=Paper=
{{Paper
|id=Vol-1241/paper05
|storemode=property
|title=Experimenting With Polylines on the Visualization of Eye Tracking Data From Observations of Cartographic Lines
|pdfUrl=https://ceur-ws.org/Vol-1241/paper05.pdf
|volume=Vol-1241
|dblpUrl=https://dblp.org/rec/conf/giscience/KaragiorgouKVN14
}}
==Experimenting With Polylines on the Visualization of Eye Tracking Data From Observations of Cartographic Lines==
https://ceur-ws.org/Vol-1241/paper05.pdf
ET4S 2014
Experimenting with polylines on the visualization of eye
tracking data from observations of cartographic lines
Sophia Karagiorgou1, Vassilios Krassanakis1, Vassilios Vescoukis2, Byron Nakos1
1
National Technical University of Athens (NTUA).
School of Rural and Surveying Engineering.
2
Eidgenössische Technische Hochschule Zürich (ETH).
Institute of Cartography and Geoinformation.
Corresponding authors: sokaragi@mail.ntua.gr, vvescoukis@ethz.ch
Abstract. Several visualization methods for eye tracking data exist to help re-
searchers from many disciplines depict data collected in eye tracking experi-
ments. Focusing on eye tracking data from observations of cartographic lines, in
this paper we propose a new visualization of eye tracking data using polylines
inferred from the analysis of samples. This visualization depicts the “average”
line that is actually seen by subjects; such a line can be useful in the study of
various optical representation concepts, such as the assessment of the effects of
alternative cartographic line attributes, distractions, abstraction levels and more,
as well as in other cases such as the study of visual computer interfaces.
Keywords. Eye tracking, visualization, cartography.
1 Introduction and related work
Eye tracking is a widely used methodology in many scientific fields, as it reveals
important findings about the human cognitive processes during the observation of a
visual stimulus. In cartographic research, eye tracking is a valuable tool for the execu-
tion of experiments related to the study of map reading and cartographic design eval-
uation. An important element of eye movement analysis is the visualization of eye
tracking data using techniques referred to the gaze behavior of either individuals or all
the subjects in an experiment. Considering that the amount of data collected can blur
the reference with the visual stimulus, visualization techniques are usually applied
after clustering the gaze recordings in fixations and saccades. A typical visualization
is the scan path graph, where fixations are depicted as circles with radical values re-
lated to their durations, and saccades are presented as connector line segments among
fixations. Other techniques include heat maps and scan path graphs, using variables
such as duration and number of fixations [1]. The idea of using polylines as reference
on eye tracking research has been discussed before from a different perspective ac-
cording to which the reference polyline is known in advance [2].
ET4S 2014, September 23, 2014, Vienna, Austria
Copyright © 2014 for the individual papers by the papers' authors. Copying permitted for private and
academic purposes. This volume is published and copyrighted by its editors.
22
ET4S 2014
In this paper, we report early progress on the depiction of the gaze route history using
a polyline, which is feasible, as the visual trace is generated from sequential raw eye
tracking data [3]. The nodes of such a polyline contain information about the duration
of fixations or other statistical values, which can also be attributed to line sections that
represent saccadic movements. Generally, the reconstruction of gaze route history can
be very useful in the study of several cartographic concepts as a gaze polyline depicts
the line that is actually perceived from subjects.
The motivation for this work stems from methods used in the inference of graph ge-
ometries such as transportation networks, from GPS tracking data. Several such
methods rely on trajectory clustering. Some of the algorithms in the literature [4], [5]
operate on point data and do not take the temporal aspect into consideration. Others
infer curved paths using k-means clustering of raw tracking data along with distance
measures [6]; others transform tracking data to discretized images using Kernel Den-
sity Estimation (KDE). They function well for frequently sampled and redundant
tracking data [7], but are sensitive to noise. Other approaches, relying on computa-
tional geometry techniques [8], operate on tracks of high-resolution and accuracy. The
final category involves trace-clustering approaches that derive a connected road net-
work from vehicle trajectories [9] of different movement types. This work applies
such a technique in eye tracking data to automatically extract “hubs” and construct a
polyline that corresponds to the observed geometry of cartographic lines.
2 Inference of polylines from eye tracking data
The aim of this work is to derive a single polyline geometry from sampled eye track-
ing data from multiple users. Fig. 1 plots data used in our experiment as samples (left)
and tracks (right) at varying shades of gray for each subject, along with the actual
cartographic line that the subjects have been asked to follow, which is shown in blue.
Fig. 1. Eye tracking data from 3 subjects
23
ET4S 2014
2.1 A first version of the proposed algorithm
The proposed algorithm to derive the polylines from eye tracking data involves three
steps; (i) identifying hubs, (ii) connecting hubs, and (iii) reducing the links into a
single geometry, which are discussed in the sequel.
Phase 1: Hubs and spatial fixation. A hub represents the spatial fixation that the eye
creates near an area of interest. Indicators for hub recognition are the number of track-
ing samples, the number of different users and the coverage of an extended area of
focus. The algorithm takes as input the eye tracking data and determines the k-Nearest
Neighbors (k-NN) of each sample, which are subsequently filtered according to the
number of users. On these filtered samples, we apply the DBSCAN clustering algo-
rithm using a distance threshold and a minimum number of samples, depending on the
specifics of the experiment. The centroids of the resulting clusters are the hubs. Fig. 2
shows the hubs derived after applying the hubs inference algorithm in our test dataset.
Phase 2: Connecting hubs. Next, we connect hubs by links. A fringe benefit of the
hubs computation based on spatial fixation is that for all data we know which samples
helped in identifying hubs. To derive links we exploit this knowledge: for each hub
we record the outgoing and/or incoming tracking portions connecting this hub to oth-
ers by scanning all eye tracking data to discover sequences of hubs. The result of this
step is the creation of a sample polyline set that connects hubs with links. In our rep-
resentation of eye tracking data, all tracking samples that are also hubs are marked as
such. Hence, performing a linear scan of all tracking data reveals the respective track-
ing portions that connect hubs.
Phase 3: Compacting links. To this point, we have hubs connected by links derived
from eye tracking data that exhibit spatial fixation at these hubs. In a nutshell, the
algorithm identifies tracking portions that are close to existing links by means of a
buffer region and merges their geometry into the existing link geometry. The size of
the buffer region depends on the specifics of the data; in our case we used 15 pixels as
buffer region. In this step, we neither introduce new hubs nor do we add new links.
We only adjust the geometry of existing links using a three-step algorithm: (i) sort
existing link samples, (ii) determine relevant tracking portions using a buffer region
around link samples, and (iii) adjust the geometry of links based on the tracking data
geometry.
In our experimentation so far we first sort all links according to their length so as to
process longer links first as they may be more significant for polyline construction,
which remains to be further tested future work. In step (ii) the algorithm uses a buffer
region around the examined link sample and retrieves all intersecting portions of other
links. New links are created by interpolating link samples and introducing hubs. New
links are assigned a weight that is the number of the merged links. Link samples are
updated several times during this phase. While the examined links are reconstructed,
new link samples are created to replace links in previous iterations.
24
ET4S 2014
2.2 Polylines inference results
The cartographic line that we try to infer consists of 6595 links (edges) and 6607 hubs
(vertices). The edges have a length of 4041 pixels, as the reference system is in pixels.
Sampling of eye tracking data is at 60 Hz (0.017 sec). Data comes from 3 different
users with a total length of 89880 pixels (Fig. 1). Following the various stages of the
polylines inference algorithm, the following output is produced. During the first
phase, i.e., hubs extraction and connection, 109 hubs and 300 link samples are gener-
ated. The second polylines inference phase, i.e. compacting links, produces 119 hubs,
79 links and a length of 2990 pixels. This result shows that during the second phase of
the algorithm, the number of hubs remains largely constant but only the length of the
links connecting them is significantly reduced since we radically merge links during
this phase. Fig. 3 visualizes the inferred polylines in blue and the actual cartographic
data in grey color.
Fig. 2. Inference of hubs Fig. 3. Compacting links and final inferred polyline
3 Discussion and further work
We briefly presented a polyline-based visualization of eye tracking data that depicts
the “average” cartographic line observed by subjects, along with the algorithm that is
used to infer this polyline. This visualization is useful in cases where the context of
eye tracking has reference to lines, paths, etc. that subjects are required or expected to
follow. One such case is cartography where borders, navigation routes and all kinds of
curves, are used to represent useful information on a map. Studying the effects of
different visualization attributes of cartographic lines in the concentration of the eye’s
attention to a central linear entity can benefit from using the representation of eye
tracking data introduced in this paper. Other applications may relate to the study of
25
ET4S 2014
user interfaces and computer visualizations in general, where following the path of
gaze attention is useful.
The proposed visualization can be further improved by adding color attributes to the
inferred polyline using calculations such as data density of eye tracking samples near
the line, or other statistical metrics. Considering that it is the mind that actually does
the cognitive interpretation of lines observed, it is rather impossible to infer a polyline
that very closely matches the initial cartographic line. However, studying the devia-
tions of individual observers’ tracks from the “average” polyline, and combining the
results with semantics from the experiment and subject context may produce some
interesting results, too.
Application of the proposed visualization in other kinds of lines, whose eye tracking
makes sense, as is the case with some medical images, is another area that is defini-
tively worth exploring. Last but not least, the production algorithm of the polyline
needs further experimentation on bigger data sets and possibly improvement in few
operational aspects. All of the above are future directions of this research.
4 References
1. Bojko, A. Informative or Misleading? Heatmaps Decostruction. In: Jacko, J.A. (Eds.) Hu-
man-Computer Interaction. pp. 30-39, Springer-Verlag, Berlin (2009)
2. J.A.de Bruin, K.M.Malan, and J.H.P.Eloff. Saccade deviation indicators for automated eye
tracking analysis. In: Proceedings of the 2013 Conference on Eye Tracking South Africa,
pp 47-54, Cape Town (2013).
3. Bargiota, T., Mitropoulos, V., Krassanakis, V., Nakos, B.: Measuring locations of critical
points along cartographic lines. In: Proceedings of the 26th International Cartographic-
Conference, Dresden (2013)
4. Zhang, T., Ramakrishnan, R., Livny, M. Birch: An efficient data clustering method for
very large databases. In: Proceedings of the 1996 ACM SIGMOD International Confer-
ence on Management of Data, pp 103-114, Montreal, CA (1996)
5. Ester, M., Kriegel, H.P., S, J., Xu, X. A density-based algorithm for discovering clusters in
large spatial databases with noise. In: Proceedings of the 2nd International Conference on
Knowledge Discovery and Data Mining, pp 226-231, Portland, OR (1996)
6. Edelkamp, S. and Schroedl, S. Route planning and map inference with global positioning
traces, In: Computer Science in Perspective, Springer-Verlag, pp 128-151, New York
(2003)
7. Biagioni, J. and Eriksson, J. Map inference in the face of noise and disparity. In: Proceed-
ings of the 20th ACM SIGSPATIAL GIS International Conference, pp 79-88, Redondo
Beach, CA (2012)
8. Ahmed, M. and Wenk, C. Constructing street networks from gps trajectories. In: Proceed-
ings of the 20th Annual European Symposium on Algorithms, pp 60-71, Ljubljana, Slove-
nia (2012)
9. Karagiorgou, S. and Pfoser, D. On vehicle tracking data-based road network generation.
In: Proceedings of the 20th ACM SIGSPATIAL GIS International Conference, pp 89-98,
Redondo Beach, CA (2012)
26