FREE ELECTRONIC LIBRARY - Dissertations, online materials

Pages:     | 1 |   ...   | 7 | 8 || 10 | 11 |   ...   | 21 |


-- [ Page 9 ] --

To further establish the difficulty in detecting the eye features, especially the iris boundaries and the eye corners, the performance of several standard edge and corner detectors and other similar algorithms is presented below. The aforementioned advantage of infrared illumination of the separation between the pupil and iris is also a disadvantage in detecting the iris boundaries. This is because in some subjects, the iris appears as a very light shade of grey and it can hardly be distinguished from the sclera (see Figure 8 for such examples). This is because the iris usually darkens on the border to the sclera but this dark border is less evident (thinner) in some people.

Further, because the camera is placed close to the eye, the inner and outer corners appear to have a different shape. This appears to be due to the appearance of the tear gland which is part of the inner eye corner. Thus, the inner eye corner extends further than the outer eye corner, as measured from the eyeball centre.

The small white circle that is the direct reflection of the infrared LED on the cornea and appears as red in 4 the heat-mapped images.

–  –  –

 Sobel edge operator (Gonzalez and Woods, 2001), with a 3x3 kernel.

 Canny edge detector (Canny, 1986) with two different set of parameters.

 Laplace operator (Gonzalez and Woods, 2001), with filter aperture of 3.

 Laplacian of Gaussian (LoG) filter (Gonzalez and Woods, 2001) – 5x5 Gaussian kernel, sigma set to 1.4 and Laplace filter aperture of 3. Both the absolute value of the filter output and the zero-crossings are illustrated.

 Minimum and maximum Eigen-values of the 2x2 gradient (or covariance) matrix extracted on a 3x3 window for each pixel. These Eigen values can be a reliable source of corners or other features and are used in other popular vision algorithms such as the Harris operator (Harris and Stephens, 1988) and the KLT feature tracker (Shi and Tomasi, 1994).

 Harris edge and corner detector (Harris and Stephens, 1988).

 Variance map – computed as the variance of a 5x5 window for each pixel.

 Partial x- and y-derivatives, with and without local maxima suppression.

As can be seen in Figure 9, most complex corner detectors perform poorly both for the detection of the iris boundaries as well as the detection of the eye corners. In fact, in most cases the iris boundaries are not preserved as strong edges in the output whereas the filtered output near the eye corners is heavily obstructed by edges from shadows on the sclera and eyelashes. In the next section, where the feature detection is explained in depth, the partial x- and y- derivatives were chosen over the more complex detectors because they were found to give similar output at a fraction of the time.

The original image has been smoothed using a Gaussian 5x5 kernel before being processed with each 5 algorithm.

–  –  –






–  –  –




= 3)




–  –  –





–  –  –

The REACT eye-tracker has been designed to detect a set of eye features (pupil location, iris radius and eye corners location) that can potentially be used to set up various different eyetracking projects. While other configurations can be easily implemented, the default configuration of the REACT eye-tracker uses these features to determine the two-dimensional direction of gaze, in degrees. A high-level block diagram of the feature detection and 2D gaze calculation is shown in Figure 10. A classification scheme (e.g. up and to the left) is proposed and implemented for the case study in Chapter 6. The eye-tracker processes images off-line although this is an implementation detail and not a requirement.

Taking advantage of the dark pupil effect discussed earlier, estimating the location of the pupil can be achieved using a simple thresholding technique and ellipse fitting. Global thresholding (versus adaptive) is used and therefore further refinement of the pupil contour is performed using an active contour (also known as a “snake”; Blake and Isard, 1998).

Then, using the concept of grey level edge strength used in cell segmentation tasks (Zhou and Pycock, 1997) the iris radius is detected and quantified. The eye feature detection is completed by initiating a search for the eye corners at the iris boundaries; edge information is obtained by performing local-maxima suppression of the y-derivative of the input image.

Both the iris and eye corner detection require that the subject is looking approximately straight ahead. This requirement is because of previously-mentioned reasons: a) as the eye moves, the iris is easily obscured by the eyelids/eyelashes and will also change shape b) as the eye moves, the eyelids near the eye corners will change shape and reveal or obscure more of the eyeball, thus severely altering the appearance of the corners.

Detecting the iris serves both to localize the eye corners position and to calculate the 3D gaze when a calibrated camera is available, like in the system presented by Wang et al. (2005).

Further, the eye corners serve as reference points and are used to calculate the reference-axes for the calculation of the gaze angle as well as re-calibration of the eye-tracker over long sequences when the glasses may have changed position on the subject’s head.

–  –  –

In the following sub-sections, the eye feature detection and direction of gaze calculation algorithms are extensively discussed.


The first step in detecting the pupil is thresholding the input image to a binary image

with a threshold value such that:

{ Example output of thresholding is demonstrated in Figure 11.

Using connected components labelling (Gonzalez and Woods, 2002), the binary image can be converted to a higher-level description of the image content. In brief, the labelling algorithm works by scanning the image top-to-bottom, left-to-right and labelling each pixel and those neighbour pixels that are connected to it with the same label, either with a 4- or 8-way connectivity criterion (see Figure 12). A second image scan is done in order to adjust equivalent labels. The resulting image contains the same intensity value (label) for each set of connected pixels (blob).

As an extension to the standard connected components labelling algorithm, during the labelling process, the following information is collected about each blob: a) the bounding rectangle, defined by the top-, bottom-, left- and right-most pixels, b) the area, defined by the number of pixels that have this blob’s label and c) a list of the locations of all the pixels that compose this blob. This information is used to assist consequent connected component processing; for example, a minimum and maximum blob area is defined ( and respectively) and used to filter blobs resulting from random noise (too small) and blobs that are formed in the darker areas of the image (where the infrared illumination fades rapidly and the resulting blobs occupy a very large area).

–  –  –

Mathematically put, from the original set of blobs, a new set is formed by discarding any elements where is now the set of pupil candidates. If the set is empty, the default threshold value is adjusted to,, and and the thresholding and labelling processes are applied again. This adjustment is necessary because the threshold is not chosen adaptively and in some cases may include too much or too little of the pupil.

At this stage, it is possible to take advantage of the pupil morphology. If we were to point a camera lens at a subject who is looking straight ahead, the pupil would appear approximately circular. The exact eyeball morphology is fairly complex and such approximations are necessary to design reduced complexity systems (Villanueva and Cabeza, 2007). In the case of the REACT

eye-tracker, the pupil always takes an approximately elliptical shape for two reasons:

 the camera is pointed to the eyeball from below and is off-centre  the video is de-interlaced by splitting the odd and even field into separate frames, thus vertically “cutting” the image in half (see Figure 6)

–  –  –


After performing an initial estimation of the pupil contour from the thresholded image, it is necessary to further refine it; since global thresholding is used, the results are very much dependent on the threshold value used. Whilst the pupil is always dark in the images, just how dark it is will depend on the camera topology (distance from the eyeball and angle of positioning i.e. exactly how the infrared light falls on the pupil), its pupil reflection properties and how they vary between subjects.

Explained simply, a strict threshold value, that is too low, may mark less or even no pixels from the pupil region as foreground. On the other hand, a permissive hypothetical threshold, that is too high, may mark more pixels than those inside the pupil region as foreground.

On the other hand, a more balanced threshold performs much better as shown in the comparison of Figure 15. As mentioned in the above section, different values of T are tried if a match for the pupil is not found; only in this sense the thresholding algorithm is adaptive.

–  –  –



An alternative route to global thresholding would have been to use local or adaptive thresholding (Gonzalez and Woods, 2002). Local thresholding is where the thresholding function applied depends on some local property of each pixel. Adaptive thresholding is where the thresholding function applied also depends on the spatial coordinates of each pixel. For example, one possible scheme of adaptive thresholding would be to examine small or medium-size neighbourhoods around each pixel and set the threshold to the mean value, median value or the average of the minimum and maximum value. While these operations are simple compared to other, more complex, schemes (e.g. Chow and Kaneko, 1972 cited by Gonzalez and Woods, 2002), they still increase the computational complexity 6 thus making them less suitable for real-time applications. At 59.97 frames a second, even small additions can make a significant difference in processing speed. The results from the next chapter (evaluation) support that this global thresholding scheme performs sufficiently well for this application despite its simplicity.

Now that the need for contour refinement has been established, more detail can be added as to how this is done in the REACT eye-tracker through the use of active contours.

For example, a scheme that examines the mean value of a neighborhood, would require 6 more additions and one division per pixel.

70 An active contour or “snake” is an energy minimizing spline that deforms to fit local minima (Blake and Isard, 1998); since active contours match local minima, their initial position must be explicitly defined. Snakes are versatile and can be adapted to wrap around various types of objects; thus, they are widely used in computer vision problems. Ramadan et al. (2002) used a snake to perform pupil tracking; that is, the task of tracking the pupil was performed solely with a modified snake (new pressure model and curvature formulation). In our case, a standard snake is used only to refine the contour obtained from thresholding.

The energy function that the snakes minimize is:

∫ ( ) ( ) is a parametric representation of the snake’s position, is the internal energy of the spline due to stretching and bending, is a measure of the attraction of image features such as contours and is a measure of the external constraints imposed either from higher-level shape information or user applied energy. For further details on the theoretical grounding of snakes and how these energies may be generically derived, the interested reader is referred to Blake and Isard (1998).

An adjustable, generic form of the above snake can be modelled as (Williams and Shah, 1992):

∫( ) In this form of the snake function, the first and second terms are first- and second-order continuity constraints and correspond to in the original snake function. The third and last term of the snake function is equivalent to and can measure some image quantity such as edge strength or image intensity. Finally, the weights are used to control the relative influence of each term to the snake energy – thus they usually default to and are adjusted depending on the particular application.

The greedy snake algorithm implemented in the REACT eye-tracker is derived from the work by Williams and Shah (1992) and is as follows.

71 The snake is initialized with the set of outmost points of the selected blob in the previous step,. When considering new locations for a point of the snake, a 7x7 neighbourhood is considered. Since a very good initial approximation of the spline is given from the thresholding, is computed such that the snake will not shrink but rather will favour points which

have distance near the average distance between consecutive points:

–  –  –

Pages:     | 1 |   ...   | 7 | 8 || 10 | 11 |   ...   | 21 |

Similar works:

«COGNITIVE SCIENCE 8, 1-26 (1984) Discrimination Nets as Psychological Models* LAWRENCE W. BARSALOU Emory University GORDON H. B O W E R Stanford University Simulations of human cognitive processes often employ discrimination nets to model the access of permanent memory. We consider two types of discrimination netsEPAM and positive-property-only netsand argue that they have insufficient psychological validity. Their deficiencies arise from negative properties, insufficient sensitivity to the...»

«REWARD AND PUNISHMENT PROCESSING IN SUBGROUPS OF YOUTH WITH CONDUCT PROBLEMS: CHARACTERIZATION OF THE NEURAL RESPONSE AND IMPLICATIONS FOR INTERVENTION by Amy Lynn Byrd, Ph.D. B.S. in Psychology, College of Charleston, 2006 M.S. in Clinical Psychology, University of Pittsburgh, 2010 Submitted to the Graduate Faculty of The Kenneth P. Dietrich School of Arts and Sciences in partial fulfillment of the requirements for the degree of Doctor of Philosophy University of Pittsburgh 2015 UNIVERSITY OF...»

«General-purpose localization of textured • • Image regions Rutb Rosenboltz· XeroxPARC 3333 Coyote Hill Rd. Palo Alto, CA 94304 Abstract We suggest a working definition of texture: Texture is stuff that is more compactly represented by its statistics than by specifying the configuration of its parts. This definition suggests that to fmd texture we look for outliers to the local statistics, and label as texture the regions with no outliers. We present a method, based upon this idea, for...»

«The Power of Being Thankful The Power of Being Thankful 365 Devotions for Discovering the Strength of Gratitude JOYCE MEYER N E W YO R K BOSTON NA SH V ILLE Copyright © 2014 by Joyce Meyer All rights reserved. In accordance with the U.S. Copyright Act of 1976, the scanning, uploading, and electronic sharing of any part of this book without the permission of the publisher constitute unlawful piracy and theft of the author’s intellectual property. If you would like to use material from the...»

«January 2010 Wartime Rescue of Jews by the Polish Catholic Clergy The Testimony of Survivors Edited and compiled by Mark Paul Polish Educational Foundation in North America Toronto 2009 Rescue activities on behalf of Jews were carried out by priests, nuns and monks in more than one thousand Roman Catholic Church institutions throughout Poland during World War II. The number of priests and religious involved in these activities was many times higher. This effort is all the more remarkable since...»

«3rd Edition, 2009 Introduction to the 3rd Edition nd It has been just over a decade since the 2 Edition of the Mush with P.R.I.D.E. Sled Dog Care Guidelines were published. During that time, scientists have made great strides in their understanding of dog physiology, psychology and behavior. Researchers have studied working sled dogs, with the support of their mushers, some of whom are also Mush with P.R.I.D.E. members. Many of these research projects have validated sled dog care methods that...»

«ecipe policy brief — 02/2016 POLICY BRIEF No. 2/2016 Xi Jinping’s long road to somewhere? China’s OBOR initiative and how Europe should respond by Guy de Jonquières, Senior Fellow at ECIPE The mention of China’s One Belt One Road initiative, also known as Belt and Road, brings to mind the Indian fable of the three blind men and the elephant. The men have never seen an elephant, so have no idea what one looks like. One grasps its trunk and insists it is a tree. Another puts his arms...»

«VARIATION OF FEEDING REGIMES: EFFECTS ON GIANT PANDA (AILUROPODA MELANOLEUCA) BEHAVIOR A Thesis Presented to The Academic Faculty By Estelle A. Sandhaus In Partial Fulfillment Of the Requirements for the Degree Master of Science in Psychology Georgia Institute of Technology September, 2004 Variation of Feeding Regimes: Effects on Giant Panda (Ailuropoda melanoleuca) Behavior Approved by: Dr. Terry L. Maple, Advisor School of Psychology Georgia Institute of Technology Dr. Mollie A. Bloomsmith...»

«From Soulard’s Notebooks Assistant Editor: Kassandra Soulard Poetry by Ric Amante 1 Notes from New England [Commentary] by Raymond Soulard, Jr. 9 Poetry by Judih Haggai 27 Poetry by Joe Ciccone 32 Paelo Redemption [Travel Journal] by Charlie Beyer 35 Poetry by Joe Coleman 45 Baby Steps Up into the Sky [Travel Journal] by Nathan D. Horowitz 55 Many Musics [Poetry] by Raymond Soulard, Jr. 59 Psychedelics and Lucid Dreaming: Doorways in the Mind [Essay] by A.S. Kay 91 Poetry by Tom Sheehan 99...»

«Seeing Your First Child with PANDAS/PANS Margo Thienemann, MD and The PANDAS Physicians Network Diagnostics and Therapeutics Committee Seeing Your First Child with PANDAS/PANS I. What Is PANS? The term Pediatric Acute-onset Neuropsychiatric Syndrome (PANS) describes the clinical presentation of a subset of pediatric onset Obsessive-Compulsive Disorder (OCD). PANS may also be a subset of Avoidant/Restrictive Food Intake Disorder (ARFID). While the criteria for diagnosing PANS do not specify a...»

«Public Broadcasters, the Internet, and Democracy Comparing Policy and Exploring Public Service Media Online Hallvard Moe Dissertation for the degree philosophiae doctor (PhD) University of Bergen Public Broadcasters, the Internet, and Democracy Comparing Policy and Exploring Public Service Media Online Bergen, 2008. Copyright: Hallvard Moe (Part I, and Part II, Article 1); Sage Publications (Part II, Article 2, 4, and 5); Intellect Books (Part II, Article 3). Public Broadcasters, the Internet,...»

«CANCER SUPPORT SERVICES in Ireland DIRECTORY Leinster Irish Cancer Society DIRECTORY OF SERVICES Welcome Welcome to the Fourth Edition of the Directory of Cancer Support Services Affiliated to the Irish Cancer Society. More people are getting cancer in Ireland. It is estimated that one in three of us will be diagnosed with the disease during our lifetime. Cancer is however increasingly viewed as a condition from which people survive. Evidence shows that when people experiencing cancer receive...»

<<  HOME   |    CONTACTS
2016 www.dissertation.xlibx.info - Dissertations, online materials

Materials of this site are available for review, all rights belong to their respective owners.
If you do not agree with the fact that your material is placed on this site, please, email us, we will within 1-2 business days delete him.