Implementation of Efficient Pan-Tilt-Zoom Camera

Calibration
by Nicholas Fung and Philip David

ARL-TR-4799

Approved for public release; distribution unlimited.

April 2009

NOTICES
Disclaimers
The findings in this report are not to be construed as an official Department of the Army position
unless so designated by other authorized documents.
Citation of manufacturers or trade names does not constitute an official endorsement or
approval of the use thereof.
Destroy this report when it is no longer needed. Do not return it to the originator.

Army Research Laboratory

Adelphi, MD 20783-1197

ARL-TR-4799

April 2009

Implementation of Efficient Pan-Tilt-Zoom Camera

Calibration
Nicholas Fung and Philip David
Computational and Information Sciences Directorate, ARL

Approved for public release; distribution unlimited.

Form Approved
OMB No. 0704-0188

REPORT DOCUMENTATION PAGE

Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and maintaining the
data needed, and completing and reviewing the collection information. Send comments regarding this burden estimate or any other aspect of this collection of information, including suggestions for reducing the
burden, to Department of Defense, Washington Headquarters Services, Directorate for Information Operations and Reports (0704-0188), 1215 Jefferson Davis Highway, Suite 1204, Arlington, VA 22202-4302.
Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to any penalty for failing to comply with a collection of information if it does not display a currently
valid OMB control number.

PLEASE DO NOT RETURN YOUR FORM TO THE ABOVE ADDRESS.

1. REPORT DATE (DD-MM-YYYY)

2. REPORT TYPE

3. DATES COVERED (From - To)

April 2009

Final

October 2007 to August 2008

4. TITLE AND SUBTITLE

5a. CONTRACT NUMBER

Implementation of Efficient Pan-Tilt-Zoom Camera Calibration

5b. GRANT NUMBER
5c. PROGRAM ELEMENT NUMBER
6. AUTHOR(S)

5d. PROJECT NUMBER

Nicholas Fung and Philip David

5e. TASK NUMBER

5f. WORK UNIT NUMBER

7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES)

8. PERFORMING ORGANIZATION
REPORT NUMBER

U.S. Army Research Laboratory

ATTN: AMSRD-ARL-CI-IA
2800 Powder Mill Road
Adelphi, MD 20783-1197

ARL-TR-4799

9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES)

10. SPONSOR/MONITOR'S ACRONYM(S)

11. SPONSOR/MONITOR'S REPORT
NUMBER(S)

12. DISTRIBUTION/AVAILABILITY STATEMENT

Approved for public release; distribution unlimited.

13. SUPPLEMENTARY NOTES
14. ABSTRACT

Pan-tilt-zoom (PTZ) cameras, frequently used in both online and automated surveillance applications, require accurate
knowledge of camera parameters in order to accurately register autonomously tracked objects to a world model. Due to
imprecision in the PTZ mechanism, these parameters cannot be obtained from PTZ control commands but must be calculated
directly from camera imagery. This report describes the efforts to implement a real-time calibration system for a stationary
PTZ camera. The approach continuously tracks distinctive image feature points from frame to frame, and from these
correspondences, robustly calculates the homography transformation between frames. Camera internal parameters are then
calculated from these homographies. Finally, the external parameters can be calculated from the internal parameters and
image homographies. The calculations are performed by a self-contained program that continually monitors images collected
by the camera as it performs pan, tilt, and zoom operations. The accuracy of the calculated calibration parameters are
compared to ground truth data. The program works with a higher degree of accuracy for small changes in the cameras
external parameters. In addition, long algorithm execution time prevents the algorithm from running under all real-time
conditions.
15. SUBJECT TERMS

Camera calibration, PTZ camera, image homography, camera parameters

17. LIMITATION
OF
ABSTRACT

16. SECURITY CLASSIFICATION OF:

a. REPORT

Unclassified

b. ABSTRACT

Unclassified

c. THIS PAGE

Unclassified

UU

18. NUMBER
OF
PAGES

22

19a. NAME OF RESPONSIBLE PERSON

Nicholas Fung
19b. TELEPHONE NUMBER (Include area code)

(301) 394-3101
Standard Form 298 (Rev. 8/98)
Prescribed by ANSI Std. Z39.18

ii

Contents

List of Figures

iv

1.

Introduction

2.

Self Calibration

3.

Approach

3.1

Rooftop Camera System..................................................................................................3

3.2

Calculations .....................................................................................................................5

4.

Experiments

6.

Conclusions

12

7.

References

13

List of Symbols, Abbreviations, and Acronyms

14

Distribution List

15

iii

List of Figures
Figure 1. Testing of homography calculation algorithms under corresponding point errors. ........4
Figure 2. Focal length as calculated through the calibration routine..............................................7
Figure 3. Distortion parameter changes. .........................................................................................8
Figure 4. Aspect ratio and focal length changes over a changing zoom.........................................9
Figure 5. Principal point shift over a changing zoom...................................................................10
Figure 6. Example of a mosaic created with calibration values. ..................................................10
Figure 7. Derived pan/tilt values against camera read values.......................................................11

iv

1. Introduction
As technology has advanced, the potential for security and surveillance systems has increased.
Better communications technology allows for large amounts of data to stream quickly throughout
a network. Faster processors allow the use of advanced algorithms to process the data.
Potentially, large numbers of both stationary and mobile cameras can be used cooperatively to
form a persistent surveillance system to provide visual coverage of a large area. Algorithms to
detect, track, and place targets in a real-world model would make an impressive system.
However, the practical use of such a system is very dependent upon the ability to acquire
calibrated images from the camera with a high degree of accuracy. Inaccurate knowledge of
camera parameters can lead to missed or misidentified targets, in addition to errors in locating
the targets within a world model. As such, camera calibration is an important subject.
Proper camera calibration is important in establishing accurate images and camera positioning to
perform other visual surveillance tasks, such as target tracking and image mosaicking. The
calibration matrix consists of five values: the aspect ratio, the skew value, the focal length, and
the x and y values of the principal point. These parameters are needed to establish the position of
targets within a world model. They are also needed when manipulating several camera images,
such as when creating a mosaic of the observed area. In addition to determining the cameras
internal parameters, calculations can determine a rotation matrix, allowing for pan and tilt values
to be established for the camera. The camera image can also suffer from both radial and
tangential distortion, particularly at extreme zoom values. If the distortions can be properly
corrected, it can lead to more accurate target location and image mapping.
As part of the U.S. Army Research Laboratorys (ARL) goal of integrating various mobile
and stationary sensor assets into a comprehensive persistent surveillance system, proper camera
calibration is vital. This report documents steps taken toward achieving a real-time calibration
routine with a high degree of accuracy. While camera calibration parameters can be computed
offline using a series of images, real-time calibration will allow a higher degree of accuracy as
the routine can continually update the calibration parameters. This capability is important,
because the calibration can change over time and movement, particularly when performing
zoom functions.

2. Self Calibration
The calibration matrix is a 3x3 matrix that incorporates the cameras intrinsic calibration
parameters. The calibration matrix of frame k is
k f k
K k 0
0

sk
fk
0

xk
y k ,
1

(1)

where k is the aspect ratio, fk is the focal length, sk is the axis skew, and xk and yk are the x and y
values of the principal point, respectively. The aspect ratio is a value that relates the horizontal
pixel length to the vertical pixel length. For a camera with equal pixel width and height, the
aspect ratio is one, as is the case in many modern cameras. The focal length is directly related to
the zoom of the camera. A larger focal length indicates a higher zoom. The skew value
indicates a misalignment between the camera axes, resulting in a slanted image. The principal
point is the point in the image through which the optical axis passes. Other factors that influence
the captured images involve radial and tangential distortions. These distortions act upon the
image to alter how it is captured and are not represented in the calibration matrix. Radial
distortion consists of barrel distortions, which are large with wide angle or low zoom lenses, and
pincushion distortions, which are large for telephoto or large zoom lenses. Tangential distortions
occur under imperfect centering of lens components. Such distortions can be corrected for using
an appropriate camera model (4).
There are several methods for calibrating a camera. For a camera capable of pan, tilt, and zoom
(PTZ) operations, multiple images can be captured at different orientations. The images can be
analyzed to calculate the inter-image homography between each pair using corresponding point
pairs. The homography is a matrix that relates corresponding points from one image to another.
In the context of a PTZ camera, the homography is a matrix that allows one image to be aligned
to the next and form a mosaic. The homography can be analyzed to establish camera calibration
parameters between each pair of images.
Several algorithms can be used to determine the point correspondences between two images.
Two of the most popular methods are the Kanade-Lucas-Tomasi (KLT) Feature detector and the
Scale-Invariant Feature Transform (SIFT). For this project, we chose SIFT for the initial
approach to the problem, because the algorithm is invariant to translation, scaling, and rotation,
in addition to being partially invariant to illumination changes and affine projections (1). The
robustness of the algorithm would be appropriate for handling changes captured in the cameras.
For example, the algorithm would be robust against an illumination change resulting from a
cloud passing in front of the sun. After matching points between the current and previous image,
the homography can be computed. After some analysis, we employed a direct linear transform

(DLT) algorithm with normalization to perform this calculation. Other options were a nonlinear
algorithm and a DLT without normalization. We evaluated these approaches by creating ideal
corresponding points using known pan, tilt, zoom, and principal point settings. The
correspondences were then perturbed by a set amount and the resulting pan, tilt, zoom, and
principal point values were calculated. The results can be seen graphically in figure 1. The
graphs compare the results between the unnormalized and normalized linear approaches. For
each of these algorithms, three lines are plotted: the minimum error, the maximum error, and the
median error. Each of the parameters showed a sharp increase as the error in point
correspondences increased from 0 to 0.1 pixels. The errors then increased at a slower rate with
higher point correspondence errors. The normalized linear method performed with the lowest
overall error. This result is particularly evident with the maximum error. The nonlinear
algorithm is not depicted in the figure as it was determined to be too slow to be useable in a
real-time system.

3. Approach
3.1

Rooftop Camera System

The calibration routine was created for and tested on the rooftop camera system currently in
place at the ARL Adelphi, MD, building. The system includes four Sony Network Cameras
(SNCs) (Sony #SNC-RZ30) positioned at each corner of the building. One of the features of the
SNC is the ability to feed PTZ values to modify the orientation and zoom of the camera. In
addition, these values can be read back. While these features would seem to make calibrating the
extrinsic parameters unnecessary, the program should still perform this task so that other cameras
can be calibrated.
The cameras are all connected to an internal network and imagery can be streamed to or captured
by any computer on the network. We are in the process of creating programs that process the
data and perform functions such as target tracking and image mosaicking. Among such
programs, a procedure to perform camera calibration is important to establish accurate data.

Error in Pan Angle

10

10

-2

-2

10

10

-4

Error (degrees)

Error (degrees)

-4

10

-6

10

-8

10

10

-6

10

-8

10

-10

-10

10

10

-12

-12

10

10

Unnormalized Linear
Normalized Linear

-14

10

Error in Tilt Angle

10

10

0.1

0.2

0.3
0.4
0.5
0.6
0.7
0.8
Error in point correspondences (pixels)

0.9

10

Error in X-Coordinate of Principle Point

Unnormalized Linear
Normalized Linear

-14

0.1

0.2

0.9

Error in Y-Coordinate of Principle Point

10

0.3
0.4
0.5
0.6
0.7
0.8
Error in point correspondences (pixels)

10

10

-2

10

10

-2

10

Error (pixels)

-6

10

-6

10

-8

10

-8

10

-10

10
-10

10

-12

-12

10

10

Unnormalized Linear
Normalized Linear
0

0.1

0.2

0.3
0.4
0.5
0.6
0.7
0.8
Error in point correspondences (pixels)

0.9

Unnormalized Linear
Normalized Linear

-14

10

0.1

0.2

0.3
0.4
0.5
0.6
0.7
0.8
Error in point correspondences (pixels)

Error in Zoom

10

-5

10

Error

Error (pixels)

-4

10

-4

10

-10

10

Unnormalized Linear
Normalized Linear

-15

10

0.1

0.2

0.3
0.4
0.5
0.6
0.7
0.8
Error in point correspondences (pixels)

0.9

Figure 1. Testing of homography calculation algorithms under

corresponding point errors.

0.9

3.2

Calculations

To establish baseline calibration parameters, we used a MATLAB code created by Bouguet of

the California Institute of Technology (5). The toolbox implements a derivative of the method
developed by Heikkil and Silven at the University of Oulu in Finland (3). This method uses the
DLT to solve for the projection of an object onto the image, thus implicitly obtaining the
calibration parameters. The algorithm requires that a test pattern, resembling a checkerboard, be
captured by the camera in a series of images. The operator aligns the pattern to the image by
manually locating the corners of the grid within the image. Because the test pattern is of
known measurements, the projection can be calculated. Solving for the projection yields the
calibration parameters.
Once an initial calibration matrix has been solved, further calibrating can be performed more
easily on successive frames without the time-consuming process of capturing the test pattern.
First, we calculate the initial image of the absolute conic (IAC), 0, based on the initial
calibration matrix:

0 K 0 K 0T

(2)

We calculate the next IAC using the following formula. This equation is an application of the
Kruppa equation 6:

i H i T i 1 H i 1 .

(3)

We derive the new calibration matrix, Ki, using the Cholesky Decomposition of the IAC. The
Cholesky Decomposition returns two matrices, one lower and one upper triangular matrix, whose
product is the input:

K i chol (i ) .

(4)

Finally, we compute the rotation matrix, Ri, using the following formula:
1

Ri K i H i K i 1 .

(5)

From the rotation matrix, we can calculate pan and tilt values:
R11
R R21
R31

R`12
R22
R32

R13
R23
R33

R23

tilt tan 1
R33

pan sin 1 R13

R12
roll tan 1
R11

(6)

These values record the change of the camera orientation in that direction. We can calculate the
absolute zoom using the focal length values of the calibration matrix:
zoom

K i (1,1) K i (2,2)
.
2 K 1 (1,1)

(7)

With the new IAC, we can readily repeat the process. Each frame just requires the few simple
calculations, which are relatively quick to calculate. In addition, changes in the pan, tilt, and
zoom values can be accumulated to record the orientation of the camera.
To increase the accuracy of the procedure, we have to account for and correct the distortion
values. Barrel distortions are more prevalent at low zoom settings, and pincushion distortions
are more prevalent at high zoom settings. In the case of the Sony cameras used for testing, the
barrel distortions seemed noticeable at the lowest zooms, while the pincushion distortions were
not as noticeable at high zoom settings. In addition to the parameters of the calibration matrix,
the MATLAB code created by Bouguet also calculates five distortion parameters, ci, i = 1,,5.
The lens-distorted pixel location of a normalized (undistorted, true pinhole camera) image point
x = (x,y)T is xd = (xd,yd)T, where
x
x d d (1 c1r 2 c 2 r 4 c5 r 6 )x d x ,
yd

(8)

2c xy c 4 (r 2 2 x 2 )
dx 3 2
.
2
c 4 (r 2 y ) 2c 4 xy

(9)

The term dx represents the tangential distortion correction. The variable r represents the distance
of the point x from the center of the image. The homography can be calculated from the feature
points found in these undistorted images. By correcting for these distortions, the homography
will be more accurate and thus subsequent calculations to find the calibration parameters will
be more accurate.

4. Experiments
We encountered several difficulties during the creation of the camera calibration routine. One of
the first problems was in finding and implementing a viable algorithm both in processing time
and accuracy. The first approach toward achieving calibration involved applying a number of
assumptions to simplify the problem. We assumed the skew value to be zero, the aspect ratio to
be one, and the principal point to be centered. This approach was detailed in reference 2. We
made these assumptions, because modern, high-quality cameras tend to have attributes close to
these values. However, after some testing with this algorithm, we found the accuracy broke
down quickly, particularly under zoom changes and moderate angle changes. We determined
these inaccuracies were unacceptable, so we reworked the process to avoid making a large
number of assumptions. Further testing showed that the calibration parameters changed and
could not be assumed to be a specific value. In particular, the principal point changed greatly
under extreme zoom functions. The principal point shift can be seen graphically in figure 2.
Calculated Focal Length Over Changing Zoom
4000
3500
Focal Length

3000
2500
fc(1)

2000

fc(2)

1500
1000
500
0
0

10

20

30

40

50

60

70

Zoom Setting Change

Figure 2. Focal length as calculated through the calibration routine.

Another problem we encountered was with the current implementation of the camera calibration
algorithm. The program obtained good results under moderate to small changes in camera
orientation and zoom values. However, the accuracy broke down as the changes got larger. This
problem could have arisen due to inaccuracies in the homography that was used to calculate the
calibration parameters. To work against this problem, the calibration calculations can be taken
quickly and often to calibrate under small PTZ conditions.
The largest problem we encountered was in calculating the homography. An accurate
homography is vital to obtaining accurate calibration parameters. For this reason, we employed
the SIFT method. SIFTs ability to perform under scale changes is important for calibrating
cameras that will be performing zoom operations. However, the implementation of SIFT used

for the calibration routine proved to be time consuming. To combat this problem, we calculated
fewer points, sacrificing accuracy for expediency. However, this method still did not produce
the speed necessary for a real-time system. Other alternatives not yet implemented and tested
could include using a faster processor or a dedicated graphical processing unit (GPU) (7).
Because the distortion values are difficult to calculate on the fly, we calculated the values
beforehand using the Bouguet MATLAB code. The values were stored in a database to be used
as the camera underwent zoom functions. The captured images were corrected for distortion
before being processed to calculate the homography and calibration values. Figure 3 shows the
changes in the distortion parameters graphically. The camera returns a value between 0 and 100
to indicate the amount of camera zoom. Figure 4 provides an interpretation of this parameter.
The parameters stay relatively small for zooms under 50. However, they begin to vary widely as
the zoom is further increased. The fifth distortion parameter was zero under all measured zooms
and is not shown.
We computed the camera focal lengths at each step-of-10 increase in zoom using the Bouguet
MATLAB code. Figure 4 shows those results. We also calculated the aspect ratio. The results
show that the aspect ratio stays close to one throughout the various zoom settings. The focal
length changes at a rate that resembles exponential growth.
Distortion Value 1 Over Zoom Change

Distortion Value 2 Over Zoom Change

100
80
60

1.5

40
20
kc(2)

kc(1)

0.5

0
-20 0

20

40

60

80

100

80

100

-40
-60

-80
0

20

40

60

80

100

-100

-0.5

-120
Zoom Setting

Zoom Setting

Distortion Value 3 Over Zoom Change

Distortion Value 4 Over Zoom Change

0.04

0.035
0.03

0.03

0.025
0.02

0.02
kc(4)

kc(3)

0.015
0.01

0.01
0.005

0
0

20

40

60

80

100

-0.01

0
-0.005 0

20

40

60

-0.01
-0.02

-0.015
Zoom Setting

Zoom Setting

Figure 3. Distortion parameter changes.

Aspect Ratio Over Zoom Change

Focal Length Over Changing Zoom

1.1

12000
10000
Focal Length

Aspect Ratio

1.05

8000
fc(2)

6000

fc(1)

4000

0.95
2000
0.9

0
0

20

40

60

80

100

Zoom Setting

20

40

60

80

100

Zoom Setting

Figure 4. Aspect ratio and focal length changes over a changing zoom.

We then used the calibration routine to calculate the focal length according to a series of images
with a changing zoom. We calibrated each zoom image from the initial image of zoom 0. For
example, the image of zoom 40 was calibrated using the homography relating it to the image of
zoom 0 and the result was approximately 2000. The total results appear graphically in figure 3.
The calculated values follow the results from figure 4 closely through small zoom changes.
However, at a zoom change of 60, the vertical focal length began to diverge from the horizontal
focal length. We did not calculate zoom changes of over 70, because there were not enough
corresponding points to calculate a homography. Again, the zoom values used are built in zoom
settings of the camera from a 0 to 100 scale.
We also recorded the shift of the principal point over changes in zoom. The results are shown in
figure 5. As with the distortion parameters, the principal point is steady under low zooms, but
starts to shift greatly with higher zoom settings. Under ideal conditions, the principal point
would be the exact center of the image. In the case of this camera capturing a 640x480
resolution image, we would estimate the principal point to be 320 horizontal and 240 vertical. At
low zoom settings, the calibration routine returns numbers similar to these expected values. At
higher zoom settings, the principal point shifts to a higher degree. It is noteworthy that the
horizontal and vertical coordinates follow a similar shift pattern as the zoom setting increases.

Principal Point Shift Over Zoom Change

700

Principal Point

600
500
400

Horizontal

300

Vertical

200
100
0
0

20

40

60

80

100

Zoom Setting

Figure 5. Principal point shift over a changing zoom.

The final test was applying the algorithm to an application. Figure 6 shows an image mosaic
created with the calibrated data. We created the mosaic using 25 images at different pan and
tilt values. We held the zoom steady at the widest field of view. The figure shows good
alignment between the images as they are placed into the mosaic, indicating good calibration at
this zoom setting.

Figure 6. Example of a mosaic created with calibration values.

10

We created the mosaic in figure 6 using images taken from pan values of 25 to 45 and tilt
values of 10 to 30, each in 5 increments. Figure 7 graphically depicts the change in pan and tilt
values returned by the calibration algorithm against the values returned by the camera itself.
The top graph shows pan values while holding the tilt at five different constant values. The
bottom graph shows tilt values at each of five different pan values. The points would be
expected to follow a linear, x = y equation. However, the growth is steeper, particularly at high
pan and tilt values.
Camera Read Pan vs. Derived Pan
-35
Derived Pan (Degrees)

-30
-25

Tilt 0

-20

Tilt 5
Tilt 10

-15

Tilt 15
Tilt 20

-10
-5
0
0

-5

-10

-15

-20

-25

Camera Read Pan (Degrees)

Camera Read Tilt vs. Derived Tilt

Derived Tilt (Degrees)

60
50
0 Pan

40

5 Pan
30

10 Pan
15 Pan

20

20 Pan

10
0
0

10

15

Camera Read Tilt (Degrees)

Figure 7. Derived pan/tilt values against camera read values.

11

20

25

This result suggests that the algorithm breaks down as the extrinsic value changes increase in
magnitude. In addition, tilting the camera did not influence the pan values as much as panning
influenced the tilt values. The first graph shows that the pan values matched well, even at
different tilt values. The second graph shows that a higher pan created a higher tilt value. There
are several possible explanations for this phenomenon. One possibility is that the algorithm
breaks down with larger pan angles. Another possibility is an inconsistent camera pan/tilt
mechanism. Further study is needed to test this phenomenon.

6. Conclusions
The camera calibration routine is accurate for small PTZ operations. However, as the change in
the camera orientation grows larger, the accuracy breaks down. This conclusion is particularly
evident for calculated tilt angles under a large change in the pan angle. The accuracy breakdown
could be the result of inaccurate homographies, inaccuracies of the equations, or imprecise
movement of the camera PTZ mechanism. Further work needs to be done to improve the
performance, including the possibility of using a different feature tracker. Alternatively,
calibration could be performed at a frequent rate to avoid having to calibrate over a large pan,
tilt, or zoom operation. This method should improve the performance, because the data has
shown better accuracy over small PTZ operations. Unless the camera is changing at a fast rate,
frequent calibrations should have a small chance of calibrating over a large orientation change.
A limiting characteristic of the procedure is the speed of the SIFT algorithm. SIFT proved to be
accurate and has the positive characteristic of being robust to scale changes; however, the
implementation used required a large amount of processing time. As a result, the program
cannot run under normal real-time scenarios. Improvements can be made in terms of processor
speed or by assigning a designated processor, such as a GPU, to perform the task. Alternatively,
we are exploring other implementations of SIFT and other methods of obtaining a homography.

12

7. References
1. Lowe, D. Object Recognition from Local Scale-Invariant Features. Proceedings of the 7th
IEEE International Conference on Computer Vision, Kerkyra, Greece, Sept 1999, 1150
1157, vol. 2.
2. Kim, J.; Hong, K. S. A Practical Self-Calibration Method of Rotating and Zooming
Cameras. IEEE 2000, 354357.
3. Heikkila, J.; Silven, O. A Four-step Camera Calibration Procedure with Implicit Image
Correction. Proceedings from the 1997 IEEE Computer Society Conference on Computer
Vision and Pattern Recognition, San Juan, Puerto Rico, 1719 June 1997, 11061112.
4. Brown, D. C. Close-Range Camera Calibration. Photogrammetric Engineering 1971, 37
(8), 855866.
5. Bouguet, J.-Y. Camera Calibration Toolbox for Matlab.
http://www.vision.caltech.edu/bouguetj/calib_doc/index.html, 2 June 2008.
6. Hartley, R.; Zisserman, A. Multiple View Geometry in Computer Vision, 2nd ed.; Cambridge
University Press: Cambridge, UK, 2004.
7. Heymann, S.; Mller, K.; Smolic, A.; Frhlich, B.; Wiegand, T. SIFT Implementation and
Optimization for General-purpose GPU. 15th International Conference in Central Europe
on Computer Graphics, Visualization and Computer Vision, Plzen, Czech Republic,
29 Jan1 Feb 2007, 317322.

13

List of Symbols, Abbreviations, and Acronyms

ARL

U.S. Army Research Laboratory

DLT

direct linear transform

GPU

graphical processing unit

IAC

image of the absolute conic

KLT

Kanade-Lucas-Tomasi Feature

PTZ

pan, tilt, and zoom

SIFT

Scale-Invariant Feature Transform

SNCs

Sony Network Cameras

14

No. of
Copies

No. of
Copies

Organization

1
ELEC

ADMNSTR
DEFNS TECHL INFO CTR
ATTN DTIC OCP
8725 JOHN J KINGMAN RD STE 0944
FT BELVOIR VA 22060-6218

DARPA
ATTN IXO S WELBY
3701 N FAIRFAX DR
ARLINGTON VA 22203-1714

1 CD

OFC OF THE SECY OF DEFNS

ATTN ODDRE (R&AT)
THE PENTAGON
WASHINGTON DC 20301-3080

US ARMY RSRCH DEV AND ENGRG

CMND
ARMAMENT RSRCH DEV AND
ENGRG CTR
ARMAMENT ENGRG AND
TECHNLGY CTR
ATTN AMSRD AAR AEF T J MATTS
BLDG 305
ABERDEEN PROVING GROUND MD
21005-5001

PM TIMS, PROFILER (MMS-P)

AN/TMQ-52
ATTN B GRIFFIES
BUILDING 563
FT MONMOUTH NJ 07703

US ARMY INFO SYS ENGRG CMND

ATTN AMSEL IE TD F JENIA
FT HUACHUCA AZ 85613-5300

COMMANDER
US ARMY RDECOM
ATTN AMSRD AMR
W C MCCORKLE
5400 FOWLER RD
REDSTONE ARSENAL AL 35898-5000

US GOVERNMENT PRINT OFF

DEPOSITORY RECEIVING SECTION
ATTN MAIL STOP IDAD J TATE
732 NORTH CAPITOL ST NW
WASHINGTON DC 20402

Organization

US ARMY RSRCH LAB

ATTN AMSRD ARL CI OK TP
TECHL LIB T LANDFRIED
BLDG 4600
ABERDEEN PROVING GROUND MD
21005-5066

DIRECTOR
US ARMY RSRCH LAB
ATTN AMSRD ARL RO EV
W D BACH
PO BOX 12211
RESEARCH TRIANGLE PARK NC
27709

US ARMY RSRCH LAB

ATTN AMSRD ARL CI IA N FUNG
ATTN AMSRD ARL CI IA P DAVID
ATTN AMSRD ARL CI IA S YOUNG
ATTN AMSRD ARL CI OK PE
TECHL PUB
ATTN AMSRD ARL CI OK TL
TECHL LIB
ATTN IMNE ALC HR
MAIL & RECORDS MGMT
ADELPHI MD 20783-1197

TOTAL: 16 (1 ELEC, 1 CD, 14 HCs)

15

INTENTIONALLY LEFT BLANK.

16