Identifying Fixations and Saccades in Eye-Tracking Protocols

Identifying Fixations and Saccades in
Eye-Tracking Protocols
Dario D. Salvucci

Joseph H. Goldberg

Nissan Cambridge Basic Research
Four Cambridge Center
Cambridge, MA 02142 USA
+1 617 374 9669

Dept. of Industrial and Manufacturing Engineering
Pennsylvania State University
310 Leonhard Building
University Park, PA 16802 USA
+1 814 863 2370

dario@cbr.com

jgoldberg@psu.edu
ABSTRACT
The process of fixation identification—separating and labeling
fixations and saccades in eye-tracking protocols—is an
essential part of eye-movement data analysis and can have a
dramatic impact on higher-level analyses.
However,
algorithms for performing fixation identification are often
described informally and rarely compared in a meaningful way.
In this paper we propose a taxonomy of fixation identification
algorithms that classifies algorithms in terms of how they
utilize spatial and temporal information in eye-tracking
protocols. Using this taxonomy, we describe five algorithms
that are representative of different classes in the taxonomy and
are based on commonly employed techniques. We then
evaluate and compare these algorithms with respect to a number
of qualitative characteristics. The results of these comparisons
offer interesting implications for the use of the various
algorithms in future work.

Keywords
Fixation identification, eye tracking, data analysis algorithms.

1. INTRODUCTION
Eye tracking has been gaining in popularity over the past
decade as a window into observers’ visual and cognitive
processes. For instance, researchers have utilized eye tracking
to study behavior in such domains as image scanning [e.g.,
12], driving [e.g., 11], arithmetic [e.g., 20], analogy [e.g.,
17], and reading [e.g., 14]. In these and other domains,
researchers typically analyze eye movements in terms of
fixations (pauses over informative regions of interest) and
saccades (rapid movements between fixations). Common
analysis metrics include fixation or gaze durations, saccadic
velocities, saccadic amplitudes, and various transition-based
parameters between fixations and/or regions of interest.

The analysis of fixations and saccades requires some form of
fixation identification (or simply identification)—that is, the
translation from raw eye-movement data points to fixation
locations (and implicitly the saccades between them) on the
visual display. Fixation identification significantly reduces
the size and complexity of the eye-movement protocol,
removing raw saccade data points and collapsing raw fixation
points into a single representative tuple. This reduction is
useful for at least two reasons. First, little or no visual
processing can be achieved during a saccade [6], and thus the
actual paths traveled during saccades are typically irrelevant for
many research applications. Second, smaller eye movements
that occur during fixations, such as tremors, drifts, and flicks
[1, 4], often mean little in higher-level analyses. Thus,
fixation identification is a convenient method of minimizing
the complexity of eye-tracking data while retaining its most
essential characteristics for the purposes of understanding
cognitive and visual processing behavior.
Fixation identification is an inherently statistical description
of observed eye movement behaviors. While it is generally
agreed upon that visual and cognitive processing do occur
during fixations [e.g., 9], it is less clear exactly when fixations
start and when they end. Thus, regardless of the precision and
flexibility associated with identification algorithms, the
identification problem is still a subjective process. Therefore
one efficient way to validate these algorithms is to compare
resultant fixations to an observer’s subjective impressions.
Though widely employed, fixation identification and its
various implementations have often been given short shrift in
the eye-movement literature, particularly in literature
concerning the interaction of eye movements and higher-level
cognition. However, identification is often a critical aspect of
eye-movement data analysis that can have significant effects
on later analyses. For instance, poorly defined identification
algorithms may produce too many or too few fixations, or may
be overly sensitive to outlier data points, thus biasing
interpretation. Karsh and Breitenbach [10] provided a rigorous
demonstration of how different identification algorithms can
produce vastly different interpretations even when analyzing
identical protocols. Thus, good identification algorithms
ensure valid fixation and saccade locations and durations which
in turn influence inferences such as processing complexity and
visual search paths.

Salvucci, D. D., & Goldberg, J. H. (2000). Identifying fixations and saccades in eye-tracking protocols. In Proceedings of the Eye
Tracking Research and Applications Symposium (pp. 71-78). New York: ACM Press.

To demonstrate the usefulness of the taxonomy, we identify and
describe five algorithms that are representative of different
classes in the taxonomy.
We first provide rigorous
descriptions of these algorithms in “pseudocode” form to best
illustrate the similarities and differences between the
algorithms. We then evaluate and compare the methods with
respect to several qualitative characteristics including
algorithm speed, accuracy, robustness, parameter space, and
ease of implementation. We conclude with a discussion of
implications and recommendations for how the various
methods can be best utilized for future applications.

2. A TAXONOMY OF FIXATION
IDENTIFICATION ALGORITHMS
Our taxonomy classifies fixation identification algorithms
with respect to spatial and temporal characteristics, as
summarized in Table 1. In constructing this taxonomy, we
attempted to identify a minimal set of criteria that would best
capture the differences between common existing algorithms.
While we could propose a more complex taxonomy that could
potentially account for finer distinctions and/or hybrid
algorithms, the proposed basic taxonomy provides a useful
characterization of the primary types of identification
algorithms. The taxonomy will thus serve as a good starting
point for comparing and evaluating existing identification
algorithms, as we discuss in the following two sections.
For spatial characteristics, we identify three criteria that
distinguish three primary types of algorithms: velocity-based,
dispersion-based, and area-based. Velocity-based algorithms
emphasize the velocity information in the eye-tracking
protocols, taking advantage of the fact that fixation points
have low velocities and saccade points have high velocities.
(Note that, assuming a constant sampling rate, velocities are
simply distances between sampled points and thus we can
ignore the temporal component implicit in velocities.)
Dispersion-based algorithms emphasize the dispersion (i.e.,
spread distance) of fixation points, under the assumption that

X
X

Area-based

X

Duration sensitive
Locally adaptive

I-AOI

X

Dispersion-based

I-MST

X

I-DT

Velocity-based

I-HMM

Criteria

I-VT

Representative Algorithms

Spatial

In this paper we address these problems by proposing a novel
taxonomy of fixation identification algorithms and evaluating
representative algorithms in the context of this taxonomy.
The taxonomy classifies algorithms with respect to five spatial
and temporal criteria. The spatial criteria divide algorithms in
terms of their use of velocity, dispersion, and area-of-interest
information. The temporal criteria divide algorithms in terms
of their use of duration information and their local adaptivity.
The taxonomy provides a meaningful labeling and
classification of existing algorithms so that they may be more
easily compared in a systematic way to guide the choice of
algorithms for particular applications.

Table 1 : T a x o n o m y o f f i x a t i o n i d e n t i f i c a t i o n
algorithms, sample citations for their use, and
the sample algorithms presented here.

Temporal

In addition to the frequent underspecification of identification
algorithms, little work has been done in evaluating and
comparing different possible algorithms. Researchers and
practitioners often have minimal information to guide their
decision of which algorithm to use in particular situations. Not
surprisingly, this problem leads to the somewhat haphazard
application of different algorithms, making it difficult to
compare results derived by different methods of identification.

X
X

X

X
X

fixation points generally occur near one another. Area-based
algorithms identify points within given areas of interest
(AOIs) that represent relevant visual targets.
These
algorithms, unlike the others described here, provide both
lower-level identification and higher-level assignment of
fixations to AOIs [see 15, 16]. Because fixations can also be
used as inputs to AOI algorithms, these can also represent
higher levels of attentional focus on a display. These dwell
times can be considered ‘macro-fixations’, in that they
organize fixations into a larger picture.
For temporal characteristics, we include two criteria: whether
the algorithm uses duration information, and whether the
algorithm is locally adaptive. The use of duration information
is guided by the fact that fixations are rarely less than 100 ms
and often in the range of 200-400 ms. The incorporation of
local adaptivity allows the interpretation of a given data point
to be influenced by the interpretation of temporally adjacent
points; this is useful, for instance, to compensate for
differences between ‘steady-eyed’ individuals and those who
show large and frequent eye movements.

3. REPRESENTATIVE ALGORITHMS
To enable us to rigorously compare the classes of algorithms in
our taxonomy, it is convenient to formalize sample algorithms
that serve to represent the essential ideas embodied by each
class of algorithms. This section describes five algorithms—
I-VT, I-HMM, I-DT, I-MST, and I-AOI—that extract the
important aspects of numerous existing algorithms and express
their basic techniques as simply as possible. Table 1 shows
how the representative algorithms map onto the proposed
taxonomy.
Of course, there are many other possible
algorithms that we could include in this analysis, such as
hybrid algorithms with multiple heuristics; however, due to
space constraints, we limit this exposition to five algorithms
with fairly different approaches and implementations.

3.1 Velocity-Based Algorithms
3.1.1 Velocity-Threshold Identification (I-VT)
Velocity-threshold fixation identification (I-VT) is the
simplest of the identification methods to understand and
implement. I-VT is a velocity-based method that separates
fixation and saccade points based on their point-to-point
velocities [e.g., 5, 18]. The velocity profiles of saccadic eye
movements show essentially two distributions of velocities:
low velocities for fixations (i.e., <100 deg/sec), and high
velocities (i.e., >300 deg/sec) for saccades. This aspect of
saccadic eye movements makes velocity-based discrimination
fairly straightforward and robust. While it is possible to create
a locally adaptive velocity-based fixation identification
algorithm, velocity profiles have strong physical and
physiological underpinnings, and thus static criteria are
usually sufficient here.
I-VT begins by calculating point-to-point velocities for each
point in the protocol. Each velocity is computed as the
distance between the current point and the next (or previous)
point. I-VT then classifies each point as a fixation or saccade
point based on a simple velocity threshold: if the point’s
velocity is below threshold, it becomes a fixation point,
otherwise it becomes a saccade point. The process then
collapses consecutive fixation points into fixation groups and
discards saccade points. Finally, I-VT translates each fixation
group to a representation <x ,y ,t,d> using the centroid (i.e.,
center of mass) of the points as x and y , the time of the first
point as t, and the duration of the points as d. Pseudocode for
the I-VT method is shown in Table 2.

3.1.2 HMM Identification (I-HMM)
Hidden Markov model fixation identification (I-HMM) uses
probabilistic analysis to determine the most likely
identifications for a given protocol [15, 16]. Hidden Markov
models (HMMs) are probabilistic finite state machines that
have been employed extensively in the fields of speech and
handwriting recognition [see 13]. I-HMM uses a two-state
HMM in which the states represent the velocity distributions
for saccade and fixation points.
This probabilistic
representation helps
I-HMM perform
more
robust
identification than a fixed-threshold method such as I-VT.
The crux of I-HMM is the two-state HMM shown in Figure 1.
The HMM includes two sets of probabilities: observation and
transition probabilities. The observation probabilities for
each state (the small v distributions) represent the distribution
of expected velocities in that state. The first state represents
saccade points, and thus contains a distribution centered around
higher velocities; the second state represents fixation points,
and thus contains a distribution centered around lower
velocities. The transition probabilities for each state (the
arrows exiting the states) represent the likelihood of remaining
in the state or making a transition to another state. The
transition probabilities in Figure 1 show a large likelihood of
remaining in each state (.95) and a small likelihood or making
a transition (.05). Thus, the HMM provides a probabilistic
representation of the observations (i.e., velocities) generated
during saccadic eye movements.

Table 2: Pseudocode for the I-VT algorithm.
I-VT (protocol, velocity threshold)
Calculate point-to-point velocities for each
point in the protocol
Label each point below velocity threshold as
a fixation point, otherwise as a saccade
point
Collapse consecutive fixation points into
fixation groups, removing saccade points
Map each fixation group to a fixation at the
centroid of its points
Return fixations
I-VT requires the specification of one parameter, the velocity
threshold. If angular velocities can be computed (i.e., the
distance from eye to visual stimuli is known), the point-topoint velocity threshold can be approximated from a
reasonable angular velocity threshold [6]; for instance, Sen and
Megaw [18] used a threshold of 20 degrees/second. However,
when only point-to-point velocities are known, an appropriate
value for the velocity threshold may need to be inferred based
on aspects of data collection (e.g., sampling frequency) along
with some exploratory data analysis.

Figure 1 : Sample t w o - s t a t e HMM.
The f i r s t
state represents higher-velocity saccade points;
t h e s e c o n d s t a t e represents l o w e r - v e l o c i t y
fixation points.
Given the two-state HMM, I-HMM determines the most likely
identification of each protocol point through a process of
decoding. HMM decoding finds an assignment of protocol
points to states that maximizes the probability of the protocol
given the HMM, using dynamic programming [13] to find this
optimal assignment efficiently. The assignment associates
each point with a state, thus providing an identification of each
point as a fixation or saccade point. Like I-VT, I-HMM then
collapses consecutive fixation points into groups and outputs
fixations at the centroid of the groups. Table 3 presents
pseudocode for this I-HMM algorithm.
The parameters of I-HMM comprise the observation and
transition parameters in the two-state HMM: two observation
probability parameters (the mean and variance of the
distribution) and two transition probabilities for each state, for
a total of eight parameters. While this parameter space is

Table 3: Pseudocode for the I-HMM algorithm.
I-HMM (protocol, HMM)
Calculate point-to-point velocities for each
point in the protocol
Decode velocities with two-state HMM to
identify points as fixation or saccade points
Collapse consecutive fixation points into
fixation groups, removing saccade points
Return fixations
Table 4: Pseudocode for the I-DT algorithm.
I-DT (protocol, dispersion threshold,
duration threshold)
While there are still points
Initialize window over first points to
cover the duration threshold
If dispersion of window points <=
threshold
Add additional points to the window
until dispersion > threshold
Note a fixation at the centroid of the
window points
Remove window points from points
Else
Remove first point from points
Return fixations
certainly more complex than that of I-VT, the parameters of IHMM can be learned through a procedure called reestimation
[13]. Given a training set (i.e., a set of protocols) reestimation
learns the values of HMM parameters that maximize the
probability of the training set given the HMM. Thus, with
protocol data from a particular eye-tracking setup, reestimation
can provide parameter values that, once estimated, can be used
for any protocols collected in that setup.

3.2 Dispersion-Based Algorithms
3.2.1 Dispersion-Threshold Identification (I-DT)
In contrast to the velocity-based identification of I-VT and IHMM, dispersion-threshold identification (I-DT) utilizes the
fact that fixation points, because of their low velocity, tend to
cluster closely together. I-DT identifies fixations as groups of
consecutive points within a particular dispersion, or maximum
separation [e.g., 19, 21]. Because fixations typically have a
duration of at least 100 ms, dispersion-based identification
techniques often incorporate a minimum duration threshold of

100-200 ms [21] to help alleviate equipment variability. The
following I-DT algorithm is based on Widdel’s [21] data
reduction algorithm.
The I-DT algorithm uses a moving window that spans
consecutive data points checking for potential fixations. The
moving window begins at the start of the protocol and initially
spans a minimum number of points, determined by the given
duration threshold and sampling frequency. I-DT then checks
the dispersion of the points in the window by summing the
differences between the points’ maximum and minimum x and y
values; in other words, dispersion D = [max(x ) – min(x )] +
[max(y ) – min(y )]. Note that alternative dispersion metrics
could be based upon spatial variance or area of samples. If the
dispersion is above the dispersion threshold, the window does
not represent a fixation, and the window moves one point to
the right. If the dispersion is below the dispersion threshold,
the window represents a fixation. In this case, the window is
expanded (to the right) until the window’s dispersion is above
threshold. The final window is registered as a fixation at the
centroid of the window points with the given onset time and
duration. This process continues with window moving to the
right until the end of the protocol is reached. Table 4 includes
pseudocode for the I-DT algorithm.
We can note that this characterization of fixations uses the
centroid and diameter. A circular area is usually assumed, and
the mean distance from each sample to the fixation centroid
provides an estimate of the radius. Also, dispersion-based
algorithms are sometimes used to locate clusters within
minimum spanning tree network representations (see the
following section). The graph forms an efficient framework for
rapid search of large sample sets.
The I-DT algorithm requires two parameters, the dispersion
threshold and the duration threshold. Like the velocity
threshold for I-VT, the dispersion threshold can be set to
include 1/2° to 1° of visual angle if the distance from eye to
screen is known. Otherwise, the dispersion threshold can be
estimated from exploratory analysis of the data. The duration
threshold is typically set to a value between 100 and 200 ms
[21], depending on task processing demands.
Note that this characterization of I-DT is conceptually different
from some dispersion-based methods that center on clustering
algorithms. For instance, the leader algorithm and the k-means
algorithm are two such methods [8]. However, most of these
algorithms require an initial estimate of the number of clusters
in the protocol, making them less useful for eye-movement
data, where the number of fixations is typically unknown.

3.2.2 MST Identification (I-MST)
MST identification (I-MST) is based on minimum spanning
trees (MSTs) — that is, a tree connecting a set of points such
that the total length of the tree’s line segments is minimized.
MSTs can provide a highly flexible and controllable
representation for dispersion-based fixation identification [7].
A two-step approach is necessary, first requiring the
construction of the MST followed by a search of the MST.
Construction uses Prim’s algorithm [2]. There is one and only
one MST for a set of points. The advantage of an MST data

Table 5: Pseudocode for the I-MST algorithm.
I-MST (protocol, edge ratio, edge sd)
Construct MST from protocol data points using
Prim’s algorithm
Find the maximum branching depth for each MST
point using a depth-first search
Identify saccades as edges whose distances
exceed predefined criteria
Define the parametric properties (µ, σ) of
local edges, identifying saccades when an
edge length exceeds a defined ratio
Identify fixations as clusters of points not
separated by saccades
Return fixations
Table 6: Pseudocode for the I-AOI algorithm.
I-AOI (protocol, duration threshold,
target areas)
Label each point as a fixation point for the
target area in which it lies, or as a saccade
point if none
Collapse consecutive fixation points for the
same target into fixation groups, removing
saccade points
Remove fixation groups that do not span the
minimum duration threshold
Return fixations
representation lies in the degree of control, flexibility, and
local adaptation for dispersion analysis, as well as improved
subsequent characterizations of defined fixations.
Fixation identification requires traversing the already defined
MST, as described in Table 5. A depth-first search is made to
determine the maximum depth of interconnectivity at each
point. Branching depths below a defined setpoint signal
locations near the edge of the MST that are not appropriate
candidates for separating fixations. If the edges connected to
each endpoint exceed the defined minimum branching depth,
then the network of edges connected to each endpoint is
gathered into an associated edge length distribution. The mean
µ and standard deviation σ of edge lengths provide a locally
adaptive comparison for separation of fixations. Separation
can occur based upon comparison of the edge under
consideration of both µ and σ of neighbor edge lengths. In this
way, a framework is achieved for controlling where fixations
may occur in an MST and for determining how local adaptivity
affects fixation decisions.

The MST format allows additional characterization parameters
for fixations. For example, the MST length is defined by the
longest path through the graph. Critical paths can be defined
by areas with minimal branching structure. These and other
parameters allow an estimate of the shape and possibly
direction of fixations; these go beyond simpler centroid and
size characterizations.

3.3 Area-based Algorithms
3.3.1 Area-of-Interest Identification (I-AOI)
The four previous identification methods can identify fixations
at any location in the visual field. In contrast, area-of-interest
fixation identification (I-AOI) identifies only fixations that
occur within specified target areas [e.g., 3]. The target areas are
rectangular regions of interest that represent units of
information in the visual field. These target areas, generally
used in later analyses like tracing, keep identified fixations
close to relevant targets. I-AOI also utilizes a duration
threshold to help distinguish fixations in target areas from
passing saccades in those areas.
I-AOI begins by associating data points with target areas: it
labels points within a target area as a fixation point for that
target, and labels points outside of all target areas as saccades.
I-AOI then collapses consecutive fixation points for the same
target into fixation groups, discarding saccade points. Finally,
it removes fixation groups that fall below a given duration
threshold and transforms each fixation group into a fixation
tuple. The pseudocode for this algorithm appears in Table 6.
As explained earlier, I-AOI can provide a larger picture than
fixations alone. Like a fixation, dwell time within an area has
a starting time and an ending time. However, since fixations
can serve as the input data into AOI determination, the time
between these dwells do not describe saccadic behavior.
Rather, there could be multiple saccades and fixations
interspersed between AOIs.
While AOI-based dwell-time
algorithms are not fixation algorithms per se, they are valuable
as a concept to help explain higher-level collections of
fixations organized about visual targets and areas.

4. EVALUATION AND COMPARISON
To evaluate and compare the various identification methods, we
now consider each method with respect to several
characteristics: interpretation speed, accuracy, robustness, ease
of implementation, and parameter setting. Table 7 shows a
summary of the described methods according to these
characteristics. To illustrate some of the issues discussed, we
utilize a sample eye-movement protocol from a task in which
students encoded four equation values and performed some
simple computation [see 15, 16], as shown in Figure 2. Each
point represents the student’s current point-of-regard sampled
at a rate of 60 Hz. Points that represent fixations, as
interpreted by the various identification algorithms, are drawn
larger than those that represent saccades. Also, points that
occur earlier in the protocol time sequence are drawn darker than
those that occur later. The fixations are numbered to facilitate
understanding of the protocol time sequence.

Table 7: Summary of identification methods.
Method

Accuracy

Speed

Robustness

Impl. Ease

Parameters

√

√√

×

√√

1

Hidden Markov Model (I-HMM)

√√

√

√√

√ / ×a

8/0

Dispersion Threshold (I-DT)

√√

√

√√

√

2

Minimum Spanning Tree (I-MST)

√

×

√√

×

2

Area-of-Interest (I-AOI)

×

√

√

√

1+ b

Velocity Threshold (I-VT)

a

Key: √√ = very good, √ = good, × = not as good.
a
I-HMM has 8 parameters in its two-state HMM that can be an be learned through reestimation. Without reestimation, I-HMM has 8
parameters but is simpler to implement; with reestimation, I-HMM effectively has no parameters but is more difficult to implement.
b
I-AOI has 1 parameter but also requires specification of target areas.

4.1 Velocity-Based Methods

4.2 Dispersion-Based Methods

I-VT is straightforward to implement, runs very efficiently, and
can easily run in real time. However, it sometimes encounters
problems when point velocities hover near threshold because
of eye-tracker noise or time-averaged data. Specifically, when
point velocities are near threshold (e.g., midway between
saccades and fixations), the strict threshold can result in
“blips” in identification—fixation or saccade groups with only
one or a few consecutive points. The protocol in Figure 2(a)
illustrates this problem: fixations 3-6 are interpreted as four
separate fixations, whereas it seems more likely that there are
only one or two actual fixations represented in these data
points. Similarly, fixations 11-13 seem to represent a single
fixation. The problem is especially prevalent when analyzing
time-averaged data, such as those in the figure. Researchers
have alleviated the problem in the past by aggregating
consecutive fixations over a single target into gazes on that
target [9], as shown in Figure 2(b). Other work required
minimum durations for either fixations or saccades; for
instance, Sen and Megaw [18] required that saccade points
remain above a velocity threshold for at least 10 ms.

I-DT is a linear-time, potentially real-time algorithm that
produces robust identification results. For the sample protocol
in Figure 2(a), it generates an almost identical analysis as IHMM (as shown in Figure 2(c)), avoiding the “blip” problem
because of its duration threshold. The primary disadvantage of
I-DT is the use of two parameters that are highly
interdependent; for instance, a small dispersion threshold with
a large duration threshold may not result in any identified
fixations. Thus, I-DT requires the most careful parameter
setting but provides good results for a fairly simple algorithm.

I-HMM employs a probabilistic model rather than a fixed
velocity threshold and utilizes the sequential information in
the protocol. Thus, I-HMM allows for more freedom in
identifying points and provides more robust analysis than IVT, especially in the presence of significant noise. Figure 2(c)
illustrates the same protocol as in Figure 1 analyzed using IHMM; the problematic fixations that I-VT separated into
multiple fixations (see Figure 2(a)) are grouped into more
plausible fixations in Figure 2(c). Another benefit of I-HMM
is that the two-state HMM can be embedded into a larger
cognitive model for even more robust interpretation of eye
movements [15, 16]. In addition, I-HMM, like I-VT, runs
efficiently in linear time [see 15, 16] and can also run in real
time. The primary disadvantage of I-HMM is the difficulty of
implementing the reestimation procedure, which is both
complex and tedious.
However, if no reestimation of
parameters is needed, the I-HMM algorithm alone (including
HMM decoding) is fairly straightforward to implement.

I-MST has a somewhat different flavor than the other
algorithms in the sense that it forms clusters of fixation points
without using sequential information in the protocols. Thus, it
is more difficult to compare with the other algorithms. I-MST
allows for control of parameter settings that define a saccade by
varying the relative importance of the edge length ratio and
mean/variance tests for clustering. Its speed is exponential in
that it becomes exponentially slower as points are added: it
requires multiple traversals over the MST for each point, one to
establish local edge distribution characteristics and one to test
each potential cut edge. It is also more difficult to make I-MST
a real-time algorithm because of its global analysis of existing
data points. However, I-MST also seems to provide additional
means of characterizing fixations beyond spatial location and
size; for instance, the MST diameter provides the longest MST
path within each fixation. Also, characteristics of the MST,
such as the ratio of path length to width, can provide
information on the shape of the defined fixations. Branching
depth provides an additional means of assessing fixation size.

4.3 Area-Based Methods
I-AOI runs efficiently in linear time or real time. However, it
suffers from a somewhat serious problem: the inclusion of
many saccade points in identified fixations. Because any point
within a target area classifies as a fixation point, saccade
points leading into a fixation are often included in the final
fixation. This problem can result in deceptively long durations
for identified fixations. An even more problematic situation

(a)

(b)

(c)

(d)
Figure 2 : F i x a t i o n i d e n t i f i c a t i o n u s i n g (a) I-VT, ( b ) I-VT w i t h f i x a t i o n s c o l l a p s e d i n t o g a z e s ,
(c) I-HMM or I-DT, and (d) I-AOI. Eye-movement protocols are taken from an equation-solving task in
which a student encodes the four equation values and computes a result [see 15, 16]. Larger points
represent f i x a t i o n s w h i l e s m a l l e r p o i n t s represent saccades ( a s interpreted b y t h e i d e n t i f i c a t i o n
algorithm). Darker points occur earlier while lighter points occur later in the protocol.
arises when long saccades through large target areas are
identified as fixations, especially when the data include some
form of time averaging. Figure 2(d) illustrates these problems:
fixation 1 includes points that likely represent a saccade (the
upper grouping of points), whereas fixation 4 fails to identify
apparent fixation points because they lie outside the target
areas used. Overall, I-AOI can provide reasonable results [e.g.,
3] for certain aggregate analyses but is generally less
preferable than the other, more robust algorithms.

5. SUMMARY AND CONCLUSIONS
To summarize our qualitative analysis, I-HMM and I-DT provide
accurate and robust fixation identification by incorporating
sequential information to aid in interpretation. I-MST also

provides robust identification but runs slower than all other
algorithms. I-VT has the simplest algorithm as thus the
smallest computational overhead; however, it can experience
severe “blip” effects when analyzing at the level of fixations
rather than gazes. I-AOI performs rather poorly on all fronts
for the purpose of identification and is best not used.
These results offer several implications for future use of these
and related algorithms. First, velocity-based and dispersionbased algorithms both fare well and provide approximately
equivalent performance. However, area-based algorithms are
too restrictive and can generate deceptive results that bias later
analyses. Second, the use of temporal information can greatly
facilitate fixation identification of protocols. The three

algorithms that utilize this information in a locally adaptive
way, namely I-HMM, I-DT, and I-MST, generate robust
interpretations even in the presence of eye-tracking equipment
noise. Third, in support of work such as Karsh and Breitenbach
[10], the choice of identification algorithms can dramatically
affect the resulting identified fixations. By describing their
chosen algorithm in terms of the proposed taxonomy,
researchers can better communicate their analysis techniques to
others and thus provide more standardized and understandable
results.
As the next step in evaluating and comparing fixation
identification algorithms, we hope to generate more
quantitative measures for comparison and examine how the
various algorithms affect later analyses in the context of
different domain applications. For instance, some preliminary
work has shown that a subset of the algorithms described here
can affect the interpretation of eye-tracking protocols with
respect to the predictions of a cognitive process model [15,
16]. This increased understanding of possible identification
algorithms will help to facilitate the use and analysis of eyetracking data for both scientific research and the development
of eye-based systems and applications.

REFERENCES
[1] Alpern, M. (1962). Type of movement. In H. Davson
(Ed.), The Eye, Vol. 3. (pp. 63-151). New York: Academic
Press.

& M. A. Just (Eds.), New Methods in Reading
Comprehension Research (pp. 151-182). Hillsdale, NJ:
Erlbaum.
[10] Karsh, R., & Breitenbach, F. W. (1983). Looking at
looking: The amorphous fixation measure. In R. Groner,
C. Menz, D. F. Fisher, & R. A. Monty (Eds.), Eye
Movements and Psychological Functions: International
Views (pp. 53-64). Hillsdale, NJ: Erlbaum.
[11] Land, M. F., & Lee, D. N. (1994). Where we look when we
steer. Nature, 369, 742-744.
[12] Noton, D., & Stark, L. (1971). Scanpaths in saccadic eye
movements while viewing and recognizing patterns.
Vision Research, 11 , 929-942.
[13] Rabiner, L. R. (1989). A tutorial on hidden Markov
models and selected applications in speech recognition.
Proceedings of the IEEE, 77 , 257-286.
[14] Rayner, K. (1995).
Eye movements and cognitive
processes in reading, visual search, and scene perception.
In J. M. Findlay, R. Walker, & R. W. Kentridge (Eds.), Eye
Movement Research: Mechanisms, Processes, and
Applications (pp. 3-21). New York: Elsevier.
[15] Salvucci, D. D. (1999). Mapping eye movements to
cognitive processes. Doctoral Dissertation, Department
of Computer Science, Carnegie Mellon University.

[2] Camerini, P. M., Galbiati, G., & Maffioli, F. (1988).
Algorithms for finding optimum trees: Description, use
and evaluation. Annals of Operations Research, 13 , 265397.

[16] Salvucci, D. D., & Anderson, J. R. (1998). Tracing eye
movement protocols with cognitive process models. In
Proceedings of the Twentieth Annual Conference of the
Cognitive Science Society (pp. 923-928). Hillsdale, NJ:
Erlbaum.

[3] DenBuurman, R., Boersma, T., & Gerrisen, J. F. (1981).
Eye movements and the perceptual span in reading.
Reading Research Quarterly, 16 , 227-235.

[17] Salvucci, D. D., & Anderson, J. R. (in press). Integrating
analogical mapping and general problem solving: The
path-mapping theory. Cognitive Science.

[4] Ditchburn, R. W. (1980). The function of small saccades.
Vision Research, 20 , 271-272.

[18] Sen, T., & Megaw, T. (1984). The effects of task variables
and prolonged performance on saccadic eye movement
parameters.
In A. G. Gale & F. Johnson (Eds.),
Theoretical and Applied Aspects of Eye Movement
Research (pp. 103-111). Amsterdam: Elsevier.

[5] Erkelens, C. J., & Vogels, I. M. L. C. (1995). The initial
direction and landing position of saccades. In J. M.
Findlay, R. Walker, & R. W. Kentridge (Eds.), Eye
Movement Research: Mechanisms, Processes, and
Applications (pp. 133-144). New York: Elsevier.
[6] Fuchs, A. F. (1971). The saccadic system. In P. Bach-yRita, C. C. Collins, & J. E. Hyde (Eds.), The Control of
Eye Movements (pp. 343-362). New York: Academic
Press.
[7] Goldberg, J. H., & Schryver, J. C. (1995).
contingent control of the
computer
Methodology and example for zoom detection.
Research Methods, Instruments, & Computers,
350.

Eye-gazeinterface:
Behavior
27 , 338-

[8] Hartigan, J. A. (1975). Clustering algorithms.
York: John Wiley & Sons.

New

[9] Just, M. A., & Carpenter, P. A. (1984). Using eye
fixations to study reading comprehension. In D. E. Kieras

[19] Stark, L., & Ellis, S. R. (1981). Scanpath revisited:
Cognitive models of direct active looking. In D. F.
Fisher, R. A. Monty, & J. W. Senders (Eds.), Eye
Movements: Cognition and Visual Perception (pp. 193226). Hillsdale, NJ: Erlbaum.
[20] Suppes, P. (1990). Eye-movement models for arithmetic
and reading performance.
In E. Kowler (Ed.), Eye
Movements and their Role in Visual and Cognitive
Processes (pp. 455-477). New York: Elsevier.
[21] Widdel, H. (1984). Operational problems in analysing
eye movements. In A. G. Gale & F. Johnson (Eds.),
Theoretical and Applied Aspects of Eye Movement
Research (pp. 21-29). New York: Elsevier.

Identifying Fixations and Saccades in Eye-Tracking Protocols

More Related Content

Similar to Identifying Fixations and Saccades in Eye-Tracking Protocols (20)

Recently uploaded (20)

Identifying Fixations and Saccades in Eye-Tracking Protocols