Systematic Review Chart
stellagirl
2017.Eyegaze.Areviewandanalysisof.Kar..pdf
Received June 21, 2017, accepted July 23, 2017, date of publication August 7, 2017, date of current version September 6, 2017.
Digital Object Identifier 10.1109/ACCESS.2017.2735633
A Review and Analysis of Eye-Gaze Estimation Systems, Algorithms and Performance Evaluation Methods in Consumer Platforms ANURADHA KAR, (Student Member, IEEE), AND PETER CORCORAN, (Fellow, IEEE) Center for Cognitive, Connected, and Computational Imaging, Department of Electrical and Electronic Engineering, National University of Ireland, Galway, Ireland
Corresponding author: Anuradha Kar ([email protected])
This work was supported in part by the Science Foundation Ireland through the Strategic Partnership Program and in part by FotoNation Ltd., Next Generation Imaging for Smartphone and Embedded Platforms, under Project 13/SPP/I2868.
ABSTRACT In this paper, a review is presented for the research on eye gaze estimation techniques and applications, which has progressed in diverse ways over the past two decades. Several generic eye gaze use-cases are identified: desktop, TV, head-mounted, automotive, and handheld devices. Analysis of the literature leads to the identification of several platform specific factors that influence gaze tracking accuracy. A key outcome from this review is the realization of a need to develop standardized methodologies for the performance evaluation of gaze tracking systems and achieve consistency in their specification and comparative evaluation. To address this need, the concept of a methodological framework for practical evaluation of different gaze tracking systems is proposed.
INDEX TERMS Eye gaze, gaze estimation, accuracy, error sources, performance evaluation, user platforms.
I. INTRODUCTION Advances in eye gaze tracking technology over the past few decades have led to the development of promising gaze estimation techniques and applications for human computer interaction. Historically, research on gaze tracking dates back to the early 1900s, starting with invasive eye track- ing techniques. These included electro-occulography using pairs of electrodes placed around the eyes or the scleral search methods that include coils embedded into a contact lens adhering to the eyes. The first video based eye track- ing study was made on pilots operating airplane controls in the 1940s [1]. Research on head-mounted eye trackers advanced in the 1960s and gaze tracking developed further in the 1970s with focus on improving accuracy and reducing the constraints on users. With increasing computing power in devices, real time operation of eye trackers became pos- sible during the 1980s. However till this time, owing to limited availability of computers, eye tracking was mainly limited to psychological and cognitive studies and medi- cal research. The application focus towards general purpose human computer interaction was sparse. This changed in the 1990s as eye gaze found applications in computer input and control [2]. Post 2000, rapid advancements in comput- ing speed, digital video processing and low cost hardware
brought gaze tracking equipment closer to users, with appli- cations in gaming, virtual reality and web-advertisements [3].
Eye gaze information is used in a variety of user platforms. The main use cases may be broadly classified into (i) desk- top computers [4]–[6], (ii) TV panels [7], [8], (iii) head mounted [9]–[12] (iv) automotive setups [13]–[17] (v) hand- held devices [18], [19]. Applications based on desktop plat- forms involve using eye gaze for computer communication and text entry, computer control and entering gaze based passwords [20]. Remote eye tracking has recently been used on TV panels to achieve gaze controlled functions, for exam- ple selecting and navigating menus and switching chan- nels. Head-mounted gaze tracking setups usually comprise of two or more cameras mounted on a support framework worn by the user. Such systems have been extensively employed in user attention and cognitive studies, psychoanalysis, occulo- motor measurements [2], virtual and augmented reality appli- cations [21], [22]. Real time gaze and eye state tracking on automotive platforms is used in driver support systems to evaluate driver vigilance and drowsiness levels. These use eye tracking setups mounted on a car’s dashboard along with computing hardware running machine vision algorithms. In handheld devices such as smartphones or tablets, the front camera is used to track user gaze to activate functions such
VOLUME 5, 2017 2169-3536 2017 IEEE. Translations and content mining are permitted for academic research only.
Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
16495
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
as locking/unlocking phones, interactive displays, dimming backlights or suspending sensors [18], [23].
Within each of these use cases there exists a wide range of system configurations, operating conditions and varying quality of imaging and optical components. Furthermore, the variations in eye-movement and biological aspects of individuals lead to challenges in achieving consistent and repeatable performance from gaze tracking methods. Thus, despite several decades of development in eye gaze research, performance evaluation and comparison of different gaze estimation techniques across different platforms is a still a difficult task [24], [25].
In order to provide insight into the current status of eye gaze research and outcomes, this paper presents a detailed literature review and analysis that considers algorithms, sys- tem configuration, user conditions and performance issues for existing gaze tracking systems. Specifically, use-cases based on five different eye gaze platforms are considered.
The aim of this work is to gain a realistic overview of the diversity currently existing in this field and to identify the factors that affect the practical usability of gaze tracking systems. Further, this review highlights the need for develop- ing standardized measurement protocols to enable evaluation and comparison of the performance and operational charac- teristics of different gaze tracking systems. In this paper, first the diversity and standardization issues in different aspects of eye gaze research are discussed and then the idea of a per- formance evaluation framework is proposed. This framework includes several planned and ongoing experiments that are aimed at practical evaluation of any gaze tracker. Our goal is to encourage further discussion and additional contributions from researchers in this field.
There have been detailed review works on eye gaze made in the last few years such as [3] and [26]–[28] which discussed about recent developments in gaze tracking methods, com- paring different estimation techniques, setups, applications and challenges involved in using gaze as an input modality. Hansen and Ji [26] provides an in-depth review on differ- ent eye models, eye detection techniques and models for gaze estimation, along with a summary of gaze applica- tions. It also discusses inaccuracies in gaze tracking arising from the eye model components, jitter and refraction due to user wearing glasses. However, our work differs on several grounds from these reviews. Firstly, our review is specifi- cally aimed towards highlighting the issues affecting realistic performance evaluation of gaze tracking systems, such as ambiguous accuracy metrics and un-accounted error sources. Secondly, we do a detailed classification of four different eye gaze research platforms. Extensive literature resources are collected and analyzed for each platform with the aim to understand the factors that affect the performance of a gaze based system in each of these. Thirdly, we present our survey in a statistical format that shows the lack of standardization in eye gaze research as a quantitative observation. Also our survey includes research works published until 2017 to make it exhaustive and up-to-date. Finally, our review not only
provides an overview of the current status of eye gaze research but also forms the foundation of a performance evaluation framework for eye gaze systems which is proposed by us in Section VI of this work and is currently under development.
The paper is organized as follows: Section II presents a brief overview on eye movements, gaze tracking sys- tems and accuracy measures used in contemporary gaze research. In Sections III and IV, several gaze tracking algo- rithms are categorized and key research works on the imple- mentation of gaze tracking in five different user platforms are reviewed. In Section V, the factors limiting practical performance of gaze tracking in different user platforms are analyzed and issues with diversity in gaze accuracy metrics are discussed. The background and concept of a methodolog- ical framework for practical evaluation of eye gaze systems is presented in Section VI. We note here that the scope of this review excludes gaze tracking for clinical and neurolog- ical directions, retinal imaging and studies on children and patients.
II. EYE GAZE TRACKING FUNDAMENTALS A. TYPES OF EYE MOVEMENTS STUDIED Several types of eye movements are studied in eye gaze research and applications to collect information about user intent, cognitive processes, behavior and attention anal- ysis [28]–[31]. These are broadly classified as follows: 1. Fixations: These are phases when the eyes are stationary between movements and visual input occurs. Fixation related measurement variables include total fixation duration, mean fixation duration, fixation spatial density, number of areas fixated, fixation sequences and fixation rate. 2. Saccades: These are rapid and involuntary eye movements that occur between fixations. Measurable saccade related parameters include saccade number, amplitude and fixation-saccade ratio 3. Scanpath: This includes a series of short fixations and saccades alternating before the eyes reach a target location on the screen. Movement measures derived from scanpath include scanpath direction, duration, length and area covered 4. Gaze duration: It refers to the sum of all fixations made in an area of interest before the eyes leave that area and also the proportion of time spent in each area. 5. Pupil size and blink: Pupil size and blink rate are measures used to study cognitive workload. Table 1 presents the characteristics of different eye movements and their applications.
B. BASIC SETUP AND METHOD USED FOR EYE GAZE ESTIMATION Video based eye gaze tracking systems comprise fundamen- tally of one or more digital cameras, near infra-red (NIR) LEDs and a computer with screen displaying a user interface where the user gaze is tracked. A typical eye gaze tracking setup is shown in Fig. 1. The steps commonly involved in passive video based eye tracking include user calibra- tion, capturing video frames of the face and eye regions
16496 VOLUME 5, 2017
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
TABLE 1. Classification of eye movements.
FIGURE 1. Schematic diagram of a typical gaze tracking system. P is the pupil of the human eye ball and G is the glint location formed on the cornea, which are imaged on the camera plane. The figure also shows the error in gaze estimation as the deviation between actual & estimated gaze locations.
of user, eye detection and mapping with gaze coordinates on screen. The common methodology (called Pupil Center Corneal Reflection or PCCR method) involves using NIR LEDs to produce glints on the eye cornea surface and then capturing images/videos of the eye region [26], [32]. Gaze is estimated from the relative movement between the pupil center and glint positions. External NIR illumination with single/multiple LEDs (wavelengths typically in the range 850+/− 30 nm with some works such as [33] using 940 nm) is often used to achieve better contrast and avoid effects due to variations induced by natural light. Webcams are mostly used; those operate at 30/60 fps frame rate and have infrared (IR) transmission filters to block out the visible light. Different gaze tracking methods are discussed in detail in Section III.
The user-interface for gaze tracking can be active or pas- sive, single or multimodal [34]–[36]. In an active user
interface, the user’s gaze can be tracked to activate a func- tion and gaze information can be used as an input modality. A passive interface is a non-command interface where eye gaze data is collected to understand user interest or attention. Single modal gaze tracking interfaces use gaze as the only input variable whereas a multimodal interface combines gaze input along with mouse, keyboard, touch, or blink inputs for command.
C. CALIBRATION A generalized structure and model of a human eye is shown in Fig. 4a. The eye parameters typically required in gaze estimation are pupil center, center of curvature of cornea, the optical and the visual axes [32]. The posterior of the eyeball is called retina and the center of the retina with highest visual sensitivity is called the Fovea. The line joining the fovea with the center of corneal curvature is called the visual axis. Optical axis is the line passing through the pupil center and center of corneal curvature as shown in Fig 4a. The visual axis determines the direction of gaze and deviates from the optical axis. This offset is known as the kappa angle and measures around 5 degrees [37] but is dependent on each user. In gaze estimation, the pan and tilt components of the kappa angle are unique to each user and the visual axis cannot be estimated directly. The visual axis and the kappa angle therefore have to be obtained for each person though a process called calibration which has to be done at the start of an eye tracking procedure. Calibration is performed by showing the user a set of specific targets distributed over the front screen (as shown in Fig. 2) and the user is asked to gaze at them for a certain amount of time [38]. The tracker camera captures the various eye positions for each target point which are then mapped to the corresponding gaze coordinates and thus the tracker learns this mapping function. Calibration routines differ in the number and layout of target points, user fixation duration at each point and type of mapping algorithm used.
VOLUME 5, 2017 16497
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
FIGURE 2. Calibration screen with 3, 5, 9, 13 target points.
D. CORRESPONDENCE OF EYE GAZE WITH HEAD POSITIONS The gaze location of a user depends both on the gaze direction and also on the head orientation [39]. In methods which use PCCR techniques, if the user moves their head with respect to the tracker-camera axis while looking at the same point on the front screen, the glint vectors with respect to the pupil centers (for two different eye locations produced by head movement) will be different from each other. There- fore the estimated gaze locations will be inaccurate. Eqn 1 presents the relationship between user reference gaze directions (dkref), head pose direction dk and actual gaze direction (dkgaze) which is a result of both head and eye rotation, as shown in Fig. 3. The effect of head movement has to be compensated before applying the gaze mapping algorithm or a chin rest for fixing head pose has to be used.
dk −dkref = k ( dkgaze −dkref
) (1)
where k is a parameter related to head pan and tilt reported in [39] with values 0.5 and 0.4 respectively.
FIGURE 3. Relation between gaze direction & head pose.
E. ESTIMATION OF GAZE TRACKING ACCURACY In a typical eye gaze tracking operation, a user gazes at an interface on a computer screen which provides them with visual stimulus in the form of a set of targets or a scene. Gaze tracking accuracy is estimated as the average difference between the real stimuli positions and the measured gaze positions, which also provides an idea about the performance of the system.
In conventional literature gaze tracking accuracy measures are presented in different ways e.g. angular accuracy in degrees, distance accuracy in cm or distances in pixels. These accuracy estimate calculations are shown below. In practice, calculations are made separately for both eyes. For brevity, single calculations are presented and the same equation holds for both right and left eye. POG·Xleft, POG·Yleft, POG·Xright, POG ·Yright are the measured X,Y coordinates of the left and
right eye’s point of gaze (PoG). The mean gaze coordinates considering both eyes are POG · X and POG · Y . dist is the distance of the eye from the screen and mean_dist is the mean distance of eye from the tracker. The x/y pixels are the pixel shifts in x/y directions and offset is the distance between the tracker sensor and lower edge of display screen. Further details on these calculations can be found in [40].
Gaze point coordinates:
POG ·X = mean ( POG ·Xleft +POG ·Xright
2
) (2)
POG ·Y = mean ( POG ·Yleft +POG ·Yright
2
) (3)
Pixel accuracy (Pix_acc):
Pix_acc
= √ ( (target ·X−POG ·X)2 + (target ·Y−POG ·Y)
2 ) (4)
On Screen Distance (OSD):
OSD = pixelsize
×
√√√√((POG·X−xpixels 2
)2 +
( ypixels−POG·Y+
offset pixelsize
)2) (5)
Angular accuracy (Ang_acc):
Gaze angle (θ) = tan−1(OSD/dist) (6)
Ang_acc = (pixelsize∗Pix_acc∗ cos(mean(θ))2)
/mean_dist (7)
III. EYE GAZE ESTIMATION ALGORITHMS Eye gaze tracking algorithms comprise of corneal reflection based methods which use NIR illumination to estimate the gaze direction or the point of gaze using polyno- mial functions, or a geometrical model of the human eye. 2D regression, 3D model, and Cross ratio based methods fall into this category. Another class of methods uses visible light and content information (e.g. local features, shape, texture of eye regions) to estimate gaze direction, e.g. appearance and shape based methods. The five different gaze tracking methods have their own advantages and disadvantages which are briefly discussed at the end of this section and summary of some key works on the development of these algorithms are presented in Table 2.
A. 2D REGRESSION BASED METHODS In regression based methods, the vector between pupil center and corneal glint is mapped to corresponding gaze coordinates on the frontal screen using a polynomial trans- formation function. This mapping function can be stated as: f: (Xe, Ye) → (Xs, Ys) where Xe, Ye & Xs, Ys are equipment and screen coordinates respectively. The relation
16498 VOLUME 5, 2017
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
TABLE 2. Classification of gaze estimation algorithms.
can be presented as described in [41] and [42]:
Xs = a0 + n∑
p=1
∗
p∑ i=0
a(i,p)X p−i e Y
i e (8)
Ys = b0 + n∑
p=1
∗
p∑ i=0
b(i,p)X p−i e Y
i e (9)
Where n represents polynomial order, ai & bi are the coef- ficients. The polynomial is optimized through calibration in which a user is asked to gaze at certain fixed points on the frontal screen. The order and coefficients are then chosen to minimize mean squared difference (ε) between the esti- mated and actual screen coordinates (with known camera
VOLUME 5, 2017 16499
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
FIGURE 4. (a) Model of a human eye ball, eye parameters and setup elements used in 3D eye gaze tracking [32], [56]. The optical axis is shown as the line joining the center of curvature of the cornea with the pupil center. The visual axis passes through the fovea and the center of corneal curvature. Kappa angle is the angular deviation between the optical and visual axis. (b) An aspherical model of the cornea, as a surface of revolution about the optical axis of the eye [59].
coordinates of user gaze), which is stated as:
ε = (Xs −Ma) T (Xs −Ma)+ (Ys −Mb)
T (Ys −Mb)
(10)
Where a and b are the coefficient vectors and M is the transformation matrix given by:
aT = [a0a1 . . .am], b T = [b0b1 . . .bm] (11)
M =
1 Xe1 Ye1 · · · Xne1 · · · X
n−i e1 Y
i e1 · · · Y
n e1
1 Xe2 Ye2 · · · Xne2 · · · X n−i e2 Y
i e2 · · · Y
n e2
... ...
... · · · ... · · · · · · · · · · · ·
1 XeL YeL · · · XneL · · · X n−i eL Y
i eL · · · Y
n eL
(12)
Where M is the transformation matrix, m the number of coefficients and L the number of calibration points [42]. The coefficients can be obtained by inverting the matrix M as [43]:
A = M−1Xs, b = M −1Ys (13)
Cherif et al. [42] used a 5×5 point calibration rou- tine and 2 higher order polynomial transformations while Cerrolaza et al. [44] studied effects of head movement on system accuracy with a 4×4 and 8×8 grid. Blignaut [41] also compared several mapping functions and calibration config- urations with a 15×9 point grid and determined that number and arrangement of calibration targets and components of the mapping function play very important roles in determining overall accuracy of tracker. Robust and accurate gaze estima- tion under head movement was also achieved using neural networks by Zhu and Ji [45] and Jian-nan et al. [46]. Some other key works in this class of gaze estimation meth-
ods include Ma et al. [47] which introduces a 2D mapping algorithm that can handle unconstrained head movements and distorted corneal reflections due to various noise effects. It uses multiple geometrical transformation based mapping of CRs and demonstrates high reliability measures for differ- ent user distances and loss of CRs due to head/eye motion. A calibration free algorithm is detailed in Zhu et al. [48]
using SVMs which is robust to natural head movement and achieves high accuracy (1.5 degrees) in presence of head movement Cerrolaza et al. [44] provided an exhaustive and detailed review of mapping equations and their impact on gaze tracking system response. The paper reports 400000 cal- ibration functions, mapping orders and features to compare their impact on accuracy measures. Another important work is by Zhu and Yang [49] which presents a very high reso- lution gaze estimation method resistant to head pose without requiring geometrical models. In [50] an improved three layer artificial neural network is used to estimate the mapping function between gaze coordinates and the pupil-glint vector. This method is shown to achieve better accuracy than simple regression based methods. A new approach involving only 2 light sources instead of four is implemented in [51] where two IR LED induced glints are real and the other two are virtual ones computed mathematically. This is done to make the algorithm more suitable for consumer applications and simplify hardware. The PoG is then estimated using map- ping function to relate the four glint locations and screen through a calibration process. A modified PCCR method is implemented in [52] to have improved tracking accuracy and suitable for indoor and outdoor use. In this, adaptive exposure control is proposed since the IR LED brightness variations within a PCCR based setup affects pupil detection and gaze tracking to a large extent.
B. 3D MODEL BASED METHODS These methods use a geometrical model of the human eye to estimate the center of the cornea, optical and visual axes of the eye (Fig. 4a) and estimate the gaze coordinates as points of intersection where the visual axes meets the scene. 3D model based methods can be categorized on the basis of whether they use single or multiple cameras and type of user calibration required. 3D model based methods using single camera have been
reported by Meyer et al. [53], Guestrin and Eizenman [32] and Hennessey etal. [54]. Single camera systems have simple
16500 VOLUME 5, 2017
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
system geometry, no moving parts and fast re-acquisition capabilities. For 3D gaze estimation in Meyer et al. [53] a single camera and LED are used to achieve an accuracy of 0.5 degrees with user calibration. Guestrin presents a mathematical model to reconstruct the optical and visual axes of the user’s eyes from the centres of the pupil and glint in the captured video frames and configuration of a remote gaze tracking system. The model considers single and multiple cameras and light sources in estimation of the point of gaze. It then demonstrates the gaze tracking performance of a system implemented using two NIR light sources and one camera using the model .Their method achieves an accu- racy of around 0.9 degrees. The system proposed by Hen- nessey includes a single camera and multiple LEDs to achieve 3D gaze tracking with free head motion.
Multi camera methods achieve high accuracy and robust- ness against head movement but require elaborate system calibration procedures including calibration of cameras for 3D measurements, estimating positioning of LEDs and deter- mining the geometric properties of the monitors and their relation with the cameras. Some key works using two or more cameras include Lai et al. [55], Ohno and Mukawa [56], Beymer and Flickner [57] and Zhu and Ji [37]. Ohno describes 3D gaze tracking allowing free head motion using simple two point calibration and a two camera system com- prising of an eye positioning unit and a gaze detection unit. The eye positioning unit uses narrow field stereo cameras and controls direction of the gaze positioning unit to achieve head motion independent tracking.
A head pose free gaze tracking system is also implemented by Beymer with a wide angle stereo system for eye posi- tioning and narrow angle stereo system for gaze detection. Pan and tilt directions of the narrow angle camera are controlled using rotating mirrors with galvo-motors. Shih and Liu [58] used a simplified eye model by Le Grand with two cameras and two LEDs to estimate the optical axis of the eye through solving linear equations. It uses single point calibration. The method in Zhu uses a gaze mapping function along with a dynamic head compensation model to update the gaze mapping function whenever the head moves to achieve tracking under natural head movement. It uses a 2 camera sys- tem with one time user calibration. With respect to the model of human eye, [32] provided evidence that a fully spherical corneal model will result in no impact of head movements on gaze estimation. They assumed an ellipsoidal model of the cornea and reported that gaze estimation errors increase with corneal asphericity and this also results in sensitivity of gaze estimation to head movements. An aspherical model of the cornea [59] is shown in Fig 4b and it is a surface of revolution about the optical axis of the eye. Use of this model showed to result in better accuracy, especially near the display corners as compared to a traditional 3D model based method.
Calibration free gaze estimation techniques have been pro- posed by Nagamatsu et al. [60], Model and Eizenman [61] and Morimoto et al. [4]. In [60], a calibration-free method is proposed using two pairs of stereo cameras, light sources
and a spherical model of the cornea. One pair of cameras and two light sources are used for each eye to estimate the eye optical axis and the position of the center of the cornea. Optical axes of both eyes are measured using a binocular 3D eye model to estimate the point of gaze, achieving an accuracy of around 2.0◦. Model & Eizenman also proposed a multiple camera based system to capture stereo images of the eye with corneal reflections. From the stereo eye images, eye features, such as the center of the pupil and corneal reflections are used to estimate subject-specific eye parameters. These parameters are then used with the eye features to estimate PoG. Their method is calibration free, has a tracking range of 3 to 5 meters and accuracy of less than 2 degrees. Morimoto proposed a method using two light sources and one camera that doesn’t require user calibration for every session. It uses the Gullstrand model of the eye and ray tracing techniques to estimate the cornea and pupil centers. It achieves an accuracy of 2-4 degrees of visual angle dependent on the position of the light sources.
A new class of 3D gaze tracking has recently emerged with the usage of depth sensors in several works. These sensors comprise of an RGB camera and an infra-red depth camera. Typically resolution for the RGB camera is 640 × 480 pixels, with 45 degrees vertical and 58 degrees horizontal field of view. The depth camera resolution is about 1.5 mm at 50 cm. Gaze tracking using the consumer grade depth sensor (Kinect) is proposed in [62]. The method uses an eye model; 3D coordinates of eye features are obtained from Kinect and eye parameters like eyeball and pupil center are derived from a user calibration process. With this, 3D gaze coordinates are tracked in real time with a simple setup. Another work [63] reports the use of Kinect and a simple low cost setup for 3D model based gaze esti- mation allowing free head motion. It derives the 3D model parameters using convolution based means of gradients iris center localization method and uses a geometric constraints- based method to estimate the eyeball center. They assume that iris center points are distributed on a sphere originated from the eyeball center and the sizes of two eyeballs of a subject are identical. Kinect data is used to obtain 3D positions of person’s head pose, iris and eyeball centers. Reference [64] also uses a Kinect sensor and a model to estimate eyeball center by making users look at a target in 3D space. Kinect is used to build a head model to determine the eyeball center, detect the pupil center and determine 3D eye gaze coordinates in conjunction with the eye model.
C. CROSS-RATIO BASED METHODS These methods work by projecting a known rectangular pat- tern of NIR lights on the eye of the user and estimating the gaze position using invariant property of projective geometry. Four LEDs on four corners of a computer screen are used to produce glints on the surface of the cornea (Fig. 5). From the glint positions, the pupil and the size of the monitor screen, gaze location is estimated using two perspective projections. The first projection comprises of the virtual images of the
VOLUME 5, 2017 16501
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
FIGURE 5. Setup for implementing cross ratio based gaze tracking [67]. Four light sources are used at four corners of the monitor screen (only one is shown here. V1 and R1 are the virtual projection and corneal reflection of L1, Rc is the reflection and the virtual projection of the LED fixed at the camera’s optical axis. V1, R1 and Rc are projected to the image plane as Uv1, Ur1, and Urc.
corneal reflections of the LEDs (scene plane). The second projection is the camera projection, that is the images of the corneal reflections on the camera’s imaging plane. With these two projections a single projective transformation relating the scene and camera image plane is obtained. Then the projection of the PoG on the scene plane to the image of the pupil center on the camera plane can be estimated [65].
Key works on the development and experimental verifi- cation of the cross ratio based methods can be found in Yoo and Chung [66] and Hansen et al. [6]. Coutinho et al. presents a detailed analysis of methods and comparison of their accuracy in [67] and [68]. They also suggest improve- ments by including a fifth LED on the optical axis of the camera and using a calibration procedure to improve accu- racy [68]. Error compensation with polynomial-based regres- sion have been proposed by Cerrolaza et al. [44] or Gaussian process regression [6]. Error correction by homography map- ping that eliminates the need of the fifth light source has been proposed by Kang et al. [69].
D. APPEARANCE BASED METHODS In appearance based methods the information from the eye region is represented using a model trained with a set of features extracted from eye images. In Bacivarov et al. [70], a statistical model is used to represent shape and texture variations and trained using images of the eye region anno- tated with landmark points (Fig.6). The shape vector is the concatenated coordinates of all landmark points, stated as
s = (x1,x2, . . . ,xL,y1,y2, . . . ,yL) T (14)
where L is the number of landmark points. The shape model is obtained by applying Principal Component Analysis (PCA) on the set of aligned shapes (equations derived from [70]):
s = S̄+ϕsbs,
S̄ = 1/Ns Ns∑ i=1
si (15)
FIGURE 6. Image fitted with an Active Appearance model of the eye region [70].
where S̄ is the mean shape vector, and Ns is the number of shape observations; ϕ s is the matrix having the eigen- vectors as its columns; bs is the set of shape parameters. Similarly, the texture vector defined for each training image is: t=(t1,t2, . . . , tp)T (p : number of texture samples). The texture model is derived by means of PCA on the texture vectors as (Nt:: number of texture observations, T̄ : mean texture vector)
t = T̄+ϕtbt
T̄ = 1/Nt Nt∑ i=1
ti (16)
The sets of shape and texture parameters (bt) describe the appearance variability of the model:
c = ( Wsbs bt
) (17)
(Ws is the vector of weights). This is the statistical model that an Active Appearance Model (AAM) algorithm uses to best fit the model to a new eye image.
An active appearance based method for retrieving eye gaze from low resolution videos is presented in [71]. Global and local appearance models are trained and fitted for the whole face as well for capturing variance of the face and eye regions. For classifying the eye gaze into six directions, two different approaches are adopted. Gaussian Mixture Models (GMMs) are trained for large changes in gaze angles and for small gaze changes a Histograms of Oriented Gradients based method are tested. A method for 3D gaze tracking without use of active illumination is proposed in [72]. In this a synthetic iris appearance fitting method is introduced that computes the 3D gaze direction from iris shape. The method synthesizes a set of iris appearances and then fits the best solution to the captured eye image. This is claimed to remove unreliable iris contour detection problems arising in simple ellipse fitting and requirement of high resolution images by other methods. Once the iris contour is accurately estimated, a 3D eyeball model is used to estimate gaze from the captured eye image using the iris center/shape information.
Several appearance based methods report use of local fea- tures with Support Vector Machine (SVM)s for classification of gaze direction. These include Wu et al. [73] in which an AAM is used to locate the eye region using 36 feature points that represent the contour of eyes, iris size, iris loca- tion, and position of pupils. Gaze direction is estimated from 2D coordinates of feature points and is classified using an SVM. In Lu et al. [74] Local-Binary-Pattern (LBP) is used
16502 VOLUME 5, 2017
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
to calculate the texture features and a dual camera system is used to detect the space coordinates of the eyes. These two sets of information are fed into an SVM to classify the gaze direction under natural head movement. A novel method based on Local Binary Pattern Histogram (LBPH), is used in [75]. LBPH and PCA are used to extract eye appearance features and several classification methods based on SVM, neural networks and k-Nearest Neighbor (k-NN)s are tested for accuracy on a collected dataset for gaze estimation. The LBPH with SVM yields best accuracy. Chen and Liu [76] reports the use of a special kind of discriminatory Haar fea- tures and efficient SVMs (eSVMs) for implementing a com- putationally efficient gaze tracking method. Haar cascade is also used in [77] for real time gaze tracking. Rectangular features of the eye region are calculated to extract eye and pupil regions in an image which are mapped with the gaze coordinates on screen.
Neural network based approaches are used in [78] for head-pose tolerant gaze tracking. Training data comprised of cropped eye images of a user gazing at a given point on a computer screen and corresponding coordinates of that point. Neural networks are used along with a skin color model to detect the face and eye regions in [79]. An improved artificial neural network optimized using the Particle Swarm Optimization approach is used for fast, high accurate and robust gaze estimation with low resolution eye images in [80].
Some methods include use of 3D face models as in Lai et al. [81] in which head pose free gaze estimation is implemented using a such a face model with head and eye coordinate systems. Both eye appearance and head pose are considered as components of a high dimensional head pose and eye appearance (HPEA) space. A Neighbourhood Approximation Forests (NAF) approach is used to model the neighbour structure of the HPEA space followed by Adaptive Linear Regression to estimate gaze direction. Other approaches report use of a deformable model [76] and Genetic algorithms [83].
In recent times, deep learning (DL) and convolutional neu- ral network (CNN) based methods have been proposed for gaze estimation. In [84], a three stage CNN model is used to classify seven gaze directions from images taken with low cost webcams without need for calibration. Gaze tracking for a near eye display robust to illumination, skin and eye color variations and occlusion is implemented using CNNs in [85]. The CNN is used to learn the mapping from eye images to gaze position and comprises of two convolutional layers and two pooling with a fully connected layer at the end.
In [86] a smartphone based app is created to collect eye images from 1450 participants which is used to train a CNN based gaze tracker that can run in real time and without calibration. The dataset comprises of images with widely varying backgrounds, lighting and head motion and the network is trained with crops of both eyes and the face region.
In [87] an extensive eye gaze dataset is built and a multi- modal CNN based method is tested. The dataset contains
more than 200,000 images with variable illumination levels and eye appearances. The CNN uses a convolutional layer followed by a max-pooling layer and second convolution layer followed by a max-pooling layer, finally with a fully connected layer. The CNN learns the mapping between input parameters, i.e., 2D head angle, eye image and gaze angle (output).
E. SHAPE BASED METHODS These methods employ deformable templates of the eye region, using two parabolas for eye contours (Fig.7) and circle for the iris, and fitting them to an eye image [88]–[91]. The procedure is to find the similarity between the template of a chosen region with images of that region.
FIGURE 7. Template of an eye -region [88]. Xc,, Yc & Xe,, Ye represent the center of the pupil and the eye respectively. P1 and P2 are foci of two parabolic sections and a, b, c and θ their parameters, r is the radius of the pupil.
This can be done by normalized cross-correlation, mod- ified cross-correlation or by mean square error calcula- tion [89]. If the template of pixel intensities of a region is represented by T(u,v) and I(i,j) represents that of the cap- tured image then S(i,j) is the similarity measure between the template and image. If cross correlation is used as similarity measure, then S is given by:
S(i, j) = < T×IT > − < T >< IT >
σ (T)σ(IT) (18) (18)
where <> is the average operator and < x > is the pixel-by -pixel product given by:
< T > = 1 n
∑ u,v
T(u,v)
< T× IT > = 1 n
∑ u,v
T(u,v)I(i+u, j+v) (19)
σ is the standard deviation of the area being matched.
σ 2 (T) =
1 n−1
∑ u,v
(T(u,v))2 −< T >2 (20)
The mean squared error similarity measure is given by:
S(i, j) = 1 n
∑ u,v
(T(u,v)−I(i+u, j+v))2 (21)
VOLUME 5, 2017 16503
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
F. SUMMARY AND DISCUSSIONS The different gaze estimation algorithms presented above have distinct characteristics, advantages and disadvantages.
The 2D regression based methods utilize the features of the human eye, like eye geometry, pupil contours and corneal reflections and can be implemented using a single camera and a few NIR LEDs. However, these techniques are very vulnerable to head movements and require users to hold their head very still using a head rest, chin rest or bite bar.
3D model based methods have tolerance towards user head movement and most of them allow free head motion. How- ever the hardware requirements for implementing 3D and stereo gaze tracking methods are high as they need several light sources or multiple cameras.
Cross ratio based methods have advantages, e.g. they do not need an eye model or hardware calibration and allow free head motion. But they are affected by problems such as increased error with distance of user and user dependent factors.
Appearance model-based algorithms are non-PCCR meth- ods that use the shape and texture properties of the eyes and position of the pupils relative to the eye corners to estimate gaze. These methods have low hardware requirements which make them suitable for implementation on platforms without a high-resolution camera or additional light sources. The disadvantage is that their accuracy is mostly lower than PCCR based methods that degrade with head movements, variation in illumination levels and for robust performance they need large training image databases.
Shape based methods have been implemented for 2D gaze estimation with low-resolution webcam images achiev- ing accuracy around 2◦. However, downside of these methods include problems due to head pose variations and eye occlusions, adapting to largely variable eye shapes, computational complexity and issues with model initialization.
In Table 2, some key research works on the above gaze tracking algorithms are presented. Column II presents the reference to the individual papers; Columns III to V presents various characteristics of the methods reported in them. The table highlights basic features and differences among different gaze tracking algorithms. 2D regression and appearance based methods have simple setups but typically offer accuracy values around 2 or 3 degrees. The accuracy of 2D regression based methods can be improved by increas- ing number of calibration points and using a chin rest to obtain fixed head position. On the other hand, 3D and cross ratio methods require more elaborate setups but offer much better accuracy (around 0.5 degrees) and most of them allow head movement.
IV. USER PLATFORMS IMPLEMENTING GAZE TRACKING In this section, the user platforms where eye gaze tracking has been implemented are described and classified.
A. DESKTOP BASED SYSTEMS Applications of eye gaze on desktop systems fall into several categories, such as computer communication, password entry and psychoanalysis. Sibert and Jacob [35] developed gaze based fast object selection as a substitute for mouse. A gaze based application called MAGIC (Manual And Gaze Input Cascaded) is presented by Zhai etal. [92] in which gaze based pointing is reported to have higher speed and accuracy than manual pointing. In Ghani et al. [93], a Hough transform based pupil detection for gaze based control of a mouse pointer is proposed. In Agustin et al. [94] evaluation studies on the use of eye gaze in video gaming control for target acquisition and tracking are made. Gaze input had a similar performance to the mouse and joysticks for big targets.
Kasprowski and K. Harężlak [95] and Kumar et al. [96] reported the use of gaze to enter a password using gaze tracking. A series of user fixations on specified digits formed the password sequence. Methods robust to shoulder surfing problem were reported by Bulling et al. [97] where a compu- tational model of visual attention is used to increase security. Applications of using eye gaze patterns for identifying user tasks, mental workload and inferring context of events and user distraction were reported by Iqbal and Bailey [98] and Doshi and Trivedi [99].
B. TV AND LARGE DISPLAY PANELS There are recent applications of long range gaze estima- tion that use corneal reflection (CR) techniques for tracking gaze on large displays and smart TVs. Gaze movements can be used to select and navigate menus, modify display properties, switch channels and understand user interests. In Gwon et al. [7] a robust pupil detection method for gaze tracking on large display is presented using a wide and a nar- row view camera with Adaboost and CAMShift algorithms. In another work, Lee et al. [8] reports a system for gaze tracking on a large-screen 60 inch TV based on a 2D method with geometric transform, using pupil center and four corneal specular reflections.
C. HEAD-MOUNTED SETUPS Head mounted gaze trackers are portable platforms with applications ranging from computer input, interactions in virtual environments, gaming controls, augmented reality and neuro/psychological research. The general setup includes two cameras; one (eye camera) pointed at the wearer’s eye, to detect the pupil; and the other (scene camera) capturing the wearer’s point of view, with sometimes additional compo- nents like NIR light sources and hot mirrors. Head-mounted gaze trackers have been implemented as attachment-free, mobile, low-cost, lightweight devices with simple hardware and software. Also they are known to provide high accuracy gaze information in unconstrained settings.
3D gaze estimation with head-mounted trackers have been reported in several papers including [9], [100]–[104]. Algorithms for high accuracy and 3D eye tracking proposed by Lee et al. [100] are based on 3D human eye model
16504 VOLUME 5, 2017
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
and Purkinje images. Pupil size and Purkinje images are fed as inputs to a multi-layer perceptron to estimate the depth location of gaze followed by 2D gaze coordinates. Takamatsu et al. [104] describes a Visual SLAM technique to estimate the user head pose and determine 3D point of regard of the user. It also includes a 3D environment to detect objects of focus and visualization of a 3D attention map in a real environment. Lanata et al. [9] presents a stereo-vision based method that implements a novel binocular system comprising of two mapping functions: linear and quadratic for depth estimation of gaze locations in 3D space.
The performance compared to other 3D methods show significant improvement in accuracy. In [103], 3D gaze track- ing with multiple calibration planes is implemented with a single eye view camera and four IR markers. The system operates in monocular and binocular mode independently. A low-cost 3D eye tracking solution is provided by [101] in which images from two eye cameras are used with intensity thresholding and blob-contour detection to determine pupil center coordinates. The 2D gaze location is then obtained by polynomial mapping and 3D gaze is obtained from the meeting point of the gaze vectors of both eyes.
2D methods and applications with head-mounted eye tracking have been proposed in a several works. Reference [10] reports an ultra-low cost 2 camera based gaze tracking system that is easy to assemble and aligns scene and eye videos by using synchronized flashing lights. A novel smooth-pursuit-based calibration methodology is proposed that works with pupil detection to track gaze. A new high speed binocular eye localization scheme using simple com- ponents was proposed by Long et al. [105] using two cam- eras and two hot mirrors. It shows speed advancement over contemporary slow systems that do offline processing. Other advanced methods are proposed in Schneider et al. [106] where a head mounted camera system tracks and follows the user’s gaze direction for natural exploration of a visual scene by capturing the perspective of the mobile user. Applications of this system include documentation of med- ical processes, sports etc. In Virtual and Augmented real- ity (VR and AR respectively) research, gaze tracking and gaze based functions are used to make the user experiences more immersive, natural and user interactions fast and effi- cient in a VR/AR environment. Applications for Augmented Reality are described in Lee et al. [107] where the device uses a scene camera and eye tracker to estimate user gaze and blink-state for interacting with the AR environment. A head mounted system that employs gaze tracking for immersive and realistic gaming experience is proposed by Lee et al. [108] where cursor aiming in the gaming environ- ment is controlled by gaze tracking. A wearable gaze track- ing device for determining effectiveness of text layout and line spacing for analyzing reading behavior is described in Toyama et al. [109]. Li et al. [11] presents the development of a head mounted system called OpenEyes with open hardware and software tools for gaze tracking. Several types of eye and head gestures, such as saccades, smooth pursuit and nod-roll
are studied in [110] as interaction methods in a head mounted VR device. A head mounted display coupled with an eye tracker is used to study which of these gestures result in better user experience in a VR environment.
A facial re-enactment method for gaze aware VR system is presented in [111]. It does real time facial motion capture of a user wearing an HMD along with monocular eye tracking. The purpose is to achieve real time photo-realistic rendering of face and eye appearances as users change facial expres- sions and gaze directions. Gaze is estimated from images taken using monocular camera and IR LEDs using a hierar- chical classification method. In [112] eye tracking is used to develop an immersive 3D user interface for VR and imple- ment several multimedia applications along with usability testing. Eye gaze pattern is extracted along with fixations to implement 3D virtual menu selection and eye based typing for mail composition using a virtual keyboard. A saliency based gaze localization approach using deep learning has been proposed in [113]. In this, image features and head movements are inputs to a convolutional neural network to estimate gaze coordinates in a VR system.
To make eye tracking in HMD systems flexible with respect to various users, HMD drifts and tracker camera adjustments, [114] presents a method for automatic calibra- tion of the eye camera in the HMD. With this, the users need not maintain a fixed head pose and relative movements of HMD and eye camera can be tolerated for reliable eye tracking. Reference [115] makes use of gaze, dwell time and half blink information for the purpose of hands-free object selection with an optical see through head mounted AR device. The multiple input parameters are used to avoid accidental/unintentional selection of items. Pupil motion is tracked to estimate a user’s viewing point, with dwell time and half blink detection successively used for the object selection process. Kalman filtering is used minimize pupil jitter and drifts. The gaze tracking accuracy of the system is reported to improve with more calibration points with the best being 0.39 degrees.
An innovative AR application combining eye tracking with a smartwatch for implementing a wearable context aware messaging service is done in [116]. A head-mounted eye tracker estimates the gaze and together with data from a scene camera it is used to track where the user is looking in the real world. The smartwatch works as a message input and output device in conjunction with the tracking device to embed/display messages with the ‘‘augmented objects. An AR application for reading and document retrieval using a head-mounted eye tracker along with a see through dis- play is shown in [109]. Eye tracking helps to estimate the section of the document the reader is focused into, in real time. Associated information on the specific part of the doc- ument is retrieved and displayed on the HMD. A calibration procedure for the HMD is also mentioned. In this, a user is presented with several dots in the HMD and he/she has to click the position of each dot in a calibration window. The homography between the scene image and the HMD is
VOLUME 5, 2017 16505
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
then estimated to map the gaze position on the scene to that on the HMD. The system achieves accuracy sufficient to distinguish between alternate lines in a document. A hard- ware and software framework for AR displays implement- ing eye tracking and gaze based interactions can be found in [117]. In this, a see through HMD is integrated with stereo eye tracking and configurable optics for various see through configurations. The system is designed to have on demand zoom and FOV expansion in a see through AR system. Eye related parameters like gaze, squint and blinks are tracked for activating different functions such as binary, sub-regional and gradual zoom and capturing snapshots of the AR view.
D. AUTOMOTIVE Visual features of the face and eye regions of an automobile driver provide cues about their degree of alertness, perception and vehicle control. Knowledge about driver cognitive state helps to predict if the driver intends to change lanes or is aware about obstacles and thereby avoid fatal accidents. Gaze related cues that indicate driver attentiveness include: blink rate, temporal gaze variation, speed of eyelid movements and degree of eye openness. Several works for driver assistance systems have been reported that are based on video based gaze tracking using a variety of classifiers.
SVM based gaze classification is common on automo- tive platforms and has been reported in Liu et al. [118], Lee et al. [119] and Chuang et al. [120]. A real time gaze tracking method robust to variable illumination levels and driver wearing glasses in an automotive environment is reported in [118]. It uses Kalman filtering and mean-shift method to track driver eyes based on their locations in a previous frame. SVM with linear, polynomial and Gaussian kernels are used for eye verification. Lee etal. [119] proposed a robust SVM based driver gaze zone estimation system that works during day and night conditions and is robust to driver wearing eyeglasses. A multiclass SVM constructed from multiple two class (binary) SVMs was constructed to classify 18 driver gaze zones using a large database of 18000 gaze feature vectors. Another SVM based gaze zone classifier that takes face parts, i.e., mouth, eyes, and nose locations is reported in Chuang et al. [120]. A multiclass linear SVM is used which inputs the feature descriptor to output 8 gaze directions. Use of other type of classifiers is reported in Tawari and Trivedi [121] and Oh and Kwak [122]. In [121] a distributed camera setup is used for estimating gaze zones combined with head pose dynamics. A random forest clas- sifier is used with static and dynamic features (head pose angles and time series statistics) to classify 8 gaze zones for the driver. In [122] Viola Jones method is used to detect the face followed by using linear discriminant analysis (LDA) to extract features from the eye region for classification. k-nearest neighbour method with Euclidean norm is used to classify the obtained features into seven gaze directions.
Gaze estimation in automotive using Purkinje images and PCCR methods are reported in Ji and Yang [17], Salvucci and Liu [123], Batista [14], Choi and Kim [124] and
Ji et al. [125]. Several papers report the application of gaze information along with other facial and visual parameters for derivation of driver psychological/cognitive state. Key works include Bergasa and Nuevo [126] which reports a method for driver vigilance estimation using fuzzy classifier taking six parameters: Percent eye closure (PERCLOS), eye closure duration, blink frequency, nodding frequency, face position, and fixed gaze. A dynamic Bayesian Neural net- work approach is used in Sung et al. [127] for driver fatigue detection using gaze location along with face detection, eye positioning and iris tracking. Ji et al. [125] combines features such as eyelid and head movement, gaze, and facial expres- sions with a Bayesian network to determine driver fatigue. Another important work that considers eye movements for deriving driver cognitive state is [16]. It takes into consid- eration eye movement features like fixations, saccades, and smooth pursuits and calculates 16 different eye movement features. These are then fed to an SVM, static and dynamic Bayesian networks to do their performance comparison for prediction of driver distraction state. Reference [128] presents a low cost system to detect eyes-off-the road condition of a driver using facial feature tracking, 3D head pose and gaze estimation. A monocular camera is installed close to the steering wheel for tracking a driver’s facial landmarks and accurate estimation of driver pose, location and gaze direction. Then, with 3D analysis of car/driver geometry the driver’s eyes off the road condition is predicted in real time. In [129], driver gaze behaviour is studied to evaluate driving performance of a user when they had to interact with a portable navigation system while driving. Glance frequency and glance time were estimated to study impact of varying display sizes and positions of the navigation device while in use during driving. A portable driver assistance system involving driver drowsiness detection and eye gaze tracking is implemented using a Raspberry Pi and machine vision algorithms in [130]. A new multi-depth calibration approach is presented in [131] for obtaining 3D user PoG with stereo face cameras and monocular scene cameras for performing driver intent and actions prediction.
Several works study the dynamics between head pose and gaze behavior of drivers. In [132] the relationship between gaze and head pose are studied and regression models are developed to predict gaze location from the position and orientation of a driver’s head. In [133] head and eye poses of 40 driving participants are estimated from monocular video and relative significance of head vs eye movements in gaze classification is studied. It is observed that driver behavior can be grouped into two cases, i.e., when the head moves a lot and gaze classification is mostly affected by head pose. The other case is when the head stays still and only the eyes move and thus accurate classification requires study of eye pose.
E. HAND-HELD DEVICES Smartphones and tablets provide a unique paradigm for gaze tracking applications. Gaze tracking on handheld
16506 VOLUME 5, 2017
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
FIGURE 8. Driver monitoring cameras mounted on car dashboard [124].
devices is done using the device front camera, one or more IR light sources and various computer vision algorithms. Key approaches for handheld-eye tracking are reported by Vaitukaitis and Bulling [134] for tracking 3D gaze location from a video using a face and eye detector followed by edge detection and ellipse fitting to determine the eye limbus boundaries. In Lukander [135] a commercial eye tracker is integrated with a magnetic position tracking device to track positions of head and eyes. Holland et al. [136] uses a single perceptron neural network that maps the features of the eye region to a position on the screen. In Imabuchi et al. [137] eye tracking on a tablet is realized with blob and contour detection for deriving iris contours. Then center and planar homography transform is used to convert the coordinate of iris center to display coordinates.
Several works use the Haar classifiers/Viola-Jones tech- nique [138] such as Elleuch et al. [139] which describes a system for android tablets to track head motions and eye gaze gestures from video captured using device front camera. The system is claimed to be robust to user movement and lighting conditions. In Pino and Kavasidis [140] eye track- ing for tablets is implemented using a Haar classifier based eye detection module in conjunction with the CAMSHIFT algorithm.
Potential applications of eye gaze in handheld devices are presented in Nagamatsu et al. [19] describing a user-interface called Mobigaze that uses gaze to operate a handheld device. The system combines gaze and touch for operating the device interface thereby eliminating the Midas touch problem [31]. The Midas touch effect [31] refers to the phenomenon in which an eye-gaze-sensitive user interface cannot distinguish between user glances for collecting visual information and those for command input. Thus every user fixation may lead to activation even without the actual user intention. Another multimodal gaze based interaction method is presented by Drewes and Schmidt [141] for monitoring applications using dwell-time and gaze gesture. The EyePhone introduced by Miluzzo et al. [142] controls the phone functions with only gaze and blinks and is free of user touch. It describes an application (Eyemenu) in which gaze direction is used to selectively access and highlight menu buttons. A gaze based user authentication scheme for smartphones is implemented
FIGURE 9. Eye tracking implemented in near real-time on a tablet. The tracking algorithm uses cascade classifiers and shape-based approaches to determine the eye region and centers. Elliptical model-fitting and 3D back-projections are then used to determine the eye optical axes and point-of-gaze [18].
in Liu et al. [143] by making the user eyes to follow a moving target on the device screen. In Sun et al. [144], a 3D image display is combined with gaze estimation to achieve adaptive 3D display on a mobile phone that shows different views to corresponding viewing angles. In Imabuchi et al. [137] a gaze based communication interface is implemented for gaze based keyboard input and gaming.
Table 3 presents key information obtained from selected papers based on five gaze user platforms. It can be seen that different research works present their results in widely different formats and describe performance of their system under varying operating conditions.
F. USER PLATFORMS FOR GAZE TRACKING: SUMMARY OF USER CONFIGURATIONS The user conditions while they are using gaze tracking in the above five platforms are completely different. Similarly, the operating environment and tracking setup are also unique to each platform. Table 4 presents typical system and user configurations for the five gaze tracking platforms including users’ postures and viewing angles, screen sizes, typical dis- tance between the user and the screen-camera setup. It can be seen that gaze tracking on different user platforms encounter wide range of operating conditions and therefore have diverse performance measures.
Apart from the differences seen in Table 4 above, Table 3 also shows the inconsistency in performance report- ing formats among the different user platforms. Accuracies are reported in varied units –such as degrees, pixels, percent- age of correct detection etc. Another feature as seen from Column 5 of this table is that only a few papers study the impact of operating conditions on the system performance. Further discussions about these aspects of gaze research are made in Section V.
Amongst the papers reporting in degree measures, a trend is seen in the accuracies for different eye gaze user platforms. Typically, it is seen that head mounted systems report better tracking accuracies of less than one degree amongst other platforms. For desktop systems it varies from 0.5 to 2 degrees of angular resolution and above 2 degrees for more dynamic platforms like automotive and handheld devices. Further, all platforms have unique setup requirements and users may
VOLUME 5, 2017 16507
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
TABLE 3. Summary of research on eye gaze use cases in various platforms.
16508 VOLUME 5, 2017
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
TABLE 3. Continued. Summary of research on eye gaze use cases in various platforms.
VOLUME 5, 2017 16509
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
assume various physical poses as seen in Table 4. Therefore, a general eye tracker may produce significantly different results depending on platforms.
TABLE 4. Features of gaze estimation systems in different user platforms.
V. PERFORMANCE METRICS AND PLATFORM SPECIFIC ERROR SOURCES IN EYE GAZE RESEARCH A. DIVERSITY OF GAZE ESTIMATION PERFORMANCE METRICS IN DIFFERENT USER PLATFORMS In contemporary literature, research works on gaze tracking present the results of their algorithm or application in several ways. The common measures used are angular resolution (in degrees), gaze recognition rates (in percentage) and minimum pixel shifts/distance between gaze and target locations. These metrics are no way correlated to each other and each research work defines some metrics, for example percent recognition rates or error rates in their own way. The consequence is that most gaze estimation methods cannot be inter-compared.
To better understand the diversity of performance measures used in gaze research, an information statistics is collected from our reviewed papers and presented in Table 5. Firstly, the papers are grouped according to platforms and then the performance measures reported in each paper are classified according to the categories: degrees, percentage and ‘‘others’’ (pixel shifts/distance in mm). In Table 5, Column 2 has the total number of research papers that was considered in the survey for each platform. Columns 3, 4 & 5 provide reference to papers that reported their performance measures in either ‘‘degrees’’, ‘‘percentage’’, or ‘‘other’’ formats respectively. Fig. 10 presents this statistics i.e. number of papers that reported accuracies in various formats for four different user platforms.
What can be observed from Table 5 is that there is no standard convention that is used to represent performance scores of eye gaze estimation algorithms in contemporary eye gaze research. For example, as seen in the table, out of total 69 papers from desktop/TV platform, 44 papers report tracking accuracy in degrees while 16 papers report them as gaze recognition rates and 9 papers in mixed units. Also this discrepancy is found in literature for all user platforms
FIGURE 10. Diversity in metrics used for representing performance in different platforms (Head-M is head-mounted, Auto is for automotive systems and Hand-H represents handheld devices like smartphones and tablets that implement gaze tracking). The figure shows that although degree measures are most common format of measurement, results are reported in several other formats which cannot be linked to each other. The inhomogeneity is majorly observed in automotive and handheld platforms.
as seen from the rows 3 to 5 of the table and Fig. 10. From Fig. 10, we observe that degree measures are common units used for desktop and head-mounted platforms, whereas the representation is completely heterogeneous for automo- tive or handheld devices. The result of this inhomogeneity is that, the stated performance from a large volume of research in this field can neither be compared nor interpreted quanti- tatively.
B. PLATFORM SPECIFIC FACTORS AFFECTING USABILITY OF GAZE TRACKING SYSTEMS Eye gaze estimation systems on various user platforms and applications face a broad range of operating conditions that are rarely taken into consideration or characterized during their development. The practical performance of a gaze track- ing system in reality may be affected by several factors (or we may call them: sources of errors) that are common or unique to each platform. Some of these factors and their effects have been discussed in our earlier work [24].
For example, the major sources of errors in desktop are head movements. In head-mounted trackers, errors may arise due to ‘‘Midas-touch’’ effect [31], miss-calibration or tracker latency. Errors in automotive systems may arise from plat- form and user head movements, variable illumination, occlu- sion due to shadows or user wearing glasses. In handheld devices, eye tracking gets highly challenged by changing positions of the user with respect to the device, head pose, hand jitter, variable illumination, and Midas-touch.
In gaze research, effects of some of these conditions such as head-pose changes, user distance and viewing angle are
16510 VOLUME 5, 2017
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
TABLE 5. Diversity in performance metrics in gaze estimation systems.
studied but other factors such as display properties i.e. size and pixel resolution of the screen where gaze is tracked, platform movement and jitter, illumination changes, camera quality and human eye limitations are very sparsely reported. To study this scenario, Table 6 presents an information statis- tics on practical conditions that may affect gaze tracking and the extent to which they have been studied. Firstly the different error sources are listed in Column I. Then the papers are categorized according to platforms to identify the research works where these error sources have been reported. This information is presented in the different columns of Table 6.
It is seen from the table that each user platform encounters at least 5-6 different conditions which can affect their perfor- mance but their impacts are very sparsely studied for all user platforms. Head-pose is the most studied factor out of all but very few papers analyze the impact of others error sources.
Further in presence of these factors, the real performance of an eye tracker can deviate significantly and unpredictably from the scores reported under ideal conditions. Hence, unless these error sources are adequately evaluated, the accu- racy achieved by a gaze tracking system cannot be reliably specified.
VI. A PERFORMANCE EVALUATION FRAMEWORK FOR EYE GAZE ESTIMATION SYSTEMS A. NEED AND RATIONALE FOR DEVELOPING COMPREHENSIVE PERFORMANCE EVALUATION STRATEGIES FOR GAZE ESTIMATION SYSTEMS From the sections and tables above it is seen that currently there exists a large diversity among gaze tracking methods, setup, implementation platforms and accuracy metrics.
An important issue that becomes apparent is that there is currently no comprehensive practice for realistic performance evaluation of gaze tracking systems. Most research works
do not assess their systems under the impact of various error sources or provide adequate details about their sys- tem configurations. For example, Table 6 shows that only 35 out of 69 papers on desktop based systems and 16 out of 57 papers on head-mounted systems report effect of head pose variations. Only 2 works in each of desktop and head- mounted platforms report effect of display properties of the gaze tracking setup. Effect of illumination changes is reported in 4 papers in desktop and one in head-mounted systems. In dynamic platforms like automotive and hand-held devices, where external conditions are more variable, the evaluation statistics is even poorer, as seen from Table 6. Another issue observed is that the accuracy scores for gaze based sys- tems are presented using varied formats and metrics, such as angular resolution, correct detection rate, pixel and physical distances etc. which makes them difficult to interpret and inter-compare. For example: Table 5 and Fig. 10 show that out of 182 total research works across all platforms, 95 papers report gaze accuracy scores in degrees whereas 41 papers report in percentages (correct detection rates) and 46 papers use other heterogeneous units.
There is also considerable ambiguity in terminologies used in eye gaze research. For example, mentioning ‘‘slight head movement’’ (as in [43] [78], and [145]) or ‘‘large head move- ment’’ (in [45], [47], and [66]) and free head movement (in [53], [56], and [67]) gives no quantitative idea about the real extent of head pose variations tolerated by these gaze tracking systems.
Considering these factors, the development of comprehen- sive evaluation strategies for gaze tracking systems seems necessary for several reasons: a. to study impact of various error sources on system performance b. to report system performance quantitatively in uniform formats c. compare results from different eye tracking systems under different operating conditions d. to identify the main bottlenecks for
VOLUME 5, 2017 16511
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
TABLE 6. Sources of errors in gaze estimation systems and literature where they are reported.
each platform. We present here the concept of such an evalu- ation framework in the sections below.
B. CONCEPT OF A PERFORMANCE EVALUATION FRAMEWORK FOR GAZE ESTIMATION SYSTEMS The framework is to be built around a set of standardized experiments for evaluating various gaze tracking systems. Through the experiments, practical performance limits of a given gaze tracking algorithm or system can be tested under the influence of various parameters: such as variations of head pose, viewing angle, screen size and resolution, eye- occlusion, platform movement and illumination changes. The structure of the framework is outlined in Fig.11. The advan- tage of having such a framework is that it can answer sev- eral critical queries related to eye gaze system design and performance.
For example, which system parameters affect the perfor- mance in a particular use-case? How does a particular system perform when compared with similar systems under certain operating conditions or in a particular use-case? Can an algo- rithm designed for one platform be ported and implemented effectively in another platform? What are the performance bottlenecks of individual algorithms? At present it is chal- lenging to answer such questions as there are no resources that allow us to do practical comparative testing of gaze estimation systems.
C. METHODOLOGY A typical eye tracking setup comprises of the user, gaze tracker and the tracking environment and each of these com- ponents may influence the overall eye tracking performance. A schematic diagram of such a setup and these factors are listed in Fig.11. The proposed experimental framework aims to test impacts of these factors on a tracker’s accuracy. A typ- ical ‘‘experiment’’ consists of the following steps: a user is asked to sit in front of the eye tracker and their eyes are calibrated for a session. The user is presented with a graphical user interface when the tracker records their gaze coordinates as the user gazes at several points on the screen. The gaze error in degrees is calculated from the shift between ground truth and tracked gaze locations. Some evaluation experiments done and planned with commercial eye trackers are presented here. Preliminary results can be found in [24].
1) ESTIMATING IMPACT OF HEAD POSE A user is positioned in front of the eye tracker while a video camera captures the position of the user’s head simultane- ously. The user’s head pose in roll, pitch and yaw angles are obtained from the video using an appearance model as shown in Fig.12. Then the user is asked to turn their head to specific fixed positions (in roll pitch yaw angles) and their gaze is tracked on the same interface again. The gaze accu- racy scores corresponding to various head pose angles are presented in Fig 13. For a particular tracker it was observed
16512 VOLUME 5, 2017
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
FIGURE 11. Outline concept of a methodological framework for performance evaluation of eye gaze systems.
FIGURE 12. Steps of quantifying head pose tolerance limits of a given eye tracking system.
that for reliable gaze tracking, head pose must be restricted within 20 degrees of movement in 3 directions and in this way, the practical head-pose tolerance limits of the tracker could be estimated.
2) USER DISTANCE AND VIEWING ANGLE For this experiment, the users are positioned at successively increasing distances from the tracker - computer screen setup (40 cm to 100 cm in 15 cm interval) and the gaze tracking accuracy data are recorded for each user position for fixed frontal head poses. The user viewing angle decreases as the user moves away from the screen and the gaze tracking
errors are seen to increase with increasing viewing angles. Tracking stops when users are closer than 40 cm.
3) ILLUMINATION LEVELS Several different illumination levels can be introduced dur- ing the experiments. There are cool temperature fluo- rescents (color temperature ∼6400K), warm incandescent lamps (color temperature∼2500 K) and mixed lamps (color temperature∼5500K) that are available at various intensities (100-3500 lux). Eye tracking experiments will be repeated under these illumination levels to study impact on tracking accuracy.
VOLUME 5, 2017 16513
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
FIGURE 13. The plot shows the gaze tracking errors arising due to user head motion while using a commercial eye tracker during our experiments. The human head has a degree of flexibility in the order of ∼40 degrees of angular movement in each of the roll, pitch and yaw directions. However the plot implies that for reliable gaze tracking with the given tracker, a user needs to keep their head position limited within a fixed angular range in each of these three directions. Through our evaluation process, we estimated this head movement limitation to be equivalent to ∼20 degrees about the central position. If the user’s head moves beyond these angles, gaze tracking errors of the tracker rises above acceptable levels. However, another feature observed from this plot is that the impact of head roll and pitch on tracking error are relatively more pronounced than yaw variations.
4) DISPLAY SIZE AND RESOLUTION The eye tracking experiment is run on displays of various sizes (9, 11, 13.5, 15 inches) and resolutions (800×600, 1024×768, 1280×768, 1366×768) and the corresponding gaze tracking errors are estimated. We have plans for includ- ing a 44 inch TV with a specialized tracker intended for gaze tracking for large screens in our testing framework and study impacts of user viewing angles and display sizes.
5) OCCLUSION We have already observed certain eye trackers having prob- lems with data collection from a user wearing glasses while another tracker has no such issues. Therefore, the level of tolerance of a tracker to user wearing glasses have to be evaluated by operating both trackers simultaneously while doing the gaze experiment for users with and without glasses.
6) PLATFORM MOVEMENTS The operating conditions faced by a gaze tracker on a static platform like desktop is largely different from that of a dynamic platform like a smartphone or head-mounted setup. We therefore plan to evaluate performance of eye trackers running on such dynamic platforms to realistically observe the difference in performance, study impact of platform movements and other influencing factors.
D. STUDYING DYNAMIC EYE MOVEMENT CHARACTERISTICS In reality, during a given task the human eyes are constantly moving and therefore sequential eye movements can be stud- ied as a statistical process. Through our framework, we aim
to do dynamic eye measurements in a video for studying smooth pursuits besides regular fixations. This is a planned inclusion in which the user will be presented with a ’moving target’ for the eyes to follow while capturing video of the eye- movements.
This framework is under development at the moment and preliminary results from some of the experiments have been published by us in [24]. The focus of this paper is mainly to highlight the issue of realistic performance evaluation of eye gaze systems through the literature review. Therefore only limited details about the methodology and implementation of the framework are presented here. A more comprehensive paper on the technical details of the framework is under preparation and will have more information.
VII. CONCLUSION AND FUTURE WORK Eye gaze estimation is an interdisciplinary area of research and development which has received quite a lot of interest from academic, industrial and general user communities in the last decades owing to the ease of availability of computing and hardware resources and increasing demands for human computer interaction methods. In this paper, a detailed lit- erature review is made on the recent advances in eye gaze research, and information in statistical format is presented to highlight the diversity in various aspects such as platforms, setups, users, algorithms and performance measures existing between different branches of this field.
Several different gaze tracking algorithms and their respective advantages and disadvantages were analyzed in Section III. Currently gaze based HCI systems are capable of achieving high speed input and control operations, leading to their implementation in a variety of user platforms and applications, which were discussed in Section IV. Typically, gaze tracking systems at present are capable of determin- ing 3D point of gaze in real time with unconstrained head movement and achieve around 0.5 degrees of angular res- olution. However, limitations arising due to gaze tracking camera quality, random illumination changes, user wearing glasses and platform vibrations are not well characterized in contemporary eye gaze research.
The literature review also raises a major question with respect to the consistency and accuracy that can be obtained from the gaze estimation systems when they operate under real world conditions, if they are not properly evaluated.
A variety of factors may affect eye gaze tracking in differ- ent platforms, making their performance unpredictable and ultimately questioning their usability in present and future applications. Effects of head movement, user distance and viewing angle, display properties of the setup are still poorly studied, as discussed in Section V. In their presence, practical system performance may differ significantly from expected values and eye gaze may lose its applicability in different consumer use cases.
Further, there is a clear lack of homogeneity in gaze per- formance metrics as pointed out in the tables of Section V. Some performance measures used, for example: detection
16514 VOLUME 5, 2017
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
rate or accuracy percentage is difficult to interpret physically and the variety in reporting formats makes inter-comparisons between different systems and algorithms impossible.
Keeping these in mind, the concept of a performance evaluation framework is proposed that will provide practical performance estimates of gaze tracking systems and adopt a uniform set of accuracy metrics for specifying performance. This is an on-going research activity and the details of the evaluation methods to be included in this framework are currently under development for different gaze estimation platforms, and will be included in a subsequent paper.
ACKNOWLEDGMENT The authors would like to thank Prof. Christopher Dainty and Dr. Claudia Costache for reviewing and providing their valuable feedback on the manuscript.
REFERENCES [1] A. O. Mohamed, M. P. da Silva, and V. Courboulay. (2007). A His-
tory of Eye Gaze Tracking, Rapport Interne. [Online]. Available: https://hal.archives-ouvertes.fr/hal-00215967
[2] A. T. Duchowski, Behavior Research Methods, Instruments, & Comput- ers, vol. 34. New York, NY, USA: Springer-Verlag, 2002, p. 455.
[3] C. H. Morimoto and M. R. M. Mimica, ‘‘Eye gaze tracking techniques for interactive applications,’’ Comput. Vis. Image Understand., vol. 98, no. 1, pp. 4–24, 2005.
[4] C. H. Morimoto, A. Amir, and M. Flickner, ‘‘Detecting eye position and gaze from a single camera and 2 light sources,’’ in Proc. Int. Conf. Pattern Recogn., vol. 4. 2002, pp. 314–317.
[5] P. M. Corcoran, F. Nanu, S. Petrescu, and P. Bigioi, ‘‘Real-time eye gaze tracking for gaming design and consumer electronics systems,’’ IEEE Trans. Consum. Electron., vol. 58, no. 2, pp. 347–355, May 2012.
[6] D. W. D. Hansen, J. S. J. Agustin, and A. Villanueva, ‘‘Homography normalization for robust gaze estimation in uncalibrated setups,’’ in Proc. Symp. Eye-Tracking Res. Appl., vol. 1. Jul. 2015, pp. 13–20.
[7] S. Y. Gwon, C. W. Cho, H. C. Lee, W. O. Lee, and K. R. Park, ‘‘Robust eye and pupil detection method for gaze tracking,’’ Int. J. Adv. Robot. Syst., vol. 10, no. 2, p. 98, 2013.
[8] H. C. Lee, D. T. Luong, C. W. Cho, E. C. Lee, and K. R. Park, ‘‘Gaze tracking system at a distance for controlling IPTV,’’ IEEE Trans. Consum. Electron., vol. 56, no. 4, pp. 2577–2583, Nov. 2010.
[9] A. Lanata, G. Valenza, A. Greco, and E. P. Scilingo, ‘‘Robust head mounted wearable eye tracking system for dynamical calibration,’’ J. Eye Movement Res., vol. 8, no. 5, pp. 1–15, Nov. 2015.
[10] E. S. Kim et al., ‘‘Development of an untethered, mobile, low-cost head- mounted eye tracker,’’ in Proc.Symp.EyeTrack.Res.Appl., vol. 14. 2014, pp. 247–250.
[11] D. Li, J. Babcock, and D. J. Parkhurst, ‘‘openEyes: A low-cost head- mounted eye-tracking solution,’’ in Proc. Symp. Eye Tracking Res. Appl., New York, NY, USA, 2006, pp. 95–100.
[12] W. J. Ryan, A. T. Duchowski, and S. T. Birchfield, ‘‘Limbus/pupil switch- ing for wearable eye tracking under variable lighting conditions,’’ in Proc. ETRA, New York, NY, USA, 2008, pp. 61–64.
[13] T. Ishikawa, S. Baker, I. Matthews, and T. Kanade, ‘‘Passive driver gaze tracking with active appearance models,’’ in Proc. World Congr. Intell. Transp. Syst., 2004, pp. 1–12.
[14] J. P. Batista, ‘‘A real-time driver visual attention monitoring system,’’ in Proc. Conf. IbPRIA, vol. 3522. Estoril, Portugal, 2005, pp. 200–208.
[15] A. Tawari, K. H. Chen, and M. M. Trivedi, ‘‘Where is the driver looking: Analysis of head, eye and iris for robust gaze zone estimation,’’ in Proc. 17th Int. IEEE Conf. Intell. Transp. Syst., Oct. 2014, pp. 988–994.
[16] Y. Liang, M. L. Reyes, and J. D. Lee, ‘‘Real-time detection of driver cognitive distraction using support vector machines,’’ IEEE Trans. Intell. Transp. Syst., vol. 8, no. 2, pp. 340–350, Jun. 2007.
[17] Q. Ji and X. Yang, ‘‘Real-time eye, gaze, and face pose tracking for monitoring driver vigilance,’’ Real-Time Imag., vol. 8, no. 5, pp. 357–377, Oct. 2002.
[18] E. Wood and A. Bulling, ‘‘EyeTab: Model-based gaze estimation on unmodified tablet computers,’’ in Proc. ETRA, New York, NY, USA, 2014, pp. 207–210.
[19] T. Nagamatsu, M. Yamamoto, and H. Sato, ‘‘MobiGaze: Development of a gaze interface for handheld mobile devices,’’ in Proc. Extended Abstracts Hum. Factors Comput. Syst., New York, NY, USA, 2010, pp. 3349–3354.
[20] A. Hyrskykari, P. Majaranta, and K.-J. Räihä, ‘‘From gaze control to attentive interfaces,’’ in Proc. HCII, Las Vegas, NV, USA, Jul. 2005, pp. 1–10.
[21] T. Pfeiffer, ‘‘Towards gaze interaction in immersive virtual reality: Eval- uation of a monocular eye tracking set-up,’’ in Proc. Virtuelle Erweiterte Realität-Fünfter Workshop GI-Fachgruppe, 2008, pp. 81–92.
[22] J. Jimenez, D. Gutierrez, P. Latorre, and U. De Zaragoza, ‘‘Gaze-based interaction for virtual environments,’’ J. Univ. Comput. Sci., vol. 14, no. 19, pp. 3085–3098, 2008.
[23] S. Bazrafkan, A. Kar, and C. Costache, ‘‘Eye gaze for consumer electron- ics: Controlling and commanding intelligent systems,’’ IEEE Consum. Electron. Mag., vol. 4, no. 4, pp. 65–71, Oct. 2015.
[24] A. Kar and P. Corcoran, ‘‘Towards the development of a standardized performance evaluation framework for eye gaze estimation systems in consumer platforms,’’ in Proc. IEEE Int. Conf. Syst., Man, Cybern., Budapest, Hungary, Oct. 2016, pp. 002061–002066.
[25] A. Kar, S. Bazrafkan, C. Costache, and P. Corcoran, ‘‘Eye-gaze systems— -An analysis of error sources and potential accuracy in consumer electron- ics use cases,’’ in Proc. IEEE Int. Conf. Consum. Electron. (ICCE), Las Vegas, NV, USA, 2016, pp. 319–320.
[26] D. W. Hansen and Q. Ji, ‘‘In the eye of the beholder: A survey of models for eyes and gaze,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 32, no. 3, pp. 478–500, Mar. 2010.
[27] P. Majaranta and A. Bulling, ‘‘Eye tracking and eye-based human- computer interaction,’’ in Advances inPhysiologicalComputing. London, U.K.: Springer-Verlag, 2014, pp. 17–39.
[28] A. Poole and L. J. Ball, ‘‘Eye tracking in human-computer interaction and usability research: Current status and future prospects,’’ in Encyclopedia of Human-Computer Interaction. Pennsylvania, PA, USA: Idea Group, Inc, 2005, pp. 211–219.
[29] J. H. Goldberg and X. P. Kotval, ‘‘Computer interface evaluation using eye movements: Methods and constructs,’’ Int. J. Ind. Ergon., vol. 24, no. 6, pp. 631–645, Oct. 1999.
[30] D. D. Salvucci and J. H. Goldberg, ‘‘Identifying fixations and saccades in eye-tracking protocols,’’ in Proc. Eye Tracking Res. Appl. Symp., 2000, pp. 71–78.
[31] B. B. Velichkovsky, M. A. Rumyantsev, and M. A. Morozov, ‘‘New solution to the midas touch problem: Identification of visual com- mands via extraction of focal fixations,’’ Procedia Comput. Sci., vol. 39, pp. 75–82, Dec. 2014.
[32] E. D. Guestrin and M. Eizenman, ‘‘General theory of remote gaze estima- tion using the pupil center and corneal reflections,’’ IEEE Trans. Biomed. Eng., vol. 53, no. 6, pp. 1124–1133, Jun. 2006.
[33] H. C. Lee et al., ‘‘Remote gaze tracking system on a large display,’’ Sensors, vol. 13, no. 10, pp. 13439–13463, 2013.
[34] O. Špakov and D. Miniotas, ‘‘Gaze-based selection of standard-size menu items,’’ in Proc. 7th Int. Conf. Multimodal Interfaces, New York, NY, USA, 2005, pp. 124–128
[35] L. E. Sibert and R. J. K. Jacob, ‘‘Evaluation of eye gaze interaction,’’ in Proc. SIGCHI Conf. Hum. Factors Comput. Syst., New York, NY, USA, 2000, pp. 281–288.
[36] M. Kumar, A. Paepcke, and T. Winograd, ‘‘EyePoint,’’ in Proc. SIGCHI Conf. Hum. Factors Comput. Syst., vol. 7. 2007, p. 421.
[37] Z. Zhu and Q. Ji, ‘‘Novel eye gaze tracking techniques under natural head movement,’’ IEEE Trans. Biomed. Eng., vol. 54, no. 12, pp. 2246–2260, Dec. 2007.
[38] K. Harezlak, P. Kasprowski, and M. Stasch, ‘‘Towards accurate eye tracker calibration—Methods and procedures,’’ Procedia Comput. Sci., vol. 35, pp. 1073–1081, Sep. 2014.
[39] S. O. Ba and J.-M. Odobez, ‘‘Multiperson visual focus of attention from head pose and meeting contextual cues,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 1, pp. 101–116, Jan. 2011.
[40] Accuracy and Precision Test Method for Remote Eye Trackers, Test Specification Version: 2.1.1, Tobii Technology AB, Sweden, Feb. 2011.
[41] P. Blignaut, ‘‘Mapping the pupil-glint vector to gaze coordinates in a simple video-based eye tracker,’’ J. Eye Movement Res., vol. 7, no. 1, pp. 1–11, 2013.
VOLUME 5, 2017 16515
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
[42] Z. R. Cherif, A. Nait-Ali, J. F. Motsch, and M. O. Krebs, ‘‘An adap- tive calibration of an infrared light device used for gaze tracking,’’ in Proc. 19th IEEE Instrum. Meas. Technol. Conf., vol. 2. May 2002, pp. 1029–1033.
[43] X. L. C. Brolly and J. B. Mulligan, ‘‘Implicit calibration of a remote gaze tracker,’’ in Proc. Conf. Comput. Vis. Pattern Recogn. Workshop, Jun. 2004, p. 134.
[44] R. J. J. Cerrolaza, A. Villanueva, and R. Cabeza, ‘‘Taxonomic study of polynomial regressions applied to the calibration of video-oculographic systems,’’ in Proc. Symp. Eye Tracking Res. Appl., vol. 1. 2008, pp. 259–266.
[45] Z. Zhu and Q. Ji, ‘‘Eye and gaze tracking for interactive graphic display,’’ Mach. Vis. Appl., vol. 15, no. 3, pp. 139–148, Jul. 2004.
[46] C. Jian-Nan, Z. Chuang, Y. Yan-Tao, L. Yang, and Z. Han, ‘‘Eye gaze calculation based on nonlinear polynomial and generalized regression neural network,’’ in Proc. 5th Int. Conf. Natural Comput., Aug. 2009, pp. 617–623.
[47] C. Ma, K.-A. Choi, B.-D. Choi, and S.-J. Ko, ‘‘Robust remote gaze estimation method based on multiple geometric transforms,’’ Opt. Eng., vol. 54, no. 8, p. 83103, Aug. 2015.
[48] Z. Zhu, Q. Ji, and K. P. Bennett, ‘‘Nonlinear eye gaze mapping function estimation via support vector regression,’’ in Proc. 18th Int. Conf. Pattern Recogn., Hong Kong, China, Aug. 2006, pp. 1132–1135.
[49] J. Zhu and J. Yang, ‘‘Subpixel eye gaze tracking,’’ in Proc. 5th IEEE Int. Conf. Autom. Face Gesture Recogn., May 2002, pp. 131–136.
[50] J. Wang, G. Zhang, and J. Shi, ‘‘2D gaze estimation based on pupil-glint vector using an artificial neural network,’’ Appl. Sci., vol. 6, no. 6, p. 174, 2016.
[51] Y.-G. Shin, K.-A. Choi, S.-T. Kim, C.-H. Yoo, and S.-J. Ko, ‘‘A novel 2-D mapping-based remote eye gaze tracking method using two IR light sources,’’ in Proc. IEEE Int. Conf. Consum. Electron., Las Vegas, NV, USA, Jan. 2015, pp. 190–191.
[52] J. Park, T. Jung, and K. Yim, ‘‘Implementation of an eye gaze tracking system for the disabled people,’’ in Proc. IEEE 29th Int. Conf. Adv. Inf. Netw. Appl., Gwangiu, South Korea, Mar. 2015, pp. 904–908.
[53] A. Meyer, M. Böhme, T. Martinetz, and E. Barth, ‘‘A single-camera remote eye tracker,’’ Perception and Interactive Technologies (Lecture Notes in Computer Science), vol. 4021. New York, NY, USA: Springer, 2006, pp. 208–211.
[54] C. Hennessey, B. Noureddin, and P. Lawrence, ‘‘A single camera eye-gaze tracking system with free head motion,’’ Measurement, vol. 1, pp. 27–29, Mar. 2006.
[55] C.-C. Lai, S.-W. Shih, and Y.-P. Hung, ‘‘Hybrid method for 3-D gaze tracking using glint and contour features,’’ IEEE Trans. Circuits Syst. Video Technol., vol. 25, no. 1, pp. 24–37, Jan. 2015.
[56] T. Ohno and N. Mukawa, ‘‘A free-head, simple calibration, gaze tracking system that enables gaze-based interaction,’’ in Proc. Eye Tracking Res. Appl. Symp. Eye Tracking Res. Appl., 2004, pp. 115–122.
[57] D. Beymer and M. Flickner, ‘‘Eye gaze tracking using an active stereo head,’’ in Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recogn., vol. 2. Jun. 2003, pp. II-451–II-458.
[58] S.-W. Shih and J. Liu, ‘‘A novel approach to 3-D gaze tracking using stereo cameras.,’’ IEEE Trans. Syst. Man, Cybern. B, Cybern., vol. 34, no. 1, pp. 234–245, Feb. 2004.
[59] T. Nagamatsu, Y. Iwamoto, J. Kamahara, N. Tanaka, and M. Yamamoto, ‘‘Gaze estimation method based on an aspherical model of the cornea: Surface of revolution about the optical axis of the eye,’’ in Proc. Symp. Eye-Tracking Res. Appl., New York, NY, USA, 2010, pp. 255–258.
[60] T. Nagamatsu, J. Kamahara, and N. Tanaka, ‘‘Calibration-free gaze tracking using a binocular 3D eye model,’’ in Proc. Extended Abstracts Hum. Factors Comput. Syst., New York, NY, USA, 2009, pp. 3613–3618.
[61] D. Model and M. Eizenman, ‘‘User-calibration-free remote eye- gaze tracking system with extended tracking range,’’ in Proc. 24th Can. Conf. Elect. Comput. Eng., Niagara Falls, ON, Canada, 2011, pp. 001268–001271.
[62] K. Wang and Q. Ji, ‘‘Real time eye gaze tracking with Kinect,’’ in Proc. 23rd Int. Conf. Pattern Recogn., Cancún, Mexico, 2016, pp. 2752–2757.
[63] X. Zhou, H. Cai, Z. Shao, H. Yu, and H. Liu, ‘‘3D eye model-based gaze estimation from a depth sensor,’’ in Proc. IEEE Int. Conf. Robot. Biomimetics, Qingdao, China, Dec. 2016, pp. 369–374.
[64] L. Jianfeng and L. Shigang, ‘‘Eye-model-based gaze estimation by RGB-D camera,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recogn. Workshops, Columbus, OH, USA, Jun. 2014, pp. 606–610.
[65] J. Huang, Q. Cai, Z. Liu, N. Ahuja, and Z. Zhang, ‘‘Towards accurate and robust cross-ratio based gaze trackers through learning from simulation,’’ in Proc. ETRA, 2014, pp. 75–82.
[66] D. H. Yoo and M. J. Chung, ‘‘A novel non-intrusive eye gaze estima- tion using cross-ratio under large head motion,’’ Comput. Vis. Image Understand., vol. 98, no. 1, pp. 25–51, Apr. 2005.
[67] F. L. Coutinho and C. H. Morimoto, ‘‘Free head motion eye gaze tracking using a single camera and multiple light sources,’’ in Proc. 19th Brazilian Symp. Comput. Graph. Image Process., 2006, pp. 171–178.
[68] F. L. Coutinho and C. H. Morimoto, ‘‘Augmenting the robustness of cross- ratio gaze tracking methods to head movement,’’ in Proc. Symp. Eye Tracking Res. Appl., 2012, pp. 59–66.
[69] J. J. Kang, E. D. Guestrin, W. J. Maclean, and M. Eizenman, ‘‘Simplifying the cross-ratios method of point-of-gaze estimation,’’ in Proc. 30th Can. Med. Biol. Eng. Conf., 2007, pp. 1–4.
[70] I. Bacivarov, M. Ionita, and P. Corcoran, ‘‘Statistical models of appear- ance for eye tracking and eye-blink detection and measurement,’’ IEEE Trans. Consum. Electron., vol. 54, no. 3, pp. 1312–1328, Aug. 2008.
[71] P. Koutras and P. Maragos, ‘‘Estimation of eye gaze direction angles based on active appearance models,’’ in Proc. IEEE Int. Conf. Image Process., Quebec City, QC, Canada, Sep. 2015, pp. 2424–2428.
[72] F. Lu, Y. Gao, and X. Chen, ‘‘Estimating 3D gaze directions using unlabeled eye images via synthetic iris appearance fitting,’’ IEEE Trans. Multimedia, vol. 18, no. 9, pp. 1772–1782, Sep. 2016.
[73] Y.-L. Wu, C.-T. Yeh, W.-C. Hung, and C.-Y. Tang, ‘‘Gaze direction estimation using support vector machine with active appearance model,’’ Multimed. Tools Appl., pp. 1–26, 2012.
[74] H. L. H. Lu, C. W. C. Wang, and Y. C. Y. Chen, ‘‘Gaze tracking by binocular vision and LBP features,’’ in Proc. 19th Int. Conf. Pattern Recogn., Tampa, FL, USA, 2008, pp. 1–4.
[75] C. M. Yilmaz and C. Kose, ‘‘Local binary pattern histogram features for on-screen eye-gaze direction estimation and a comparison of appear- ance based methods,’’ in Proc. 39th Int. Conf. Telecommun. Signal Process., Vienna, Austria, 2016, pp. 693–696.
[76] S. Chen and C. Liu, ‘‘Eye detection using discriminatory Haar features and a new efficient SVM,’’ Image Vis. Comput., vol. 33, pp. 68–77, Jan. 2015.
[77] Y. Li, X. Xu, N. Mu, and L. Chen, ‘‘Eye-gaze tracking system by haar cascade classifier,’’ in Proc. IEEE 11th Conf. Ind. Electron. Appl., Hefei, China, Jun. 2016, pp. 564–567.
[78] S. Baluja and D. Pomerleau, ‘‘Non-intrusive gaze tracking using artificial neural networks,’’ Carnegie Mellon Univ. Pittsburgh, Pittsburgh, PA, USA, Tech. Rep. FS-93-04, 1994, pp. 753–760.
[79] T. Schneider, B. Schauerte, and R. Stiefelhagen, ‘‘Manifold alignment for person independent appearance-based gaze estimation,’’ in Proc. 22nd Int. Conf. Pattern Recogn., 2014, pp. 1167–1172.
[80] L. Yu, J. Xu, and S. Huang, ‘‘Eye-gaze tracking system based on particle swarm optimization and BP neural network,’’ in Proc. 12th World Congr. Intell. Control Autom., Guilin, China, 2016, pp. 1269–1273.
[81] C. C. Lai, Y. T. Chen, K. W. Chen, S. C. Chen, S. W. Shih, and Y. P. Hung, ‘‘Appearance-based gaze tracking with free head movement,’’ in Proc.Int. Conf. Pattern Recogn., vol. 1. 2014, pp. 1869–1873.
[82] D. Benavides and D. L. Borges, ‘‘Eye detection in unrestrained set- tings using efficient match kernels and SVM classification,’’ in Proc. IEEE Int. Conf. Syst., Man, Cybern. (SMC), Budapest, Hungary, 2016, pp. 002801–002806.
[83] H. Huang and J. Wechsler, ‘‘Eye location using genetic algorithm,’’ in Proc. 2nd Int. Conf. Audio Video-Based Biometric Pers. Authentication, 1999, pp. 130–135.
[84] A. George and A. Routray, ‘‘Real-time eye gaze direction classification using convolutional neural network,’’ in Proc. Int. Conf. Signal Process. Commun., Bengaluru, India, 2016, pp. 1–5.
[85] R. Konrad, Near-Eye Display Gaze Tracking Via Convolutional Neu- ral Networks. [Online]. Available: https://web.stanford.edu/class/cs231a/ prev_projects_2016/eye-display-gaze-2.pdf
[86] K. Krafka et al., ‘‘Eye tracking for everyone,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recogn., Las Vegas, NV, USA, Jun. 2016, pp. 2176–2184.
[87] X. Zhang, Y. Sugano, M. Fritz, and A. Bulling, ‘‘Appearance-based gaze estimation in the wild,’’ in Proc. IEEE Int. Conf. Comput. Vis. Pattern Recogn., Jun. 2015, pp. 4511–4520.
16516 VOLUME 5, 2017
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
[88] W. Wang, Y. Huang, and R. Zhang, ‘‘Driver gaze tracker using deformable template matching,’’ in Proc. IEEE Int. Conf. Veh. Electron. Safety, Jul. 2011, pp. 244–247.
[89] M. J. T. Reinders, ‘‘Eye tracking by template matching using an automatic codebook generation scheme,’’ in Proc. 3rd Annu. Conf. Adv. School Comput. Imag., Heijen, The Netherlands, Jun. 1997, pp. 215–221.
[90] S. Ramadan, W. Abd-Almageed, and C. Smith, ‘‘Eye tracking using active deformable models,’’ in Proc. 3rd Indian Conf. Comput. Vis., Graph. Image Process., India, 2002.
[91] I. F. Ince and J. W. Kim, ‘‘A 2D eye gaze estimation system with low- resolution webcam images,’’ EURASIP J. Adv. Signal Process., vol. 2011, no. 1, p. 40, Dec. 2011.
[92] S. Zhai, C. Morimoto, and S. Ihde, ‘‘Manual and gaze input cascaded (MAGIC) pointing,’’ in Proc. SIGCHI Conf. Hum. Factors Comput. Syst., New York, NY, USA, 1999, pp. 246–253.
[93] M. U. Ghani, S. Chaudhry, M. Sohail, and M. N. Geelani, ‘‘Gaze- Pointer: A real time mouse pointer control implementation based on eye gaze tracking,’’ in Proc. 16th Int. Multi Topic Conf., 2013, pp. 154–159.
[94] J. S. Agustin, J. C. Mateo, J. P. Hansen, and A. Villanueva, ‘‘Evaluation of the potential of gaze input for game interaction,’’ PsychNology J., vol. 7, no. 2, pp. 213–236, 2009.
[95] P. Kasprowski and K. Harężlak, ‘‘Cheap and Easy PIN Entering Using Eye Gaze,’’ Ann. UMCS, Inf., vol. 14, no. 1, pp. 75–83, 2014.
[96] M. Kumar, T. Garfinkel, D. Boneh, and T. Winograd, ‘‘Reducing shoulder-surfing by using gaze-based password entry,’’ in Proc.3rdSymp. Usable Privacy Secur., vol. 7. 2007, pp. 13–19.
[97] A. Bulling, F. Alt, and A. Schmidt, ‘‘Increasing the security of gaze-based cued-recall graphical passwords using saliency masks,’’ in Proc. ACM Annu. Conf. Hum. Factors Comput. Syst., vol. 12. 2012, p. 3011.
[98] S. T. Iqbal and B. P. Bailey, ‘‘Using eye gaze patterns to identify user tasks,’’ in Grace Hopper Celebration of Women in Computing. Lausanne, Switzerland: Frontiers, 2004, p. 6.
[99] A. Doshi and M. M. Trivedi, ‘‘Head and gaze dynamics in visual atten- tion and context learning,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recogn., Jun. 2009, pp. 77–84.
[100] J. W. Lee, C. W. Cho, K. Y. Shin, E. C. Lee, and K. R. Park, ‘‘3D gaze tracking method using Purkinje images on eye optical model and pupil,’’ Opt. Lasers Eng., vol. 50, no. 5, pp. 736–751, 2012.
[101] W. W. Abbott and A. A. Faisal, ‘‘Ultra-low-cost 3D gaze estimation: An intuitive high information throughput compliment to direct brain- machine interfaces,’’ J. Neural Eng., vol. 9, no. 4, p. 46016, 2012.
[102] E. Schneider et al., ‘‘Gaze-aligned head-mounted camera with pan, tilt, and roll motion control for medical documentation and teaching appli- cations,’’ in Proc. IEEE Int. Conf. Syst., Man Cybern., Taipei, Taiwan, Oct. 2006, pp. 327–331.
[103] M. Y. Kim, S. Yang, and D. Kim, ‘‘Head-mounted binocular gaze detec- tion for selective visual recognition systems,’’ Sens. Actuators A, Phys., vol. 187, pp. 29–36, Nov. 2012.
[104] K. Takemura, K. Takahashi, J. Takamatsu, and T. Ogasawara, ‘‘Esti- mating 3-D point-of-regard in a real environment using a head-mounted eye-tracking system,’’ IEEE Trans. Human-Mach. Syst., vol. 44, no. 4, pp. 531–536, Aug. 2014.
[105] X. Long, O. K. Tonguz, and A. Kiderman, ‘‘A high speed eye tracking system with robust pupil center estimation algorithm,’’ in Proc. 29th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., Lyon, France, Aug. 2007, pp. 3331–3334.
[106] E. Schneider, T. Dera, K. Bard, S. Bardins, G. Boening, and T. Brand, ‘‘Eye movement driven head-mounted camera: It looks where the eyes look,’’ in Proc. IEEE Int. Conf. Syst., Man Cybern., vol. 3. Oct. 2005, pp. 2437–2442.
[107] S. H. Lee, J.-Y. Lee, and J. S. Choi, ‘‘Design and implementation of an interactive HMD for wearable AR system,’’ in Proc. 17th Korea- Jpn. Joint Workshop Frontiers Comput. Vis., Ulsan, South Korea, 2011, pp. 1–6.
[108] E. C. Lee, Y. J. Ko, and K. R. Park, ‘‘Gaze tracking based on active appearance model and multiple support vector regression on mobile devices,’’ Opt. Eng., vol. 48, no. 7, pp. 077002-1–077002-11, Jul. 2009.
[109] T. Toyama, A. Dengel, W. Suzuki, and K. Kise, ‘‘Wearable read- ing assist system: Augmented reality document combining document retrieval and eye tracking,’’ in Proc. Int. Conf. Doc. Anal. Recogn., 2013, pp. 30–34.
[110] T. Piumsomboon, G. Lee, R. W. Lindeman, and M. Billinghurst, ‘‘Explor- ing natural eye-gaze-based interaction for immersive virtual reality,’’ in Proc. IEEE Symp. 3D User Interfaces, Los Angeles, CA, USA, Mar. 2017, pp. 36–39.
[111] M. Niener, M. Stamminger, J. Thies, C. Theobalt, and M. Zollhöfer. (2016). ‘‘FaceVR: Real-time facial reenactment and eye gaze control in virtual reality.’’ [Online]. Available: https://arxiv.org/abs/1610.03151
[112] N. Sidorakis, G. A. Koulieris, and K. Mania, ‘‘Binocular eye-tracking for the control of a 3D immersive multimedia user interface,’’ in Proc. IEEE 1st Workshop Everyday Virtual Reality, Arles, France, Mar. 2015, pp. 15– 18.
[113] A. M. Soccini, ‘‘Gaze estimation based on head movements in vir- tual reality applications using deep learning,’’ in Proc. IEEE Virtual Reality, Los Angeles, CA, USA, Mar. 2017, pp. 413–414.
[114] A. Plopski, J. Orlosky, Y. Itoh, C. Nitschke, K. Kiyokawa, and G. Klinker, ‘‘Automated spatial calibration of HMD systems with unconstrained eye- cameras,’’ in Proc. IEEE Int. Symp. Mixed Augmented Reality, Mérida, Mexico, Sep. 2016, pp. 94–99.
[115] J. Y. Lee, H. M. Park, S. H. Lee, T. E. Kim, and J. S. Choi, ‘‘Design and implementation of an augmented reality system using gaze interaction,’’ in Proc. Int. Conf. Inf. Sci. Appl., Jeju Island, 2011, pp. 1–8.
[116] M. Bâce, T. Leppänen, D. G. de Gomez, and A. R. Gomez, ‘‘ubiGaze: Ubiquitous augmented reality messaging using gaze gestures,’’ in Proc. SIGGRAPH ASIA Mobile Graph. Interact. Appl., 2016, pp. 11:1–11:5.
[117] J. Orlosky, T. Toyama, K. Kiyokawa, and D. Sonntag, ‘‘ModulAR: Eye- controlled vision augmentations for head mounted displays,’’ IEEETrans. Vis. Comput. Graph., vol. 21, no. 11, pp. 1259–1268, Nov. 2015.
[118] X. Liu, F. Xu, and K. Fujimura, ‘‘Real-time eye detection and tracking for driver observation under various light conditions,’’ in Proc. IEEE Intell. Vehicle Symp., vol. 2. Jun. 2002, pp. 344–351.
[119] S. J. Lee, J. Jo, H. G. Jung, K. R. Park, and J. Kim, ‘‘Real-time gaze estimator based on driver’s head orientation for forward collision warning system,’’ IEEE Trans. Intell. Transp. Syst., vol. 12, no. 1, pp. 254–267, Mar. 2011.
[120] M.-C. Chuang, R. Bala, E. A. Bernal, P. Paul, and A. Burry, ‘‘Estimat- ing gaze direction of vehicle drivers using a smartphone camera,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recogn. Workshops, Jun. 2014, pp. 165–170.
[121] A. Tawari and M. M. Trivedi, ‘‘Robust and continuous estimation of driver gaze zone by dynamic analysis of multiple face videos,’’ in Proc. IEEE Intell. Vehicles Symp., Jun. 2014, pp. 344–349.
[122] J. H. Oh and N. Kwak, ‘‘Recognition of a Driver’s gaze for vehi- cle headlamp control,’’ IEEE Trans. Veh. Technol., vol. 61, no. 5, pp. 2008–2017, Jun. 2012.
[123] D. D. Salvucci and A. Liu, ‘‘The time course of a lane change: Driver control and eye-movement behavior,’’ Transp. Res. F, Traffic Psychol. Behaviour, vol. 5, no. 2, pp. 123–132, 2002.
[124] I. H. Choi and Y. G. Kim, ‘‘Head pose and gaze direction tracking for detecting a drowsy driver,’’ Appl. Math. Inf. Sci., vol. 9, no. 2, pp. 505–512, 2015.
[125] Q. Ji, Z. Zhu, and P. Lan, ‘‘Real-time nonintrusive monitoring and pre- diction of driver fatigue,’’ IEEE Trans. Veh. Technol., vol. 53, no. 4, pp. 1052–1068, Jul. 2004.
[126] L. M. Bergasa and J. Nuevo, ‘‘Real-time system for monitoring driver vigilance,’’ in Proc. IEEE Int. Symp. Ind. Electron., vol. 3. Jun. 2005, pp. 1303–1308.
[127] X. Sun, L. Xu, and J. Yang, ‘‘Driver fatigue alarm based on eye detection and gaze estimation,’’ Proc. SPIE, vol. 6786, pp. 678612-1–678612-6, Nov. 2007.
[128] F. Vicente, Z. Huang, X. Xiong, F. D. L. Torre, W. Zhang, and D. Levi, ‘‘Driver gaze tracking and eyes off the road detection system,’’ IEEE Trans. Intell. Transp. Syst., vol. 16, no. 4, pp. 2014–2027, Aug. 2015.
[129] R. Zheng, K. Nakano, H. Ishiko, K. Hagita, M. Kihira, and T. Yokozeki, ‘‘Eye-gaze tracking analysis of driver behavior while interacting with navigation systems in an urban area,’’ IEEE Trans. Human-Mach. Syst., vol. 46, no. 4, pp. 546–556, Aug. 2016.
[130] O. Stan, L. Miclea, and A. Centea, ‘‘Eye-gaze tracking method driven by raspberry PI applicable in automotive traffic safety,’’ in Proc. 2nd Int. Conf. Artif. Intell., Modelling Simulation, Madrid, Spain, 2014, pp. 126–130.
[131] T. Kowsari, S. S. Beauchemin, M. A. Bauer, D. Laurendeau, and N. Teasdale, ‘‘Multi-depth cross-calibration of remote eye gaze trackers and stereoscopic scene systems,’’ in Proc. IEEE Intell. Vehicles Symp., Dearborn, MI, USA, Jun. 2014, pp. 1245–1250.
VOLUME 5, 2017 16517
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
[132] S. Jha and C. Busso, Analyzing the Relationship Between Head Pose and Gaze to Model Driver Visual Attention. Rio de Janeiro, Brazil: IEEE, 2016, pp. 1–6.
[133] L. Fridman, J. Lee, B. Reimer, and T. Victor, ‘‘‘Owl’ and ‘Lizard’: Patterns of head pose and eye pose in driver gaze classification,’’ IET Comput. Vis., vol. 10, no. 4, pp. 308–313, 2016.
[134] V. Vaitukaitis and A. Bulling, ‘‘Eye gesture recognition on portable devices,’’ in Proc. ACM Conf. Ubiquitous Comput., vol. 12. 2012, p. 711.
[135] K. Lukander, ‘‘A system for tracking gaze on handheld devices,’’ Behav- ior Res. Meth., vol. 38, no. 4, pp. 660–666, 2006.
[136] C. Holland, A. Garza, E. Kurtova, J. Cruz, and O. Komogortsev, ‘‘Usabil- ity evaluation of eye tracking on an unmodified common tablet,’’ in Proc. Extended Abstracts Hum. Factors Comput. Syst., vol. 13. 2013, p. 295.
[137] T. Imabuchi, O. Dicky, A. Prima, H. Kikuchi, Y. Horie, and H. Ito, ‘‘Visible-spectrum Remote Eye Tracker for Gaze Communication,’’ Proc. SPIE, vol. 9443, p. 944333, Mar. 2015.
[138] P. Viola and M. Jones, ‘‘Rapid object detection using a boosted cascade of simple features,’’ in Proc. Comput. Vis. Pattern Recogn., vol. 1. 2001, pp. I-511–I-518.
[139] H. Elleuch, A. Wali, and A. M. Alimi, ‘‘Smart tablet monitoring by a real-time head movement and eye gestures recognition system,’’ in Proc. Int. Conf. Future Internet Things Cloud, Barcelona, Spain, 2014, pp. 393–398.
[140] C. Pino and I. Kavasidis, ‘‘Improving mobile device interaction by eye tracking analysis,’’ in Proc. Comput. Sci. Inf., 2012, pp. 1199–1202.
[141] H. Drewes and A. Schmidt, ‘‘Interacting with the computer using gaze gestures,’’ in Proc. Int.Conf.Hum.-Comput. Interact., 2007, pp. 475–488.
[142] E. Miluzzo, T. Wang, and A. T. Campbell, ‘‘EyePhone: Activating mobile phones with your eyes,’’ in Proc. Workshop Netw., Syst., Appl. Mobile Handhelds, 2010, pp. 15–20.
[143] D. Liu, B. Dong, X. Gao, and H. Wang, ‘‘Exploiting eye tracking for smartphone authentication,’’ in Applied Cryptography and Network Security (Lectures Notes Computer Science), vol. 9092, T. Malkin, V. Kolesnikov, A. Lewko, and M. Polychronakis, Eds. New York, NY, USA: Springer, 2015.
[144] Y.-W. Sun, C.-K. Chiang, and S.-H. Lai, ‘‘Integrating eye tracking and motion sensor on mobile phone for interactive 3D display,’’ Proc. SPIE, vol. 8856, p. 88560F, Sep. 2013.
[145] F. Lu, Y. Sugano, T. Okabe, and Y. Sato, ‘‘Adaptive linear regression for appearance-based gaze estimation,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 36, no. 10, pp. 2033–2046, Oct. 2014.
[146] A. Villanueva, R. Cabeza, and S. Porta, ‘‘Eye tracking: Pupil orientation geometrical modeling,’’ Image Vis. Comput., vol. 24, no. 7, pp. 663–679, 2006.
[147] S.-W. Shih, Y.-T. Wu, and J. Liu, ‘‘A calibration-free gaze tracking technique,’’ in Proc. Int. Conf. Pattern Recogn., vol. 4. Sep. 2000, pp. 201–204.
[148] R. Newman, Y. Matsumoto, S. Rougeaux, and A. Zelinsky, ‘‘Real-time stereo tracking for head pose and gaze estimation,’’ in Proc. 4th IEEE Int. Conf. Autom. Face Gesture Recogn., Grenoble, France, Mar. 2000, pp. 122–128.
[149] Y. Sugano, Y. Matsushita, and Y. Sato, ‘‘Learning-by-synthesis for appearance-based 3D gaze estimation,’’ in Proc.IEEEComput.Soc.Conf. Comput. Vis. Pattern Recogn., Jun. 2014, pp. 1821–1828.
[150] K. Tan, D. J. Kriegman, and N. Ahuja, ‘‘Appearance-based eye gaze estimation,’’ in Proc. WACV, 2002, pp. 191–195.
[151] X. Zhao, X. Zou, and Z. Chi, ‘‘A 3D gaze estimation method based on facial feature tracking,’’ in Proc. Int. Conf. Comput. Healthcare, 2012, pp. 9–12.
[152] Z. Zhang and Q. Cai, ‘‘Improving cross-ratio-based eye tracking tech- niques by leveraging the binocular fixation constraint,’’ in Proc. Symp. Eye Tracking Res. Appl., vol. 14. 2014, pp. 267–270.
[153] C. Yu, M. Scheutz, and P. Schermerhorn, ‘‘Investigating multimodal real- time patterns of joint attention in an HRI word learning task,’’ in Proc. 5th ACM/IEEE Int. Conf. Hum.-Robot Interact., 2010, pp. 309–316.
[154] V. Rantanen et al., ‘‘A wearable, wireless gaze tracker with inte- grated selection command source for human-computer interaction,’’ IEEE Trans. Inf. Technol. Biomed., vol. 15, no. 5, pp. 795–801, Sep. 2011.
[155] M. Stengel, S. Grogorick, M. Eisemann, E. Eisemann, and M. A. Magnor, ‘‘An affordable solution for binocular eye tracking and calibration in head-mounted displays,’’ in Proc. 23rd ACM Int. Conf. Multimedia, New York, NY, USA, 2015, pp. 15–24.
[156] B. R. Pires, M. Hwangbo, M. Devyver, and T. Kanade, ‘‘Visible-spectrum gaze tracking for sports,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recogn. Workshops, Jun. 2013, pp. 1005–1010.
[157] L. Fletcher and A. Zelinsky, ‘‘Driver inattention detection based on eye gaze—Road event correlation,’’ Int. J. Robot. Res., vol. 28, no. 6, pp. 774–801, 2009.
[158] K. Yamashiro et al., ‘‘Automatic calibration of an in-vehicle gaze track- ing system using driver’s typical gaze behavior,’’ in Proc. IEEE Intell. Vehicles Symp., Xi’an, China, Jun. 2009, pp. 998–1003.
[159] A. Doshi and M. M. Trivedi, ‘‘On the roles of eye gaze and head dynamics in predicting driver’s intent to change lanes,’’ IEEE Trans. Intell. Transp. Syst., vol. 10, no. 3, pp. 453–462, Sep. 2009.
[160] P. Smith, M. Shah, and N. da Vitoria Lobo, ‘‘Determining driver visual attention with one camera,’’ IEEETrans. Intell.Transp.Syst., vol. 4, no. 4, pp. 205–218, Dec. 2003.
[161] J. Wu and M. M. Trivedi, ‘‘Simultaneous eye tracking and blink detec- tion with interactive particle filters,’’ EURASIP J. Adv. Signal Process., vol. 2008, Jan. 2008, Art. no. 114.
[162] R. Oyini Mbouna, S. G. Kong, and M.-G. Chun, ‘‘Visual analysis of eye state and head pose for driver alertness monitoring,’’ IEEE Trans. Intell. Transp. Syst., vol. 14, no. 3, pp. 1462–1469, Sep. 2013.
[163] I. Lissoboi and H. Kasai, ‘‘Development of an efficient method for eye detection on mobile CE devices,’’ in Proc. Int. Symp. Comput. Consum. Control, 2012, pp. 337–340.
[164] K. Shibasato, C. Tateyama, H. Ohtsuka, and Y. Shimada, ‘‘Implementa- tion of application by gaze interaction on a tablet computer for challenged people,’’ in Proc. 9th Int. Conf. Sens. Technol., Auckland, New Zealand, 2015, pp. 369–373.
[165] L. H. Yu and M. Eizenman, ‘‘A new methodology for determining point- of-gaze in head-mounted eye tracking systems,’’ IEEE Trans. Biomed. Eng., vol. 51, no. 10, pp. 1765–1773, Oct. 2004.
[166] C. Morimoto, D. Koons, A. Amir, and M. Flickner, ‘‘Pupil detection and tracking using multiple. Light sources,’’ Image Vis. Comput., vol. 18, no. 4, pp. 331–335, 2000.
[167] S. J. Baek, K. A. Choi, C. Ma, Y. H. Kim, and S. J. Ko, ‘‘Eyeball model- based iris center localization for visible image-based eye-gaze tracking systems,’’ IEEE Trans. Consum. Electron., vol. 59, no. 2, pp. 415–421, May 2013.
[168] F. Lu, T. Okabe, Y. Sugano, and Y. Sato, ‘‘A head pose-free approach for appearance-based gaze estimation,’’ in Proc. Brit. Mach. Vis. Conf., 2011, pp. 126.1–126.11.
[169] A. Nakazawa and C. Nitschke, ‘‘Point of gaze estimation through corneal surface reflection in an active illumination environment,’’ in Computer Vision (Lecture Notes in Computer Science), vol. 7573. New York, NY, USA: Springer-Verlag, 2012, pp. 159–172.
[170] D. LeBlanc, A. Forget, and R. Biddle, ‘‘Guessing click-based graphical passwords by eye tracking,’’ in Proc. 8th Int. Conf. Privacy, Secur. Trust, 2010, pp. 197–204.
[171] N. M. M. Hassan and W. Mansor, ‘‘Detection of eye movements for controlling a television,’’ in Proc. IEEE 10th Int. Colloq. Signal Process. Appl., Mar. 2014, pp. 257–260.
[172] R. Alonso, M. Causse, F. Vachon, R. Parise, F. Dehais, and P. Terrier, ‘‘Evaluation of head-free eye tracking as an input device for air traffic control,’’ Ergonomics, vol. 56, no. 2, pp. 246–255, 2013.
[173] J. Weaver, K. Mock, and B. Hoanca, ‘‘Gaze-based password authentica- tion through automatic clustering of gaze points,’’ in Proc. IEEEInt.Conf. Syst., Man Cybern., Oct. 2011, pp. 2749–2754.
[174] K. R. Newman and C. R. Sears, ‘‘Eye gaze tracking reveals different effects of a sad mood induction on the attention of previously depressed and never depressed women,’’ Cognit. Therapy Res., vol. 39, no. 3, pp. 292–306, 2015.
[175] B. Noris, J. B. Keller, and A. Billard, ‘‘A wearable gaze tracking system for children in unconstrained environments,’’ Comput. Vis. Image Under- stand., vol. 115, no. 4, pp. 476–486, 2011.
[176] A. Carbone, F. Martínez, E. Pissaloux, D. Mazeika, and R. Velázquez, ‘‘On the design of a low cost gaze tracker for interaction,’’ Procedia Technol., vol. 3, pp. 89–96, 2012.
[177] Z. Ye et al., ‘‘Detecting eye contact using wearable eye-tracking glasses,’’ in Proc. ACM Conf. Ubiquitous Comput., vol. 12. 2012, p. 699.
[178] N. Kumar, S. Kohlbecher, and E. Schneider, ‘‘A novel approach to video- based pupil tracking,’’ in Proc. IEEE Int. Conf. Syst., Man Cybern., Oct. 2009, pp. 1255–1262.
16518 VOLUME 5, 2017
A. Kar, P. Corcoran: Review and Analysis of Eye-Gaze Estimation Systems, Algorithms, and Performance Evaluation Methods
[179] S. H. Kwon and M. Y. Kim, ‘‘Selective attentional point-tracking through a head-mounted stereo gaze tracker based on trinocular epipolar geom- etry,’’ in Proc. IEEE Int. Instrum. Meas. Technol. Conf., May 2015, pp. 1617–1621.
[180] X. Li and W. G. Wee, ‘‘An efficient method for eye tracking and eye-gazed fov estimation,’’ in Proc. Int. Conf. Image Process., 2009, pp. 2597–2600.
[181] A. Lanata, A. Greco, G. Valenza, and E. P. Scilingo, ‘‘On the tridi- mensional estimation of the gaze point by a stereoscopic wearable eye tracker,’’ in Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., Aug. 2015, pp. 2283–2286.
[182] D. J. Mack, P. Schönle, S. Fateh, T. Burger, Q. Huang, and U. Schwarz, ‘‘An EOG-based, head-mounted eye tracker with 1 kHz sampling rate,’’ in Proc. IEEE Biomed. Circuits Syst. Conf., Atlanta, GA, USA, Oct. 2015, pp. 1–4.
[183] H. Heo, E. Lee, K. Park, C. Kim, and M. Whang, ‘‘A realistic game sys- tem using multi-modal user interfaces,’’ IEEE Trans. Consum. Electron., vol. 56, no. 3, pp. 1364–1372, Aug. 2010.
[184] G. Schiavone, E. Guglielmelli, F. Keller, L. Zollo, and F. Chersi, ‘‘A wear- able ergonomic gaze-tracker for infants,’’ in Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., vol. 10. Aug. 2010, pp. 1283–1286.
[185] R. Mantiuk, M. Kowalik, A. Nowosielski, and B. Bazyluk, ‘‘Do-it- yourself eye tracker: Low-cost pupil-based eye tracker for computer graphics applications,’’ in Advances in Multimedia Modeling (Lect. Notes Comput. Sci.), vol. 7131, K. Schoeffmann, B. Merialdo, A. G. Haupt- mann, C.-W. Ngo, Y. Andreopoulos, and C. Breiteneder, Eds. Berlin, Germany: Springer-Verlag, 2012, pp. 115–125.
[186] H. Murphy and A. T. Duchowski, ‘‘Gaze-contingent level of detail,’’ in Eurographics, J. Roberts, Ed. Manchester, U.K.: Univ. Manchester, 2001, pp. 219–228.
[187] A. T. Duchowski, E. Medlin, N. Cournia, A. Gramopadhye, B. Melloy, and S. Nair, ‘‘3D eye movement analysis for VR visual inspection train- ing,’’ in Proc. Symp. Eye Tracking Res. Appl., vol. 2. 2002, p. 103.
[188] S. S. Mozafari Chanijani, S. S. Bukhari, and A. Dengel, ‘‘Analysis of text layout quality using wearable eye trackers,’’ in Proc. IEEE Int. Conf. Multimedia Expo Workshops, Turin, Italy, 2015, pp. 1–6.
[189] C. Topai, A. Dogan, and N. Gerek, ‘‘A wearable head-mounted sensor- based apparatus for eye tracking applications,’’ in Proc. IEEE Conf. Virtual Environ., Hum.-Comput. Interfaces Meas. Syst., Jul. 2008, pp. 136–139.
[190] Y. Itoh and G. Klinker, ‘‘Interaction-free calibration for optical see- through head-mounted displays based on 3D Eye localization,’’ in Proc. IEEE Symp. 3D User Interfaces, Minneapolis, MN, USA, Mar. 2014, pp. 75–82.
[191] A. Steptoe et al., ‘‘Eye tracking for avatar eye gaze control during object-focused multiparty interaction in immersive collaborative vir- tual environments,’’ in Proc. IEEE Virtual Reality Conf., Mar. 2009, pp. 83–90.
[192] M. Köles and K. Hercegfi, ‘‘Eye tracking precision in a virtual CAVE environment,’’ in Proc. 6th IEEE Int. Conf. Cogn. Infocommun., Oct. 2015, pp. 319–322.
[193] R. G. Bozomitu, V. Cehan, R. G. Lupu, C. Rotariu, and C. Barabaşa, ‘‘A new technique for improving pupil detection algorithm,’’ in Proc. Int. Symp. Signals, Circuits Syst., Iasi, Romania, Jul. 2015, pp. 1–4.
[194] T. Kocejko, J. Ruminski, J. Wtorek, and B. Martin, ‘‘Eye tracking within near-to-eye display,’’ in Proc. 8th Int. Conf. Hum. Syst. Interact., 2015, pp. 166–172.
[195] K. A. I. Essig, M. Pomplun, and H. Ritter, ‘‘A neural network for 3D gaze recording with binocular eye trackers,’’ Int. J. Parallel, Emergent Distrib. Syst., vol. 21, no. 2, pp. 79–95, Apr. 2006.
[196] S. Hillaire, A. Lecuyer, R. Cozot, and G. Casiez, ‘‘Using an eye-tracking system to improve camera motions and depth-of-field blur effects in virtual environments,’’ in Proc. IEEE Virtual Reality Conf., Reno, NE, USA, Mar. 2008, pp. 47–50.
[197] V. Tanriverdi and R. J. K. Jacob, ‘‘Interacting with eye movements in virtual environments,’’ in Proc. SIGCHI Conf. Hum. Factors Comput. Syst., New York, NY, USA, 2000, pp. 265–272.
[198] A. Plopski, Y. Itoh, C. Nitschke, K. Kiyokawa, G. Klinker, and H. Takemura, ‘‘Corneal-imaging calibration for optical see-through head- mounted displays,’’ IEEE Trans. Vis. Comput. Graphics, vol. 21, no. 4, pp. 481–490, Apr. 2015.
[199] J. Turner, A. Bulling, and H. Gellersen, ‘‘Extending the visual field of a head-mounted eye tracker for pervasive eye-based interaction,’’ in Proc. Symp. Eye Tracking Res. Appl., vol. 1. 2012, pp. 269–272.
[200] Y. Kim and S. Jo, ‘‘Wearable hybrid brain-computer interface for daily life application,’’ in Proc. 3rd Int. Winter Conf. Brain-Comput. Interface, Sabuk, South Korea, 2015, pp. 1–4.
[201] X. Fu, Y. Zang, and H. Liu, ‘‘A real-time video-based eye tracking approach for driver attention study,’’ Comput. Informat., vol. 31, no. 4, pp. 805–825, 2012.
[202] T. Poitschke, F. Laquai, S. Stamboliev, and G. Rigoll, ‘‘Gaze-based interaction on multiple displays in an automotive environment,’’ in Proc. IEEE Int. Conf. Syst., Man, Cybern., Oct. 2011, pp. 543–548.
[203] A. Doshi and M. Trivedi, ‘‘A comparative exploration of eye gaze and head motion cues for lane change intent prediction,’’ in Proc. IEEE Intell. Vehicles Symp., Eindhoven, the Netherland, Jun. 2008, pp. 49–54.
[204] M. Sodhi, B. Reimer, and I. Llamazares, Behavior Research Methods, Instruments, & Computers, vol. 34. New York, NY, USA: Springer- Verlag, 2002, p. 529.
[205] Y.-L. Chen et al., ‘‘Real-time eye detection and event identification for human-computer interactive control for driver assistance,’’ in Proc. IEEE Int. Conf. Syst., Man, Cybern., Oct. 2014, pp. 2144–2149.
[206] D. Kern, A. Mahr, S. Castronovo, A. Schmidt, and C. Müller, ‘‘Making use of drivers’ glances onto the screen for explicit gaze-based interac- tion,’’ in Proc. 2nd Int. Conf. Autom. User Interfaces Interact. Veh. Appl., 2010, p. 110.
[207] T. W. Victor, J. L. Harbluk, and J. A. Engström, ‘‘Sensitivity of eye- movement measures to in-vehicle task difficulty,’’ Transp. Res. F, Traffic Psychol. Behaviour, vol. 8, no. 2, pp. 167–190, Mar. 2005.
[208] Z. Wuhe, Z. Lei, and D. Ning, ‘‘Sensing driver awareness by combining fisheye camera and Kinect,’’ Proc.SPIE, vol. 9276, p. 927624, Nov. 2014.
[209] S. Han, S. Yang, J. Kim, and M. Gerla, ‘‘EyeGuardian: A framework of eye tracking and blink detection for mobile device users,’’ in Proc. 12th Workshop Mobile Comput. Syst. Appl., New York, NY, USA, 2012, Art. no. 6.
[210] Z. Li, G. Sun, F. Zhang, L. Jia, K. Zheng, and D. Zhao, ‘‘Smartphone- based fatigue detection system using progressive locating method,’’ IET Intell. Transp. Syst., vol. 10, no. 3, pp. 148–156, 2016.
[211] R. Biedert, A. Dengel, G. Buscher, and A. Vartan, ‘‘Reading and estimat- ing gaze on smart phones,’’ in Proc. Symp. Eye Tracking Res. Appl., New York, NY, USA, 2012, pp. 385–388.
[212] S. Cheng, ‘‘The research framework of eye-tracking based mobile device usability evaluation,’’ in Proc. 1st Int. Workshop Pervasive Eye Tracking Mobile Eye-Based Interact., New York, NY, USA, 2011, pp. 21–26.
[213] A. Lopez-basterretxea and A. Mendez-zorrilla, ‘‘Eye/head tracking tech- nology to improve HCI with iPad applications,’’ Sensors, vol. 15, no. 2, pp. 2244–2264, 2015.
[214] J. V. Singh and G. Prasad, ‘‘Enhancing an eye-tracker based human- computer interface with multi-modal accessibility applied for text entry,’’ Int. J. Comput. Appl., vol. 130, no. 16, pp. 16–22, 2015.
ANURADHA KAR is currently pursuing the Ph.D. degree with the National University of Ireland Galway and the Center for Cognitive, Connected, and Computational Imaging. Her research inter- ests include human computer interaction and com- putational imaging. She is involved in eye gaze tracking-addressing the issues of accuracy and per- formance evaluation of gaze estimation systems in various platforms.
PETER CORCORAN (F’10) is currently a Pro- fessor and a Personal Chair with the College of Engineering and Informatics, NUI Galway. In addition to his academic career, he is also an Occasional Entrepreneur, an Industry Consultant, and a Compulsive Inventor. He has co-authored over 300 technical publications and co-inventor on more than 250 granted U.S. patents. His research interests include biometrics, cryptography, com- putational imaging, and consumer electronics.
He is the Editor-in-Chief and the Founding Editor of the IEEE Consumer Electronics Magazine.
VOLUME 5, 2017 16519
