte, A. (2020). Human-in-the-Loop

Evaluation of a Manned-Unmanned System Approach

to Derive Operational Requirements for Military Air

Missions. In D. Harris & W.-C. Li (Eds.), Engineering

Psychology and Cognitive Ergonomics. Cognition and

Design (pp. 341–356). Springer International

Publishing.

Nyström, M., Andersson, R., Holmqvist, K., & van de

Weijer, J. (2013). The influence of calibration method

and eye physiology on eyetracking data quality.

Behavior Research Methods, 45(1), 272–288.

https://doi.org/10.3758/s13428-012-0247-4

Papoutsaki, A., Laskey, J., & Huang, J. (2017).

Searchgazer: Webcam Eye Tracking for Remote

Studies of Web Search. In R. Nordlie (Ed.), ACM

Digital Library, Proceedings of the 2017 Conference

on Conference Human Information Interaction and

Retrieval (pp. 17–26). ACM. https://doi.org/

10.1145/3020165.3020170

Peysakhovich, V., Lefrançois, O., Dehais, F., & Causse, M.

(2018). The Neuroergonomics of Aircraft Cockpits:

The Four Stages of Eye-Tracking Integration to

Enhance Flight Safety. Safety, 4(1), 8.

https://doi.org/10.3390/safety4010008

Pfeuffer, K., Vidal, M., Turner, J., Bulling, A., &

Gellersen, H. (2013). Pursuit Calibration: Making Gaze

Calibration Less Tedious and More Flexible. In

UIST ’13, Proceedings of the 26th Annual ACM

Symposium on User Interface Software and Technology

(pp. 261–270). Association for Computing Machinery.

https://doi.org/10.1145/2501988.2501998

Schwerd, S., & Schulte, A. (2020). Experimental

Validation of an Eye-Tracking-Based Computational

Method for Continuous Situation Awareness

Assessment in an Aircraft Cockpit. In D. Harris & W.-

C. Li (Eds.), Lecture Notes in Computer Science.

Engineering Psychology and Cognitive Ergonomics.

Cognition and Design (Vol. 12187, pp. 412–425).

Springer International Publishing. https://doi.org/

10.1007/978-3-030-49183-3_32

Sugano, Y., & Bulling, A. (2015). Self-Calibrating Head-

Mounted Eye Trackers Using Egocentric Visual

Saliency. In C. Latulipe, B. Hartmann, & T. Grossman.

(Eds.), Proceedings of the 28th Annual ACM

Symposium on User Interface Software & Technology

(pp. 363–372). ACM. https://doi.org/10.1145/

2807442.2807445

Sugano, Y., Matsushita, Y., & Sato, Y. (2013).

Appearance-based gaze estimation using visual

saliency. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 35(2), 329–341. https://

doi.org/10.1109/TPAMI.2012.101

Sugano, Y., Matsushita, Y., Sato, Y., & Koike, H. (2008).

An Incremental Learning Method for Unconstrained

Gaze Estimation. European Conference on Computer

Vision, 5304, 656–667. https://doi.org/10.1007/978-3-

540-88690-7_49

van de Merwe, K., van Dijk, H., & Zon, R. (2012). Eye

Movements as an Indicator of Situation Awareness in a

Flight Simulator Experiment. The International

Journal of Aviation Psychology, 22(1), 78–95.

https://doi.org/10.1080/10508414.2012.635129

Zhang, X., Huang, M. X., Sugano, Y., & Bulling, A.

(2018). Training Person-Specific Gaze Estimators from

User Interactions with Multiple Devices. In R. Mandryk,

M. Hancock, M. Perry, & A. Cox (Eds.), Proceedings

of the 2018 CHI Conference on Human Factors in

Computing Systems (pp. 1–12). ACM. https://

doi.org/10.1145/3173574.3174198

Ziv, G. (2016). Gaze Behavior and Visual Attention: A

Review of Eye Tracking Studies in Aviation. The

International Journal of Aviation Psychology, 26(3-4),

75–104. https://doi.org/10.1080/10508414.2017.

1313096

HUCAPP 2023 - 7th International Conference on Human Computer Interaction Theory and Applications