INDOOR PTZ CAMERA CALIBRATION

WITH CONCURRENT PT AXES

Jordi Sanchez-Riera, Jordi Salvador and Josep R. Casas

Image Processing Group, UPC – Technical University of Catalonia

Jordi Girona 1-3, ediﬁci D5, 08034 Barcelona

Keywords:

Pan, Tilt, Zoom, Active camera calibration.

Abstract:

The introduction of active (pan-tilt-zoom or PTZ) cameras in Smart Rooms in addition to ﬁxed static cameras

allows to improve resolution in volumetric reconstruction, adding the capability to track smaller objects with

higher precision in actual 3D world coordinates. To accomplish this goal, precise camera calibration data

should be available for any pan, tilt, and zoom settings of each PTZ camera. The PTZ calibration method pro-

posed in this paper introduces a novel solution to the problem of computing extrinsic and intrinsic parameters

for active cameras. We ﬁrst determine the rotation center of the camera expressed under an arbitrary world

coordinate origin. Then, we obtain an equation relating any rotation of the camera with the movement of the

principal point to deﬁne extrinsic parameters for any value of pan and tilt. Once this position is determined,

we compute how intrinsic parameters change as a function of zoom. We validate our method by evaluating the

re-projection error and its stability for points inside and outside the calibration set.

1 INTRODUCTION

Smart Rooms equipped with multiple calibrated cam-

eras allow visual observation of the scene for appli-

cation to multi-modal interfaces in human-computer

interaction environments. Static cameras with wide-

angle lenses may be used for far-ﬁeld volumetric anal-

ysis of humans and objects in the room. Wide angle

lenses might provide maximum room coverage, but

the resolution of the images from these far-ﬁeld static

cameras is somewhat limited. The introduction of

pan-tilt-zoom (PTZ) cameras permits a more detailed

analysis of moving objects of interest by zooming in

on the desired positions. In this context, calibration

of the PTZ cameras for any pan, tilt, and zoom set-

tings is fundamental to referring the images provided

by these cameras to the working geometry of the ac-

tual 3D world, commonly employed by static cam-

eras.

Our research group has built a room equipped with six

wide-angle static cameras. The purpose of the static

cameras is the computation of a 3D volumetric re-

construction of the foreground elements in the room

by means of background subtraction techniques fol-

lowed by shape-from-silhouette. The volumetric data

resulting from this sensor fusion process is exploited

by different tracking and analysis algorithms working

in the actual 3D world coordinates of the room. With

wide angle lenses spatial room coverage from multi-

ple cameras is almost complete, but the precision of

the volumetric data is restricted by the resolution of

the camera images. Feature extraction analysis op-

erates at a coarse spatial scale, and cannot obtain de-

tailed information, such as the positions of ﬁngers and

hands, which might correspond to convex volumet-

ric blobs in the volumetric representation. The intro-

duction of PTZ cameras in the Smart Room helps to

provide a higher resolution volumetric reconstruction

for hand and gesture recognition algorithms. Hand

movements are closely followed by means of real-

time adjustment of pan and tilt. The calibration of

PTZ cameras is fundamental to referring their images

to the common 3D geometry of the computed volu-

metric data.

Camera calibration methods can be divided into two

groups. The ﬁrst group makes use of a calibration

pattern, such as a checkerboard, located at a known

3D world position. Calibration parameters are then

inferred from the detected pattern points in order to

calibrate either a single camera (Tsai, 1987; Zhang,

2000; Heikkila, 2000), or several cameras simultane-

ously (Svoboda et al., 2005). The disadvantage of

methods based on knowledge of the calibration pat-

tern position is that they might not be convenient for

Sanchez-Riera J., Salvador J. and R. Casas J. (2009).

INDOOR PTZ CAMERA CALIBRATION WITH CONCURRENT PT AXES.

In Proceedings of the Fourth International Conference on Computer Vision Theory and Applications, pages 10-15

DOI: 10.5220/0001754900100015

 SciTePress

large focal lengths.

The second group of methods, also known as auto-

calibration or self-calibration, may also use a calibra-

tion pattern, but in any case the 3D world position

is not known (Hartley and Zisserman, 2000; Agapito

et al., 1999). These methods are more difﬁcult to

implement due to geometrical complexity problems,

such as ﬁnding the absolute conic.

Most of the methods cited above have been developed

for static cameras. Calibration of active cameras faces

new challenges, such as computing variation of intrin-

sic and extrinsic parameters, respectively, as a func-

tion of the zoom value and the rotation (pan and tilt)

angles. Effective calibration of the varying parame-

ters for the PTZ cameras requires adaptation of the

algorithms developed for static cameras.

Intrinsics Calibration for PTZs. One approach for

zoom calibration considers each zoom position as a

static camera and then calibrates the PTZ as multi-

ple static cameras. (Willson, 1994) provides a study

on zoom calibration describing the changes in focal

length, principal point, and focus based on Tsai cal-

ibration (Tsai, 1987). The large number of possible

combinations for zoom and focus values makes the

computation of intrinsic parameters ineffective. Other

approaches limit the re-projection error as described

in (Chen et al., 2001), which proposes the calibration

of extreme settings followed by the comparison of the

re-projection error with an arbitrary threshold. Posi-

tions with errors above the threshold are re-calibrated

with a mid point and the extreme. The process is

repeated until the range of zoom and focus settings

is below the re-projection error threshold. A quar-

tic function is ﬁnally found to best match the ob-

tained parameters. However, the strategy of multi-

ple static cameras presents problems with increasing

zoom values, that reduce the constraints needed to

solve the equations making results inconsistent (Ruiz

et al., 2002). Huang (Huang et al., 2007) proposes an

algorithm for cameras with telephoto lenses to face

this problem.

An alternative approach for calibrating intrinsic pa-

rameters under varying zoom values is the separate

computation of focal length and principal point al-

though, for the latter, (Li and Lavest, 1996) claim that

calibration results are not signiﬁcantly affected if the

principal point is assumed to be constant.

Extrinsics Calibration for PTZs. Extrinsic param-

eters are important for the correct application of in-

trinsic calibration results, but the methods mentioned

so far focus on the computation of intrinsic parame-

ters without describing how to compute extrinsics in

much detail. (Sinha and Pollefeys, 2004) proposes

computing both intrinsics and extrinsics in an outdoor

environment using an homography between images

acquired at different pan, tilt and zoom. (Davis and

Chen, 2003) compute extrinsics by ﬁnding the cam-

era rotation axis. Further references for computing

both extrinsics and intrinsics for surveillance applica-

tions are (Kim et al., 2006; Senior et al., 2005; Lim

et al., 2003).

Proposal. To overcome the problems of the differ-

ent methods mentioned above for the calibration of

active cameras, we propose to ﬁnd the rotation center

of the camera. Once the rotation center is known, ex-

trinsic parameters can be found by simply applying a

geometric formula, and then intrinsic parameters can

be determined by a simple bundle adjustment.

In the following, we ﬁrst review the speciﬁcs of the

active camera model, then describe the proposed cal-

ibration method in Section 3, and ﬁnally we present

the experimental results in Section 4 and draw con-

clusions to close the paper.

2 THE PTZ CAMERA MODEL

Camera calibration is important to the relation of the

physical world with the images captured by the cam-

era. This relation is deﬁned by a mathematical model.

The pin-hole camera model is based on the perspec-

tive transform (a 3x3 matrix K) explaining the rela-

tion from the 3D world to the 2D image plane. Unfor-

tunately, real camera optics introduce distortion that

must be modeled and corrected. The camera model

is not complete without positioning the camera in the

physical world according to arbitrarily chosen world

coordinates. For this we need to know the rotation (a

3x3 matrix R) and translation (a 3x1 vector T) of the

camera coordinates with respect to the world coordi-

nates.

The equation resulting from the above description re-

lates the 3D homogeneous points X in world coor-

dinates with 2D homogeneous points x in the image

plane as follows:

x = K[R|T]X (1)

2.1 Extrinsic Parameters

For PTZ cameras, R and T depend on how the prin-

cipal point changes with pan and tilt movements. If

the principal point is different from the rotation cen-

ter of the camera (e.g. by a shift r in the optical axis),

then any change in the pan (or tilt) angle of α degrees

will produce a rotation and a translation of the prin-

cipal point, which, due to the mechanics of the PTZ

INDOOR PTZ CAMERA CALIBRATION WITH CONCURRENT PT AXES

camera, will describe a sphere of radius r centered

on the rotation center. Considering R

and T

, the ro-

tation and translation describing the position of the

rotation center as reference position, any pan (or tilt)

movement will produce an additional rotation R

and

translation T

describing the position of the principal

point.

The reference is deﬁned to be at pan and tilt zero,

then the rotation R

and translation T

should be ap-

plied to the 3D homogeneous points X. Note that this

transform is not a trivial operation. When pan and

tilt displacement occur at the same time, two rota-

tion matrices are needed and the order in which ma-

trices are multiplied produces different results. Thus,

the transform necessary to express the 3D homoge-

neous points X in correct principal point origin coor-

dinates must be applied in the following order: Rotate

the points R

and translate T

, then apply rotation R

and after the translation T

, ﬁnally translate the points

back to the origin with transform T.

= (R

) (2)

R = R

(3)

= T

+ T

+ [0,0,r]

(4)

T = T

+ T

(5)

In the next subsections we describe the parameters in

Equations (2-5) and the effects of pan and tilt inde-

pendently. Our coordinate axis election is described

considering z for the optical axis (in the depth direc-

tion) and x, y for the Cartesian coordinates on the im-

age plane.

2.1.1 Pan and Tilt Movements

Figure 1 represents the displacement of the princi-

pal point when either pan or tilt changes. When pan

changes by α degrees, it introduces a rotation R

(α)

and a translation T

(α) parameterized by α:

(α) =





cos(α) 0 sin(α)

0 1 0

−sin(α) 0 cos(α)





(6)

(α) =













rsin(α)

−r(1− cos(α))





(7)

When tilt changes by β degrees, it introduces a rota-

tion R

(β) and a translation T

(β) parameterized by β:

(β) =





1 0 0

0 cos(β) −sin(β)

0 sin(β) cos(β)





(8)

Figure 1: Principal point displacement (black dot) due to

pan (left image) and tilt (right image).

(β) =













rsin(β)

−r(1− cos(β))





(9)

2.2 Intrinsic Parameters

The perspective transform provides a 3x3 matrix, K,

mapping the 3D space to the 2D camera plane assum-

ing the pin-hole camera model. This mapping is char-

acterized by the principal point and the focal length.

The principal point (p

, p

) is located at the intersec-

tion of the optical axis with the image plane. Usually

the camera is designed so that the principal point is

at the center of the image, but it may shift depending

on the camera optics and zoom settings. The focal

length is the distance between the principal point and

the focal point. Measured in pixels, the focal length is

represented by two components ( f

, f

) (the “aspect

ratio” f

/ f

is different from 1 only for non-square

pixels). For PTZ cameras, we may assume that all the

intrinsic parameters are a function of zoom:

K(z) =





(z) sk(z) p

(z)

0 f

(z) p

(z)

0 0 1





(10)

sk represents the skew coefﬁcient deﬁning the angle

between the x and y image axes. Further, we must

consider the distortion introduced by camera lenses,

represented by four parameters, two for radial and two

for tangential distortion, as described in (Bouguet,

2007).

3 CALIBRATION METHOD

The ﬁrst goal of our calibration method is to ﬁnd the

rotation center of the camera and its distance to the

VISAPP 2009 - International Conference on Computer Vision Theory and Applications

principal point. Then, with just a geometric trans-

form, we can obtain the extrinsic parameters for any

given pan, tilt position. Once this is done, we can

focus on the intrinsic parameters. Using the extrinsic

paramters already found as the valid camera world po-

sition, we can obtain the intrinsic parameters consid-

ering that, when zoom changes, intrinsic parameters

also change.

To compute the rotation center and its distance to the

principal point, ﬁrst we place the calibration pattern

in a known and measured position in the room respect

our world origin, and we take some images at sev-

eral known pan and tilt positions at zoom zero (widest

lens). One of the images is taken as reference and the

others are described by its relative pan and tilt position

with respect to the reference. Then a bundle adjust-

ment, shown in Equation (11) is applied to minimize

the reprojection error.

∑

i=1

∑

j=1

||(x

− ˆx(R

,kc,a

)||

(11)

where ˆx(R

,..) is the projection of point X

in image

i according to Equation (1) with modiﬁcations ex-

plained in Section 2.1 and, using four parameters for

lens distortion kc. R

, T

are the rotation and transla-

tion matrices for the reference image. R

, R

and T

are the additional rotation and translation matrices

due to the relative pan and tilt movements, described

in Equations (6-9). Our formulation assumes there

are n images with m points in each image. Depend-

ing on the camera position, it will not be possible to

take images of the calibration pattern at pan and tilt

zero, thus a

and a

describe the pan and tilt of the

reference image respect the zero position.

In order to reach convergence, it is important to start

from a good initial guess of the parameters’ values. In

our case the initial guess for the intrinsic and extrin-

sic parameters is obtained with (Bouguet, 2007), al-

though any other method, such as (Tsai, 1987), could

be equally valid.

With the same principle, we can consider the values

of rotation center and radius just found after the previ-

ous error minimization as an initial guess for a second

bundle adjustment. In this case, we aim at minimizing

error with respect to all the parameters: Intrinsic and

extrinsic (rotation center and radius). This minimiza-

tion is computed on the functional shown in Equa-

tion (12).

∑

i=1

∑

j=1

||(x

− ˆx(K, R

,kc,a

)||

(12)

Note that the intrinsic parameters K, are not present

in the Equation (11). With a calibrated PT (pan-

tilt) camera after this second minimization, we need

to calibrate the intrinsic parameters as a function of

zoom. For this purpose, we take some images at dif-

ferent zoom values of the calibration pattern in the

previous reference position. As we increase the zoom,

the calibration pattern ends up outside of the image,

therefore, a smaller calibration pattern is used. This

second pattern must be aligned with the previous one

to keep it in the same measured position. In this case,

the minimized functional is shown in Equation (13).

∑

j=1

||(x

− ˆx(K(z),X

)||

(13)

Note that only one image is used each time and only

intrinsic parameters are minimized. In our particular

case, the camera has 1000 step positions for zoom.

This is a large number of values and it would not be a

good idea to perform the minimization for all the pos-

sible values. Therefore, we take the images sampling

the whole zoom range at equal intervals. With this, we

obtain the intrinsic parameters for some of the values

of zoom and then, we can ﬁt a function. This function

is as simple as a polynomial of n degrees and n+ 1

coefﬁcients.

4 RESULTS

The proposed algorithm has been implemented in

Matlab as an extension of the (Bouguet, 2007) tool-

box. In order to test the correctness and limits of

the method, we have studied how pan and tilt range

affects the re-projection error and have determined

which is the optimal number of degrees to ﬁt a func-

tion and the optimal sampling of the zoom range. An-

other test performed to ensure that the correct param-

eters are found consists in the evaluation of the repro-

jection error of a known 3D position different from

those used to calibrate the camera, in the Smart Room.

For all recordings we have used two different calibra-

tion patterns consisting of black and white squares.

The images have been grabbed using a Sony-EVI D70

camera, controlled with the Evi-Lib library (EVI-Lib,

2006) for adjustment of pan, tilt, and zoom.

Depending on the distance from the calibration pat-

tern to the camera, the range of pan and tilt, such that

the camera is able to grab an image of the whole pat-

tern, will vary. In the selected pattern position, we are

able to capture a maximum range of 35 degrees from

one side to another of pan, and a maximum range of

20 degrees from one side to another of tilt. In Figure 2

we have represented the re-projection error (y-axis)

as a function of pan range (x-axis) and tilt range (the

INDOOR PTZ CAMERA CALIBRATION WITH CONCURRENT PT AXES

lines drawn). Note that, if for any reason, we are not

able to work with a tilt range larger than 10, the pan

range needs to be at least 25 to keep re-projection er-

ror below 3 pixels. For a tilt range larger than 10, the

algorithm is more ﬂexible, allowing more variability

for pan range and still keeping a re-projection error

below 1.5 pixels.

5 10 15 20 25 30 35

0.5

1.5

2.5

pan range

reprojection error

tilt 20

tilt 15

tilt 10

Figure 2: Reprojection error for extrinsic parameters as a

function of pan and tilt.

In Figure 3 we have represented the reprojection error

(y-axis) as a function of the number of coefﬁcients

(x-axis) and intervals of zoom sampling (the lines

drawn). For intervals of 25 or 50 steps, the results

are very similar, except for coefﬁcients larger than 14,

when re-projection error of the line corresponding to

50 steps increases. This is because with 50 steps, over

a range of 1000 possible steps, there is not enough

data to ﬁt a function with more than 14 coefﬁcients.

For similar reasons, when data is taken from intervals

of 100 steps, it is not possible to ﬁt a function with an

arbitrary large number of coefﬁcients. With a small

number of coeffs. (e.g. 4), the error is larger than 5

pixels.

4 6 8 10 12 14 16 18

2.5

3.5

4.5

reprojection error

degrees

steps 25

steps 50

steps 100

Figure 3: Reprojection error for intrinsic parameters for dif-

ferent polynomial degrees.

With camera parameters within correct re-projection

error margins, it is also important to acknowledge to

what extent the model is able to reproduce the same

results for different positions than those used to cal-

ibrate the camera. For this purpose a second set of

images has been acquired with the calibration pat-

tern at different position. Figure 4 shows two exam-

ples of the re-projection error at different zoom val-

ues. Taking 40 different images at random positions

with zoom zero (shortest focal length), we get a re-

projection error of 1.20 pixels with a standard devia-

tion of 0.56 pixels while taking 40 images at different

zoom positions the re-projection error is 2.45 pixels

with a standard deviation of 2.64 pixels. This proves

that the re-projection error is very close to the values

obtained during calibration.

Figure 4: Grid projection. Left image zoom 0, error = 0.93

pixels. Right image zoom 550, error 1.83 pixels.

5 CONCLUSIONS

Calibration of active cameras has to overcome some

difﬁculties that are not present for ﬁxed cameras, such

as solving calibration equations for telescopic lenses

or large focal lengths and the fact that the rotation cen-

ter of the PTZ camera is not co-located with the prin-

cipal point. The goal of this work has been to obtain a

calibration algorithm for PTZ cameras that improves

the precision of volumetric reconstruction algorithms

applied to a combination of ﬁxed and PTZ camera im-

ages in our Smart Room. The principal features we

wanted were real-time availability of calibration pa-

rameters for all range of pan, tilt and zoom values,

with re-projection errors smaller than one pixel spac-

ing in order to have precise 3D volumetric reconstruc-

tion.

State of the art methods did not provide a solution for

these goals. Methods for static cameras use to fail

for large focal lengths and require large tables to store

calibration values for the pan-tilt-zoom range. Meth-

ods for active cameras were more adequate but pro-

vided a good solution for only a small range of pan

and tilt.

The presented algorithm is ﬂexible in the sense that it

provides readily available calibration parameters for

any combination of pan, tilt, and zoom, with a small

re-projection error, even though it is slightly larger

than one pixel.

Some sources of error have been identiﬁed. Probably,

the most inﬂuential error source is the assumption of

a nearly constant 3D position of the principal point of

VISAPP 2009 - International Conference on Computer Vision Theory and Applications

the camera with respect to the zoom when, in practice,

the focal length in the zoom range of the camera can

vary up to a few centimeters. The consideration of

constant lens distortion for the zoom range is not true

either, although its inﬂuence in the ﬁnal results might

be insigniﬁcant.

Another possible source of error is the position of the

calibration pattern in the Smart Room. The Smart

Room has some predeﬁned points, precisely mea-

sured to put the calibration pattern on according to

arbitrary coordinates. For each attempt to position

the calibration pattern at predeﬁned 3D world coordi-

nates, there is an error of about 0.5cm in the physical

location, which is small considering the actual dimen-

sions of the Smart Room(4m × 5m), but still enough

to affect camera calibration results.

ACKNOWLEDGEMENTS

We want to thank to Jeff Mulligan for the ideas given

in the algorithm deﬁnition and to Albert Gil for the

support given for camera grabbing software.

REFERENCES

Agapito, L. d., Hayman, E., and Reid, I. (1999). Self-

calibration of a rotating camera with varying intrinsic

parameters. In Proc. 9th British Machine Vision Con-

ference, Southampton., pages 105–114.

Bouguet, J. Y. (2007). Camera calibration tool-

box for matlab. (Available for download from

http://www.vision.caltech.edu/bouguetj/calib doc).

Chen, Y.-S., Shih, S.-W., Hung, Y. P., and Fuh, C. S.

(2001). Simple and efﬁcient method of calibrating a

motorized zoom lens. Image and Vision Computing,

19(14):1099–1110.

Davis, J. and Chen, X. (2003). Calibrating pan-tilt cameras

in wide-area surveillance networks. In Ninth IEEE

International Conference on Computer Vision, 2003.

Proceedings, number 1, pages 144–149.

EVI-Lib (2006). C++ library for controlling the serial in-

terface with sony color video cameras evi-d30, evi-

d70, evi-d100 (http://sourceforge.net/projects/evilib/).

http://sourceforge.net/projects/evilib/.

Hartley, R. I. and Zisserman, A. (2000). Multiple View Ge-

ometry in Computer Vision. Cambridge University

Press, ISBN: 0521623049.

Heikkila, J. (2000). Geometric camera calibration using

circular control points. IEEE Transactions on Pat-

tern Analysis and Machine Intelligence, 22(10):1066–

1077.

Huang, X., Gao, J., and Yang, R. (2007). Calibrating pan-

tilt cameras with telephoto lenses. In Computer Vision

- ACCV 2007, 8th Asian Conference on Computer Vi-

sion, pages I: 127–137, Tokyo, Japan.

Kim, N., Kim, I., and Kim, H. (2006). Video surveillance

using dynamic conﬁguration of multiple active cam-

eras. In IEEE International Conference on Image Pro-

cessing (ICIP06), pages 1761–1764.

Li, M. and Lavest, J.-M. (1996). Some aspects of zoom

lens camera calibration. IEEE Transactions on Pat-

tern Analysis and Machine Intelligence, 18(11):1105–

1110.

Lim, S. N., Elgammal, A., and Davis, L. S. (2003).

Image-based pan-tilt camera control in a multi-

camera surveillance environment. In Proceedings

International Conference on Multimedia and Expo

(ICME’03), volume 1, pages I–645–8.

Ruiz, A., L´opez-de Teruel, P. E., and Garc´ıa-Mateos, G.

(2002). A note on principal point estimability. In

16th International Conference on Pattern Recognition

(ICPR’02), volume 2, pages 304–307.

Senior, A., Hampapur, A., and Lu, M. (2005). Ac-

quiring multi-scale images by pan-tilt-zoom control

and automatic multi-camera calibration. In Seventh

IEEE Workshops on Application of Computer Vision

(WACV/MOTION’05), pages 433–438.

Sinha, S. and Pollefeys, M. (2004). Towards calibrating

a pan-tilt-zoom camera network. In OMNIVIS 2004,

ECCV Conference Workshop CD-rom proceedings.

Svoboda, T., Martinec, D., and Pajdla, T. (2005). A con-

venient multi-camera self-calibration for virtual envi-

ronments. PRESENCE: Teleoperators and Virtual En-

vironments, 14(4):407–422.

Tsai, . R. Y. (1987). A versatile camera calibration tech-

nique for high-accuracy 3d machine vision metrology

using off-the-shelf tv cameras and lenses. IEEE J. of

Robotics and Automation, RA-3(4):323–344.

Willson, R. G. (1994). Modeling and Calibration of Au-

tomated Zoom Lenses. PhD thesis, Carnegie Mellon

University, Pittsburgh, PA, USA.

Zhang, Z. (2000). A ﬂexible new technique for camera cal-

ibration. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 22(11):1330–1334.

INDOOR PTZ CAMERA CALIBRATION WITH CONCURRENT PT AXES