1. Facial Recognition at Purdue Universitys Airport 2003-2008
Jeremy M. Morton, C. Michael Portell, Stephen J. Elliott, Ph.D.
& Eric P. Kukula Presented By: Eric Kukula Biometric Standards,
Performance, and Assurance Laboratory www.biotown.purdue.edu
Department of Industrial Technology, School of Technology, Purdue
University, West Lafayette, IN 47906 Purdue University 2006 1
2. Identifying Individuals There are three common ways of
distinguishing someones identity Through something that an
individual knows A password Something an individual has
Identification card Something they own Biometric In many airport
applications, individuals gain access to specific areas by
providing a card and personal identification number (PIN). A
combination of these identifiers provides a more robust security
option. Biometric identification is defined as: The automatic
identification or identity verification of (living) individuals
based on behavioral and physiological characteristics Purdue
University 2006 2
3. Introduction Physiological biometrics include: Facial
recognition Finger Face Eye Hand Behavioral biometrics include:
Speaker (voice) Keystroke Dynamic signature verification.
Furthermore, a biometric must be: Measurable Robust Distinctive
Within the research community, there is interest in the performance
of a biometric over an extended period of time. According to Wayman
template aging is defined as: the increase in error rates caused by
time related changes in the biometric pattern, and that longer time
intervals generally make for more difficulty in matching samples to
templates Furthermore, a study of template aging will also require
a cohort of participants who are available over an extended period
of time. In this study, the time period will last up to five years.
Purdue University 2006 3
4. Motivation Cross Disciplinary Research Opportunity with
Aviation Technology The existence of an airport facility on campus,
with a potential cohort of individuals to participate over a five
year program Flight students Faculty Staff Interdisciplinary
relationships with other departments Need to understand how images
change over a period of time Purdue University 2006 4
5. Experimental Setup Evaluation takes place in Hanger 6, the
Student Flight Operations Center at the airport Evaluation of the
area completed in conjunction with a graduate class in Biometric
Technology Constraints: Had to fit in with the existing operation
of the airport Had to be as unobtrusive as possible Had to be able
to collect data ongoing for a 5 year period, therefore setup of
camera was critical In line with the labs philosophy Therefore
commercially available software had to be used, with all images
stored Purdue University 2006 5
6. Experimental Setup Purdue University 2006 6
7. Experimental Setup The face recognition system was setup in
the Hanger 6 area with no additional environmental controls. A
Logitech QuickCam Pro 4000 was used. This camera has a video
resolution of 640x480 VGA CCD, with up to 30 frames per second. The
camera also has a still image capture of up to 1280 x 960 pixels,
1.3 mega pixels. The video camera was connected to Dell Omniplex
GX260 2.0 GHZ, 512 MB RAM computers through a 35ft cable with USB
boosters. The camera was 6 inches off the table. The angle of the
camera also accommodated all of the participants, regardless of
their height. As lighting was constant, it was not measured
continually. Purdue University 2006 7
8. Volunteer Crew Recruitment of the students occurs at the
flight operations safety briefings Held once per month. It is
anticipated that over 300 students will be enrolled over the
duration of the test. Currently in Cohorts 1 and 2 Purdue
University 2006 8
9. Cohorts Cohorts are used to describe the groupings of
students Currently 2 cohorts So far there are 2 cohort groups Group
1 71 individuals 60 males 11 females Enrolled in a 4 day period in
May 2003 Group 2 56 individuals 54 male 2 female Purdue University
2006 9
10. Enrollment Each enrollment required an individual to stand
in front of the camera, and move their head to the left, right, up
and down. 100 images were taken for each enrollment, an enrollment
taking approximately 1.5 minutes The experimenters enrolled people
in batches, so later enrollees were more habituated to the
enrollment process than those at the beginning of the enrollment
period. It was noticed that sometimes, in a batch enrollment,
people would speed up their movement, causing the FRS system to
loose track of the individual This was the only significant issue
with enrollment There was 0% Failure to Enroll (FTE) There was also
a 0% Failure to Acquire (defined as successfully acquiring 100
images). Purdue University 2006 10
11. Initial Results from the Study As the data is collected in
a 1:M classification mode, ie: there is no claim to identity, each
image collected by the system has to be manually verified against
its original template The system returns two different variables
Classify Classify Failure Purdue University 2006 11
12. Initial Results Classify The classification rate in a 1:M
scenario is currently between 70-73% We are now seeing a decline in
the classification success rate and an increase in the
classification failure rate due to the following reasons: It is a
1:M system, so the user is not presenting their token or password
to be verified as a 1:1 match. Therefore, when an individual walks
by the camera, the system takes an image, and compares it to all
known individuals in the database However, there are lessons to be
learned regarding the testing and evaluation of biometric equipment
in operational tests Purdue University 2006 12
13. Issues with 1:M operational testing The result of a
classification failure in a 1:M situation may be an understatement
of the true performance of the FRS system due to: A False Match A
Failure to Acquire An unknown person Purdue University 2006 13
14. Issues with Operational Testing The false non-match rate
used in this evaluation is defined as the error rate of the
matching algorithm from a single attempt- template comparison in a
genuine attempt. This rate will be established by the research team
using additional software currently being developed Purdue
University 2006 14
15. Issues with Operational Testing Failure to Acquire The
failure to acquire rate is the proportion of attempts for which the
system is unable to capture or locate the users face In this
situation, as the camera is in a high traffic area, the camera
cannot find a face, therefore times out The failure to acquire rate
is counted as a classification failure Purdue University 2006
15
16. Issues with Operational Testing Unknowns An unknown is also
a classification failure. An unknown is defined for this study as
someone who has not yet been entered into the system The FRS
captures an image, tries to match that image with all those
enrolled in the database, yet fails to find one It then returns
classification failure Purdue University 2006 16
17. Issues and Conclusions with Operational Testing and the
Reporting of Results The issue with operational testing is that the
commercially available software (COTS) used in this study
overstates the classification failure rate due to the issues
discussed: False Match Failure to Acquire Unknowns Research is
currently underway at the Biometrics Standards, Performance, and
Assurance Laboratory to resolve these issues, and provide a
comprehensive best practice testing and reporting guide for 1:M
applications Purdue University 2006 17