A Virtual Proctor with Biometric Authentication for

Download PDF
2 downloads 0 Views 431KB Size Report
Comment
Keywords: Distance education, Virtual proctor, Face detection, Facial recogni ... plate match, the student is authenticated and can continue the educational activity .... ing samples X1 is sent to classifier h1, and a large number of samples which make the ... The part of the code based on the cascade of OpenCV is listed in Fig.
A Virtual Proctor with Biometric Authentication for Facilitating Distance Education Zhou Zhang, El-Sayed Aziz, Sven Esche & Constantin Chassapis Department of Mechanical Engineering, Stevens Institute of Technology, Hoboken, New Jersey, USA Zhou Zhang [email protected] El-Sayed Aziz [email protected] Sven Esche (Corresponding Author) [email protected] Constantin Chassapis [email protected]

Abstract. The lack of efficient and reliable proctoring for tests, examinations and laboratory exercises is slowing down the adoption of distance education. At present, the most popular solution is to arrange for proctors to supervise the students through a surveillance camera system. This method exhibits two shortcomings. The cost for setting up the surveillance system is high and the proctoring process is laborious and tedious. In order to overcome these shortcomings, some proctoring software that identifies and monitors student behavior during educational activities has been developed. However, these software solutions exhibit certain limitations: (i) They impose more severe restrictions on the students than a human proctor would. The students have to sit upright and remain directly in front of their webcams at all times. (ii) The reliability of these software systems highly depends on the initial conditions under which the educational activity is started. For example, changes in the lighting conditions can cause erroneous results. In order to improve the usability and to overcome the shortcomings of the existing remote proctoring methods, a virtual proctor (VP) with biometric authentication and facial tracking functionality is proposed here. In this paper, a twostage approach (facial detection and facial recognition) for designing the VP is introduced. Then, an innovative method to crop out the face region from images based on facial detection is presented. After that, in order to render the usage of the VP more comfortable to the students, in addition to an eigenface-based facial recognition algorithm, a modified facial recognition method based on a realtime stereo matching algorithm is employed to track the students’ movements. Then, the VP identifies suspicious student behaviors that may represent cheating attempts. By employing a combination of eigenface-based facial recognition and real-time stereo matching, the students can move forward, backward, left,

adfa, p. 1, 2011. © Springer-Verlag Berlin Heidelberg 2011

right and can rotate their head in a larger range. In addition, the modified algorithm used here is reliable to changes of lighting, thus decreasing the possibility of false identification of suspicious behaviors. Keywords: Distance education, Virtual proctor, Face detection, Facial recognition, Stereo matching

1

Introduction

The distance education market keeps growing rapidly 1. While several research threads (i.e. on real-time creation of virtual environments for virtual laboratories 2, 3, augmentation of virtual laboratories 4, creation of smart sensor networks 5, etc.) have contributed to the continued adoption of distance education approaches, the lack of efficient and reliable proctoring is slowing this adoption process down. At present, the most popular solutions in distance education for monitoring an experiment or an examination are human proctors. Human proctors used in distance education can be teaching assistants, instructors, laboratory administrators and faculty members. There are also companies that provide the service of monitoring examinations from a distance (e.g. ProctorU 6). In most remote proctoring cases, the students take examinations and perform experiments on a computer and the proctor(s) watch(es) them from another computer through video cameras. The human proctors must monitor a screen throughout the entire process. The basic requirement for operating a remote proctor is that it needs a remote surveillance camera system mounted at the student’s site. The advantage of this method is that it is similar to traditional classroom education and therefore provides fewer challenges than technology-assisted methods such as VPs. However, this method also has two shortcomings. One disadvantage is that the operational costs are high since such proctoring services currently usually charge over $60 per student per course while the cost for setting up the surveillance system is high. Another disadvantage is that the proctoring process is laborious and tedious. With the further development of computer vision technology, VPs appeared. VPs are integrated software-hardware solutions that have the potential to contribute to bringing academic integrity to distance education. They were enabled by the proliferation of the high-speed Internet and advanced computer peripherals. They first perform the authentication of the students by scanning either their faces 7 or their fingerprints 8. Then, a camera monitors the environment and/or a microphone records the sounds within it. Virtual proctoring software used in distance education includes Remote Proctor Pro 9 , Instant-InTM Proctor 10 , Proctortrack 11 , Proctorfree 12 and Securexam Remote Proctor 13. Virtual proctoring has three advantages over human proctors. First is its low fixed cost. The students only need to set up a webcam and install the VP software which then performs the authentication of the student and the proctoring of the educational activity. Typically, the cost of the VP (including webcam, microphone and software kit) will not exceed $15 per student. The second advantage of virtual proctoring lies in its convenience. There is no need for human proctors, and thus the educational activity to be proctored can take place at anytime and anywhere. The third advantage is in the accurate authentication of the students. The utilization of biometric

technologies enables the accurate recognition of the students, thus ensuring a reliable authentication10. It should be noted that virtual proctoring is still evolving and current systems are often attracting complaints from the students, mostly because of two shortcomings exhibited by these systems. First, they impose more severe restrictions on the students than a human proctor would. The students have to sit upright and remain directly in front of their webcams at all times. Second, the reliability of the VP highly depends on the initial conditions under which the proctoring is started. For example, changes in the lighting conditions can cause mistakes in the verification of suspicious behaviors 14. In order to improve the usability and to overcome the shortcomings of the existing remote proctoring methods, a VP with biometric authentication and facial tracking functionality is proposed here. This VP is designed to authenticate the students and capture suspicious behaviors based on facial recognition and facial tracking. The workflow of this VP is depicted in Fig. 1 and is composed of two main parts: authentication and supervision. In addition, there is a database which stores the enrolled students’ face templates indexed with their campus ID. When they use this VP, the students are first required to scan their face using a webcam. Second, the scanned frame is processed, and the part of the frame containing the face is cropped out. Third, the face is then compared with the face template that was stored in the face database and could be retrieved by the index of student’s campus ID. If the face and the template match, the student is authenticated and can continue the educational activity. Otherwise, the student is logged out. After authentication, the frame used for authentication is stored in a newly allocated memory address and is taken as the new template. Then, the subsequent matching is based on this new template instead of the template stored in the face database. Then, the educational activities are monitored by the webcam. During the monitoring period, the VP samples the live video of the student with a sample rate of 30 frames per second. If the mismatching percentage between the sampled face and the new template exceeds a pre-configured threshold value, a suspicious behavior is identified and a video clip is recorded, which is then used for further verification by the instructor of the examinations or experiments.

2 2.1

Design of virtual proctor based on facial recognition Overview of proposed virtual proctor

The proposed VP was designed based on facial recognition techniques. In order to reduce the computational cost of the facial recognition while keeping its reliability, the process was divided into two stages. The first stage is the detection of the face in a sampled frame of the student’s live video. Once a face has been detected, the area of the detected face is cropped out from that frame and used for the following facial recognition. The second stage is the recognition of the detected face. The cropped face area is compared with the face template as illustrated in Fig. 1. Because the cropped face area is much smaller than the whole image area, the computational cost of recognizing the face is reduced considerably.

Below, the reason why facial recognition was selected instead of other biometric methods is explained first. Following that, various facial detection algorithms are discussed, including the algorithm selected here and the method employed to crop out the face area. Subsequently, a modified facial recognition algorithm based on stereo matching is introduced that overcomes some of the shortcomings of other facial recognition algorithms based on template matching or eigenfaces. Finally, the results of some benchmarks are presented to confirm that the proposed facial recognition algorithm is reliable. 2.2

Advantages of facial recognition for virtual proctor

A VP should have the functions of both authentication and real-time monitoring. In the authentication process, the method used to verify the students’ identity can be based on their biometric information (such as face snapshot10, fingerprint8, palm print 15, hand geometry 16, iris 17 and/or retina 18). Following the authentication, the student’ activities are monitored by consecutively sampling the webcam video. Compared with other biometrics-based methods, the facial recognition method employed here has three advantages: • •

•

The hardware is available and affordable. A common webcam, instead of special biometrics data readers, can meet the hardware requirements for real-time facial recognition and tracking. The algorithms used to implement facial recognition are much simpler than those used in biometrics-based methods, thus allowing for a higher sampling frequency. The features used for facial recognition are so notable that they can be identified very easily. Therefore, the algorithms for the facial recognition are more robust and simpler than other biometrics-based algorithms. Facial recognition is more practical than iris or retina tracking. Facial recognition is macroscopic in scale while iris recognition and retina scanning are based on microcosmic features which have strict requirements related to the distance between the scanners and the eyes.

Based on the above discussion, facial recognition was used to design the proposed VP.

Fig. 1. Workflow of virtual proctor based on facial recognition

2.3

Facial detection

2.3.1

Selection of facial recognition algorithm

Facial detection is the first step that precedes facial recognition. In fact, there are many facial detection algorithms with different complexities. The most common methods in facial detection include 19: •

• • •

Detecting faces in images with a controlled background. The most common approach in this method is to use the green screen algorithm 20 to crop out the faces. Although this method is the simplest one, it is not practical to employ it in VPs because one cannot expect the students to provide a green background. Detecting faces by color. This method uses a typical skin color to find face segments. Obviously, it is not robust when the environment lighting condition is changed. In addition, it is not universally effective for all kinds of skin colors. Finding faces by motion. This method assumes that the face is the only moving object in consecutively acquired images. Thus, it is not effective in scenarios where there are other moving objects in the background. Finding faces in unconstrained scenes. This method removes the constraints imposed on the background (for example an intended green background) or the face itself (for example the markers on faces). Hence, it represents a general and convenient method. In addition, it can be further divided into tracking based on models (e.g. model-based facial detection 21, edge-orientation matching based facial detection 22, Hausdorff distance facial detection 23) and weak classifier cascades (e.g. boosting classifier cascades 24 , asymmetric AdaBoost and detector cascades 25).

Obviously, unconstrained methods are more appropriate for VPs. Tracking based on models is not robust because of the lack of generalization in the definition of human facial expressions 26. On the other hand, facial detection using weak classifier cascades is based on the analysis of human expressions, and hence, it is more general and robust. The algorithms used here represent a modification of weak classifier cascades.

2.3.2

Innovative method to crop out face based on facial detection

A basic cascade is a degenerate decision tree. The training process is implemented by going through a sequence of weak classifiers expressed as functions (h1, h2, …, hn) with binary outputs (true = 1 and false = 0) as illustrated in Fig. 2 27. The set of training samples X1 is sent to classifier h1, and a large number of samples which make the output of h1 equal to zero are rejected. Subsequently, the remaining samples are sent to h2, and so on. After n stages, the number of samples is significantly smaller. The new classifier is composed of (h1, h2, …, hn). Then, the remaining samples can be taken as the input of other cascade processing or another detection system. After the training process described above, a series of strong and accurate classifiers was obtained. For facial detection, Haar features were used to train the classifier 28,29. In addition, the Adaboost (adaptive boosting) algorithm was employed to find the best threshold for the Haar features. For convenience reasons, the pre-trained classifier from OpenCV 30 was used to implement the facial detection. More details can be found elsewhere 31. The process described above only finds an estimated area of the faces. It renders the recognition process difficult since the information provided by the face area is insufficient for the subsequent facial recognition. In fact, the estimated area of the faces results in the loss of the entire background and part of the outline of the face. In order to compensate for the loss of information and to facilitate the following facial recognition process, a modification for the facial detection algorithms based on the localization of the mouth and eyes was implemented. First, the coordinates of the eyes are set as E1(x1, y1) and E2(x2, y2), and the coordinates of the mouth are set as M(x3, y3). All coordinates represent the center of the area of the eyes and mouth. The cross product of vectors E1 E2 and E1M is positive if M is above the line E1E2. Based on the golden ratio of the human face 32, the face area forms a golden rectangle with the eyes at its midpoint. Then, the vertical ratio equals the distance between the pupils and the mouth in relation to the distance of the hairline to the chin, i.e. E1 M / H C = 0.36. The horizontal ratio equals the distance between the pupils in relation to the width of the face, i.e. E1 E2 / L R = 0.46 (see Fig. 332, 33 ). The obtained rectangle should be L R × H C, but the rectangle actually used to facilitate the face recognition is increased by 10 pixels in each direction in order to avoid loss of the facial information. The part of the code based on the cascade of OpenCV is listed in Fig. 4.

Fig. 2. Schematic depiction of detection cascade

Fig. 3. Golden ratio of human face

Fig. 4. Part of code used to detect face, eyes and mouth

2.4

Facial recognition

2.4.1

Limitations of template matching and eigenfaces in facial recognition

Face recognition methods for the images can be divided into feature-based methods 34 and holistic methods 35. Feature-based methods lead to robust results, but they make the automatic detection of the features difficult to achieve. Obviously, these methods are inappropriate for the VPs. Holistic methods have the advantage that they concentrate on the limited regions or points of interest without the distortion of the information of the images 36. Their shortcoming is the hypothesis of equal importance of all pixels in the image. These methods are not only costly but also sensitive to the rela-

tionship between the training samples and the test data, to changes of the pose and to the illumination conditions. Typical holistic methods include temple match, eigenfaces, eigenfeatures, the combination of eigenfaces and eigenfeatures 37 and 2D matching. In the template matching algorithm, the selected patch that is taken as the template traverses the target image. Then, an error function is defined as:

R( x, y ) = f [T ( x' , y ' ), I ( x + x' , y + y ' )]

(1)

where R is the resulting error, T is the template, I is the target image, (x, y) are the coordinates of the image in pixels, and (x’, y’) are the coordinates of the template in pixels. Different error functions can be specified depending on the prevailing conditions. After comparison between the template and the target image, the best matches can be found as global minima or maxima 38. The eigenface method is an efficient approach for recognizing a face 39. A high recognition rate can be achieved with a low dimension d of the eigenvector space since the recognition rates are stable when the dimension of the eigenvector space equals 8. In order to identify the eigenvectors, a principal component analysis was used to find the directions with the greatest variances of the components of a given dataset. These variances are called principal components (and are also the eigenvalues associated with the eigenvectors used in the eigenfaces). Then, a high-dimensional dataset is described by such a series of correlated variables. The algorithm can be described as follows. Let

X = {x1 , x2 ,..., xn }, x i ∈ R d be a random vector wherein xi are the observa-

tions. The expected value μ of the observations is:

µ=

1 n

n

∑x

i

(2)

i =1

The covariance matrix S can be expressed as:

S=

1 n ∑ ( xi − µ )( xi − µ )T n i =1

(3)

The eigenvalues λi and eigenvectors νi of S are defined by:

Svi = λi vi ,

i = 1,2,..., n

(4)

If there are k principal components and the corresponding eigenvectors are labelled in descending order based on the values of the principal components, then the k principal components of the observed vector x are given by:

y = W T ( x − µ ),

W = {v1 , v2 ,..., vk }

(5)

The reconstruction of the eigenvectors from the principal component analysis (PCA) is given by: x = Wy + µ

(6)

Following the procedures outlined above, three more steps realize facial recognition. In the first step, all training samples are projected into the PCA subspace composed of eigenvectors. In the second step, the query image (i.e. the target image that will be identified) is projected into the PCA subspace. In the last step, the nearest neighbor between the projected training images (i.e. the former training samples) and the projected query image is determined. Template matching is reliable and simple when the contexts of the images are constrained. The eigenface method is designed based on a generalization of the faces, and therefore, it is robust and accurate under conditions that vary only mildly. Unfortunately, both the template matching method and the eigenface method are vulnerable to changes in the environment and pose variations of the face. Therefore, the method of facial recognition employed here is a modification of stereo matching based on the facial detection results. 2.4.2

Facial-detection-based facial recognition with stereo matching

The proposed method used to recognize and track faces is based on stereo matching. Stereo matching refers to comparing two images taken by nearby cameras and attempting to map every pixel in one image to the corresponding location in the other image. The proposed method can be described as follows: • • • • • • •

Detect the facial, eye and mouth area in the captured frame Crop out the facial area Mark the eyes, mouth as landmarks Rectify the captured frame based on the landmarks both in the template and the captured frame Compare the template stored in the face database to the rectified frame (i.e. the frame with the same row coordinates as the corresponding points in the template) using the stereo matching algorithm Compute stereo correspondence which illustrates the relationship of the points between the pair of images, and the matching cost which measures the similarity of pixels Identify the students by a pre-set threshold used to control the matching cost, and the value of this threshold is selected according to the desired accuracy

The core of these procedures is the stereo matching algorithm. The basic approach of the matching algorithm can be described as follows: The template image and the captured images are expressed in grayscale instead of in R, G, B values. Given the intensity I(x,y) of a point in the template and the intensity I’(x,y) of the assumed corresponding point in the captured image, the absolute intensity disparity d(x,y) of the two points can be computed as follows:

d ( x, y ) = I ( x, y ) − I ' ( x, y )

(7)

Then, the sum of the absolute intensity differences (SAD) of the intensity in the template and captured image is: W

W

∑ ∑ I ( x + i, y + j ) − I ' ( x + i + d , y + j )

SAD(d ) =

(8)

y = −W x = −W

If the SAD is used directly to obtain the disparity maps (which refer to the apparent pixel difference or motion between a pair of stereo images), the noise in the disparity maps is very large since the signal-to-noise ratio is too low. In order to optimize the stereo matching accuracy, a box filter based on the cross-correlation in the window areas (of size 2W × 2W) around the landmarks is used:

S C ( x, y , d ) =

W

W

∑ ∑ [ I ( x + i, y + j ) − I ' (x + i + d , y + j )]

(9)

j = −W i = −W W +1

S C ( x, y + 1, d ) =

W

∑ ∑ [ I ( x + i, y + 1 + j ) − I ' (x + i + d , y + 1 + j )]

(10)

j = −W −1 i = −W

Denoting the sum of a row in the windows as:

AC ( x, y, d , j ) =

W

∑ (( I ( x + i, y + 1 + j ) − I ' ( x + i, y + 1 + j − d )]

(11)

i = −W

Then, the cross-correlation in a 2W × 2W window centered at point (x,y) becomes:

S C ( x, y + 1, d ) = S C ( x, y, d ) + AC ( x, y, d ,W + 1) − AC ( x, y, d ,−W )

(12)

Here, d is the possible floating of the actual location of the captured image (i.e. the lag in the definition of the cross-correlation). In Fig. 5, the comparison of the disparity maps obtained with SAD and SAD with window filter is depicted. It can be seen that the noise can be reduced efficiently with the window filter. After obtaining the disparity between the template image and the captured image, the facial recognition of the users can be performed. In order to test the proposed method used to recognize the students, a benchmark analysis was conducted which compares the proposed method with the template matching and eigenface methods. In this benchmark, “the Sheffield (previously UMIST) face database” 40, 41 was used to test the performance of different methods with respect to the pose translation. 20 sample sets from this database were tested from different view points (0 to 90 degrees in 10 degree intervals). The frontal view was defined as the 0 degree viewpoint, and it was taken as the template. The test results are shown in Fig. 6. It is seen that the stereo matching method provides for a higher reliability than the other two methods when the pose translation is large.

In order to test the reliability of the proposed method under variable illumination conditions, a benchmark analysis with the “extended Yale face database B” 42,43 was conducted. The frontal faces of 15 individuals under 50 different illumination conditions were examined. The results of the test are illustrated in Fig. 7. The eigenface method is better than the template matching method, the reliability of which is critically affected by the illumination conditions. In contrast, the stereo matching method ensures a high reliability even under poor illumination conditions. As mentioned above, the final implementation of the stereo matching method focuses on three areas: two eye-centered areas and one mouth-centered area. Decreasing the sizes of the areas used for stereo matching does not only increase the efficiency of the stereo matching algorithm but also improves the reliability of the facial recognition. In addition, ‘C++ accelerated massive parallelism’ 44 was used in the implementation of the stereo matching algorithm, and therefore the speed of the execution of the program was increased significantly.

Fig. 5. Comparison of SAD and SAD with window filter

Fig. 6. Template matching, eigenfaces and stereo matching about pose translation

Fig. 7. Template matching, eigenfaces and stereo matching under various illumination conditions

3

Definition of suspicious behaviors and test results

It is difficult to define all possible suspicious behaviors that may represent cheating attempts. Therefore, a small set of such behaviors was defined in the pilot implementation of the proposed VP.

In order to understand this definition, a coordinate system was chosen as depicted in Fig. 8. First, rotations of the head about the Z axis are considered normal. Suspicious behaviors of “rotating head” correspond to rotations about either the X or Y axes. Second, “moving relative to webcam” corresponding to a translation along the X, Y or Z directions is also suspicious. The suspicious behavior of “rotating head” is judged by the matching percentage between the captured frame and the template stored in the face database. The essence of this method is that the face in frontal view is taken as the facial recognition object. Thus, rotations about the X and Y axes generate obvious differences between the captured frame and the template. In addition, these differences cannot be eliminated by face alignment. Therefore, the rotation angle can be estimated according to the face matching percentage. The calculation of translations is much easier than that of rotations. The method used here is to track the location of the face and the size of the face area. Then, the relative location of the face can be determined. In order to evaluate the performance of the VP, 50 intentional cheating attempts per criterion were tested. The test results corresponding to different criteria are listed in Table 1. These results show that the proposed VP is reliable with respect to pose translations and illumination changes. For further assessment of the proposed algorithms used in the VP, a pilot test with three volunteers was conducted. The details can be found elsewhere 45. The difference between this pilot test and the prior experiment involving 50 cheating attempts is that the illumination conditions were changed while the volunteers were asked to bow their head. Despite the changing conditions, the VP based on the proposed algorithms worked well with respect to recognizing and tracking the users.

Table 1. Test results corresponding to different criteria based on 50 cheating attempts

Head rotation and translation (rotation Rθ in deg, translation Td in cm)

∣Rθ∣ ≤ 10

∣Rθ∣ ≤ 30

∣Rθ∣ ≤ 50

∣Rθ∣ ≤ 60

∣Rθ∣ > 70

Accuracy

94%

≤ 87%

≤ 76%

≤ 54%

≤ 30%

∣Td∣ ≤ 20

∣Td∣ ≤ 40

∣Td∣ ≤ 60

∣Td∣ ≤ 70

any Td

Fig. 8. Definition of suspicious behaviors

4

Conclusions and future work

In this paper, a VP based on facial recognition was introduced. In order to overcome the shortcomings of existing facial recognition methods, a stereo matching method was proposed to improve the reliability of the facial recognition. In order to evaluate the reliability of this method, two benchmark analyses were conducted. The first analysis was designed to determine the impact of the stereo matching method on the reliability for large pose translations. The second analysis was aiming to test the performance of the stereo matching method under variable illumination conditions. The results proved that the proposed method has a higher reliability than the existing facial recognition methods (template matching and eigenfaces). In the future, the performance of the VP will be enhanced by adding voice identification and recognition functions, adding screen monitoring functionality, targeting more complicated suspicious behaviors and optimizing the recognition algorithms. Although the proposed VP still has certain limitations, it performed well under laboratory conditions. In addition, it has the potential to replace human proctors in both distance education and traditional classroom settings.

References

[1] http://wcet.wiche.edu/initiatives/research/WCET-Distance-EducationEnrollment-Report-2016, accessed in November, 2016. [2] Zhang, Z., Zhang, M., Chang, Y., Esche, S. K. & Chassapis, C., 2015, “A smart method for developing game-based virtual laboratories”, Proceedings of the

ASME International Mechanical Engineering Conference & Exposition IMECE'15, Houston, Texas, November 13-19, 2015. [3] Zhang, Z., Zhang, M., Chang, Y., Esche, S. K. & Chassapis, C., 2016, “Real-time 3D reconstruction for facilitating the development of game-based virtual laboratories”, Computers in Education Journal, Vol. 7, No. 1, pp. 85-99. [4] Zhang, Z., Zhang, M., Tumkor, S., Chang, Y., Esche, S. K. & Chassapis, C., 2013, “Integration of physical devices into game-based virtual reality”, International Journal of Online Engineering, Vol. 9, pp. 25-38. [5] Qureshi, F. & Terzopoulos, D., 2007, “Smart camera networks in virtual reality”, Proceedings of First ACM/IEEE International Conference on Distributed Smart Cameras, Vienna, Austria, September 25-28, 2007. [6] http://www.proctoru.com/, accessed in October, 2016. [7] http://www.proctortrack.com/, accessed in October, 2016. [8] http://www.softwaresecure.com/remote-proctor-pro-faq/, accessed in October, 2016. [9] http://proctorfree.com/, accessed in October, 2016. [10] http://www.biomids.com/proctoring/, accessed in October, 2016. [11] http://www.proctortrack.com/, accessed in October, 2016. [12] http://proctorfree.com/, accessed in October, 2016. [13] http://remoteproctor.com/rpinstall/orgselector/orgselector.aspx, accessed in October [14] http://www.nytimes.com/2015/04/06/technology/online-test-takers-feel-anticheating-softwares- uneasy-glare.html, accessed in September, 2016. [15] Rasmussen, K. B., Roeschlin, M., Martinovic, I. & Tsudik, G., 2014, “Authentication using pulse-response biometrics”, Proceedings of Network and Distributed System Security Symposium 2014, San Diego, California, USA, February 23-25, 2014. [16] Bača, M., Grd, P. & Fotak, T., 2012, “Basic principles and trends in hand geometry and hand shape biometrics”, In: “New Trends and Developments in Biometrics”, INTECH Open Access Publisher. [17] http://proctorfree.com/blog/the-future-is-now-iris-recognition-technology-makesid-cards-redundant, accessed in October, 2016. [18] Proctor, R. W., Lien, M. C., Salvendy, G. & Schultz, E. E., 2000, “A task analysis of usability in third-party authentication”, Information Security Bulletin, Vol. 5, No. 3, pp. 49-56. [19] https://facedetection.com/algorithms/, accessed in October, 2016.

[20] Horprasert, T., Harwood, D. & Davis, L.S., 2000, “A robust background subtraction and shadow detection”, Proceedings of 4th Asian Conference on Computer Vision, Taipei, Taiwan, January 5-8, 2000. [21] http://www.cs.rutgers.edu/~decarlo/facetrack.html, accessed in October, 2016. [22] Fröba, B. & Külbeck, C., 2001, “Real-time face detection using edge-orientation matching”, Proceedings of International Conference on Audio- and Video-based Biometric Person Authentication, Halmstad, Sweden, June 6-8, 2001. [23] Jesorsky, O., Kirchberg, K. J. & Frischholz, R. W., 2001, “Robust face detection using the Hausdorff distance”, Proceedings of International Conference on Audio- and Video-based Biometric Person Authentication, Halmstad, Sweden, June 6-8, 2001. [24] Vasconcelos, N. & Saberian, M. J., 2010, “Boosting classifier cascades”, Proceedings of Advances in Neural Information Processing Systems 23, Vancouver, British Columbia, Canada, December 6-9, 2010. [25] Viola, P. & Jones, M., 2001, “Fast and robust classification using asymmetric adaboost and a detector cascade”, Proceedings of Advances in Neural Information Processing System 14, Vancouver, British Columbia, Canada, December 3-8, 2001 [26] Gokturk, S. B., Bouguet, J. Y., Tomasi, C. & Girod, B., 2002, “Model-based face tracking for view-independent facial expression recognition” Proceedings of Fifth IEEE International Conference on Automatic Face and Gesture Recognition, Washington D.C., USA, May 20-21, 2002. [27] Viola, P. & Jones, M., 2004, “Robust real-time object detection”, International Journal of Computer Vision, Vol. 57, No. 2, pp. 137-154. [28] Wilson, P. I. & Fernandez, J., 2006, “Facial feature detection using Haar classifiers”, Journal of Computing Sciences in Colleges, Vol. 21, No. 4, pp. 127-133. [29] http://docs.opencv.org/master/d7/d8b/tutorial_py_face_detection.html, in November, 2016.

accessed

[30] http://opencv.org/, accessed in November, 2016. [31] Zhang, Z., Zhang, M., Chang, Y., Esche, S. K. & Chassapis, C., 2016, “A virtual laboratory system with biometric authentication and remote proctoring based on facial recognition”, Proceedings of the 2016 ASEE Annual Conference & Exposition, June 26-29, 2016, New Orleans, LA, USA. [32] http://www.goldennumber.net/face/, accessed in November, 2016. [33] http://www.goldennumber.net/facial-beauty-new-golden-ratio/, accessed in November, 2016

[34] Brunelli, R. & Poggio, T., 1993, “Face recognition: features versus templates”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 15, pp. 1042-1052. [35] Turk, M. & Pentland, A., 1991, “Eigenfaces for recognition”, Journal of Cognitive Neuroscience, Vol. 3, pp. 71-86. [36] Jafri, R. & Arabnia, H., 2009, “A survey of face recognition techniques”, Journal of Information Processing Systems, Vol 5, No. 2, pp. 41-68. [37] Pentland, A., Moghaddam, B. & Starner, T., 1994, “View-based and modular eigenspaces for face recognition”, Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, June 21-23, 1994, Seattle, WA. [38] http://docs.opencv.org/2.4/modules/imgproc/doc/object_detection.html, accessed in November, 2016. [39] Menezes, P., Barreto, J. C. & Dias, J., 2004, “Face tracking based on Haar-like features and eigenfaces”, Proceedings of IFAC/EURON Symposium on Intelligent Autonomous Vehicles, Técnico, Lisboa, Portugal, July 5-7, 2004. [40] https://www.sheffield.ac.uk/eee/research/iel/research/face, accessed in November, 2016. [41] Graham, D. B. & Allinson, N. M., 1998, “Characterising virtual eigensignatures for general purpose face recognition”, In: Face Recognition, Springer Berlin Heidelberg, pp. 446-456. [42] http://vision.ucsd.edu/~leekc/ExtYaleDatabase/ExtYaleB.html, accessed in November, 2016. [43] Georghiades, A. S., Belhumeur, P. N. & Kriegman, D. J., 2001, “From few to many: Illumination cone models for face recognition under variable lighting and pose”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 23, No. 6, pp. 643-660. [44] https://msdn.microsoft.com/en-us/library/hh265136.aspx, accessed in November, 2016. [45] Zhang, Z., Zhang, M., Chang, Y., Esche, S. K. & Chassapis, C., 2016, “A virtual laboratory combined with biometric authentication and 3D reconstruction”, Proceedings of the ASME International Mechanical Engineering Conference & Exposition IMECE’16, Phoenix, Arizona, USA, November 11-17, 2016.