ISPRS Journal of Photogrammetry and Remote Sensing 102 (2015) 62–72

Contents lists available at ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing

journal homepage: www.elsevier.com/locate/isprsjprs

Zoom lens calibration with zoom- and focus-related intrinsic parameters

applied to bundle adjustment
Shunyi Zheng, Zheng Wang, Rongyong Huang ⇑
School of Remote Sensing and Information Engineering, Wuhan University, No. 129 Luoyu Road, Wuhan 430079, People’s Republic of China

a r t i c l e i n f o a b s t r a c t

Article history: A zoom lens is more flexible for photogrammetric measurements under diverse environments than a
Received 5 November 2014 fixed lens. However, challenges in calibration of zoom-lens cameras preclude the wide use of zoom lenses
Received in revised form 12 January 2015 in the field of close-range photogrammetry. Thus, a novel zoom lens calibration method is proposed in
Accepted 13 January 2015
this study. In this method, instead of conducting modeling after monofocal calibrations, we summarize
Available online 6 February 2015
the empirical zoom/focus models of intrinsic parameters first and then incorporate these parameters into
traditional collinearity equations to construct the fundamental mathematical model, i.e., collinearity
Keywords:
equations with zoom- and focus-related intrinsic parameters. Similar to monofocal calibration, images
Single-lens reflex cameras
Zoom lens
taken at several combinations of zoom and focus settings are processed in a single self-calibration bundle
Intrinsic parameter adjustment. In the self-calibration bundle adjustment, three types of unknowns, namely, exterior orien-
Collinearity equation tation parameters, unknown space point coordinates, and model coefficients of the intrinsic parameters,
Self-calibration are solved simultaneously. Experiments on three different digital cameras with zoom lenses support the
Bundle adjustment feasibility of the proposed method, and their relative accuracies range from 1:4000 to 1:15,100. Further-
more, the nominal focal length written in the exchangeable image file header is found to lack reliability in
experiments. Thereafter, the joint influence of zoom lens instability and zoom recording errors is further
analyzed quantitatively. The analysis result is consistent with the experimental result and explains the
reason why zoom lens calibration can never have the same accuracy as monofocal self-calibration.
Ó 2015 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier
B.V. All rights reserved.

1. Introduction With regard to zoom lens calibration, common camera calibra-

Compared with a fixed lens, a zoom lens has inherent advanta-
ges in terms of flexibility and controllability. A zoom camera sys-
tem is capable of changing the focal length, focus distance, and
aperture value to suit different fields of view (FOVs), depths of field
(DOFs), and lighting conditions. Thus, zoom lens devices have been
extensively used in various applications, e.g., robotic vision and
active vision (Utsumi et al., 2012; Warren et al., 2013), augmented
reality (Chendeb et al., 2013; Tatzgern et al., 2014), object detec-
tion or tracking (Kim et al., 2013; Tao et al., 2009; Tarhan and
Altug, 2011), and traffic monitoring (Schmidt et al., 2009; Song
and Tai, 2006). However, in the field of close-range photogramme-
try, fixed lenses are more commonly used than zoom lenses mainly
because of difficulties in metric modeling and calibration of zoom
lens cameras (Ahmed and Farag, 2000; Wu et al., 2013).
⇑ Corresponding author.
E-mail addresses: syzheng@263.net (S. Zheng), zheng.wang@whu.edu.cn
(Z. Wang), hry@whu.edu.cn (R. Huang).
tion should be conducted first. Camera calibration is the process of
determining the intrinsic parameters of the camera, including prin-
cipal distance, principal point offsets, and lens distortions (Fraser,
1997; Tsai, 1987). The principal distance and principal offsets are
also known as interior orientation (IO) parameters in the field of
photogrammetry. IO parameters, together with extrinsic parame-
ters also known as exterior orientation (EO) parameters, are able
to obtain the space information of an object through images. Thus,
camera calibration is an indispensable procedure for photogram-
metric 3D measurements.
To date, a large number of mature methods for camera calibra-
tion are being used (Remondino and Fraser, 2006; Salvi et al.,
2002). Among these methods, self-calibration bundle adjustment
is considered to achieve superior accuracy (Remondino and
Fraser, 2006) and has been extensively used in the photogramme-
try community since its introduction in the early 1970s (Lichti
et al., 2012; Parian and Gruen, 2010; Tang et al., 2012; Zhang
et al., 2012). However, self-calibration requires the strong geomet-
ric configurations of image networks (Lichti et al., 2010; Wu et al.,
2013), which is impossible in some cases, such as vulnerable

http://dx.doi.org/10.1016/j.isprsjprs.2015.01.005
0924-2716/Ó 2015 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.
 S. Zheng et al. / ISPRS Journal of Photogrammetry and Remote Sensing 102 (2015) 62–72
cultural heritage recording, urban traffic monitoring, mud–rock
flows, and landslides. Therefore, to use zoom lenses in cases
wherein self-calibration cannot be conducted, zoom lens calibra-
tion as a stand-alone step is necessary.
With regard to zoom lens calibration, the most challenging
problem is that intrinsic parameters vary with changing lens set-
ting. To solve this problem, general strategies are based on a
lookup table (Tarabanis et al., 1992) or interpolation with fitted
functions between the lens settings and corresponding calibrated
intrinsic parameters. From the viewpoint of the number of mono-
focal calibrations involved, previous methods for zoom lens cali-
bration can be mainly divided into the following three categories.
First, the simplest method to calibrate a zoom lens is to conduct
monofocal calibration at each zoom and focus setting and then
store the setting values and calibrated results in a lookup table
(Tarabanis et al., 1992). This method involves a huge number of
monofocal calibrations and becomes extremely inefficient. To
reduce the calibration workload, (Chen et al., 2001) selected a
sampled set of lens settings for calibration and then further con-
structed a sparse table to interpolate the desired intrinsic parame-
ters. Nevertheless, this approach still needs to conduct more than
100 monofocal calibrations.
Second, (Sarkis et al., 2009) calibrated zoom lenses by fitting the
continuous local functions of intrinsic parameters with respect to
zoom and focus settings on the basis of the moving least square
scheme. This method first completes separate monofocal calibra-
tions at some scattered zoom and focus points. Thereafter, the scat-
tered data are clustered into similar small regions, followed by the
modeling of each cluster with a single bivariate polynomial. As sta-
ted previously, this strategy is relatively complex with several
steps. Furthermore, the calibration workload will increase rapidly
with a higher than expected interpolated accuracy.
Finally, in addition to the aforementioned two categories of
methods, photogrammetric practitioners prefer to model the vari-
ations of intrinsic parameters regarding lens settings on the basis
of global regression schemes, which involve less monofocal cali-
brations compared with the aforementioned methods. On the basis
of this strategy, different types of models are adopted. The first
model is the zoom-dependent model (Alvarez et al., 2012; Fraser
and Al-Ajlouni, 2006; Fraser et al., 2012), which disregards the
influence of focus on intrinsic parameters. The second model is
the zoom/focus model (Atienza and Zelinsky, 2001; Wu et al.,
2013), which considers the influence of zooming and focusing.
Compared with the zoom-dependent model, the zoom/focus model
involves more monofocal calibrations. On the basis of the zoom/
focus model, (Wu et al., 2013) used a planar checkerboard to
accomplish zoom lens calibration. This work has two main differ-
ences: (a) fixing and determining the principal point of an image
position by using the focus of expansion, which is similar to the
center of expansion (Li and Lavest, 1996; Willson, 1994); (b) use
of a scale parameter modeled with a bivariate nth-order polyno-
mial to compensate for the influence of focus changes on the
principal distance. Moreover, the third type of model is the
Fig. 1. Flowchart of the implementation of the proposed method.
63
focus-dependent model (Sanz-Ablanedo et al., 2012), which is only
for fixed lenses with alterable focus settings. Thus, this type of
model is less significant than the two aforementioned models. Fur-
thermore, some early models even contained the aperture (Hosoda
et al., 1995; Willson, 1994). However, later research showed that
intrinsic parameters do not change with changing aperture
(Chentt et al., 2000; Läbe and Förstner, 2004; Li and Lavest, 1996).
Among the aforementioned methods, except for the lookup
table-based zoom lens calibration method, other methods share
the following strategy. First, individual monofocal calibrations at
selected configurations are conducted. Second, the lens settings
and corresponding calibrated results are used to model local or glo-
bal functions. Finally, the desired intrinsic parameters are interpo-
lated on the basis of the input lens settings and model coefficients.
From the perspective of optimization theory, such methods do not
find the global optimal solution for model coefficients of intrinsic
parameters, because monofocal calibrations and the subsequent
modeling undoubtedly mean separate local optimizations. There-
fore, it is significant to develop an alternative zoom lens calibration
method with a global optimal model coefficients of intrinsic
parameters.
In such a background, a novel method for zoom lens calibration
is proposed in this study. In contrast to previous methods, this
study highlights the following aspects: (a) the development of
the fundamental mathematical model by integrating zoom- and
focus-related intrinsic parameters into conventional collinearity
equations; (b) the contribution to conduct zoom lens calibration
in only one self-calibration bundle adjustment; (c) the quantitative
analysis of zoom lens instability and zoom recording precision. The
details of this approach are specifically discussed in the following
sections.
2. Methodology
2.1. Overview of the approach
In this study, the empirical models of the intrinsic parameters
are first summarized on the basis of existing studies and then
the proposed method is implemented by combining the empirical
models with traditional collinearity equations to construct the
fundamental mathematical model and further conduct a single
self-calibration bundle adjustment. An advantage of the proposed
method is that it solves all unknowns in only one bundle adjust-
ment, thus, a global optimal solution for zoom lens calibration
can be achieved. Fig. 1 shows the framework of the approach in
detail.
2.2. Empirical zoom/focus models
Brown self-calibration model (Brown, 1971; Fraser, 1997; Tang
et al., 2012) holds up as the optimal formulation for digital camera
calibration (Remondino and Fraser, 2006) and it has been
64 S. Zheng et al. / ISPRS Journal of Photogrammetry and Remote Sensing 102 (2015) 62–72

Fig. 2. Rulers for recording the zoom and focus settings.

successfully employed for monofocal calibration of zoom lens 2.3. Collinearity equations with zoom- and focus-related intrinsic
devices (Fraser and Al-Ajlouni, 2006; Fraser et al., 2012; Wakutsu parameters
and Chikatsu, 2011; Wu et al., 2013). And a large number of studies
have already focused on modeling the variations of intrinsic param- In contrast to normal self-calibration procedures, collinearity
eters associated with lens settings (Agapito et al., 2001; Alvarez equations with zoom- and focus-related intrinsic parameters are
et al., 2012; Burner, 1995; Fraser and Al-Ajlouni, 2006; Fraser used to implement zoom lens calibration. The details of the collin-
et al., 2012; Kim et al., 2010; Li and Lavest, 1996; Sanz-Ablanedo earity equations are as follows:
et al., 2012; Sarkis et al., 2007, 2009; Tarabanis et al., 1992; Wiley
and Wong, 1995; Willson, 1994; Wu et al., 2013). On the basis of x  x0 ðzs ; f s Þ þ Dx ¼ cðzs ; f s Þ aa13 ðXXsÞþb 1 ðYYsÞþc 1 ðZZsÞ
ðXXsÞþb3 ðYYsÞþc3 ðZZsÞ
;
ð2Þ
existing studies, an extended Brown self-calibration model is used y  y0 ðzs ; f s Þ þ Dy ¼ cðzs ; f s Þ aa23 ðXXsÞþb 2 ðYYsÞþc2 ðZZsÞ
ðXXsÞþb3 ðYYsÞþc3 ðZZsÞ
;
in our calibration method:
where:
2
a a
x0 ¼ x0 ðzs ; f s Þ ¼ 0 þ 1 zs þ a 2
2 zs a a
þ 3 f s þ 4 zs f s þ 5 f s ; a Dx ¼ ðx  x0 ðzs ; f s ÞÞðk1 ðf l ; f s Þr 2 þ k2 ðf l ;f s Þ  r 4 Þ
2
2
y0 ¼ y0 ðzs ; f s Þ ¼ b0 þ b1 zs þ b2 z2s þ b3 f s þ b4 zs f s þ b5 f s ; þp1 ðzs ; f s Þ½r 2 þ 2ðx  x0 ðzs ; f s ÞÞ  þ 2p2 ðzs ; f s Þðx  x0 ðzs ; f s ÞÞðy  y0 ðzs ; f s ÞÞ;
2 Dy ¼ ðy  y0 ðzs ; f s ÞÞðk1 ðf l ; f s Þr 2 þ k2 ðf l ;f s Þ  r 4 Þ
c ¼ cðzs ; f s Þ ¼ 0 þ 1 zs þ 2 f s þ 3 z2s þ 4 zs f s þ 5 f s þ 6 z2s f s
c c c c c c c 2
þ2p1 ðzs ; f s Þðx  x0 ðzs ; f s ÞÞðy  y0 ðzs ; f s ÞÞ þ p2 ðzs ; f s Þ½r 2 þ 2ðy  y0 ðzs ; f s ÞÞ ;
2 3
þ 7 zs f s þ 8 z3s þ 9 f s ;
c c c 2 2
r ¼ ðx  x0 ðzs ; f s ÞÞ þ ðy  y0 ðzs ; f s ÞÞ ;
2

2 3 2
ð1Þ
l l l l l
k1 ¼ k1 ðf l ; f s Þ ¼ 0 þ 1 =f l þ 2 =f l þ 3 =f l þ 4 f s þ 5 f s ; l ð3Þ
2 3 2
m m m m m
k2 ¼ k2 ðf l ; f s Þ ¼ 0 þ 1 =f l þ 2 =f l þ 3 =f l þ 4 f s þ 5 f s ; m and x0 , y0 , c, k1 , k2 , p1 , and p2 are expressed in Eq. (1) of Section 2.2.
2
p1 ¼ p1 ðzs ; f s Þ ¼ k0 þ k1 zs þ k2 z2s þ k3 f s þ k4 zs f s þ k5 f s ; The collinearity equations contain three categories of
2 unknowns: EO parameters, unknown space point coordinates,
p2 ¼ p2 ðzs ; f s Þ ¼ 0 þ 1 zs þ 2 z2s þ 3 f s þ 4 zs f s þ 5 f s ;
q q q q q q
and model coefficients. Error equations corresponding to the col-
linearity equations are illustrated by first-order Taylor series
where fai ; bi ; ci ; li ; mi ; ki ; qi g are model coefficients, which have a expansion and can be expressed as follows:
total of 46 variables. zs and f s denote the zoom and focus settings
respectively, and f l in the radial distortion models denotes the focal Vx;y ¼ Ax;y Dt þ Bx;y DP þ Cx;y Dc  Lx;y ; ð4Þ
length. zs , f s and f l are described in more detail in Section 3.2. More-
over, the zoom- and focus-dependent models can be derived by where Vx;y ¼ ½ v x v y T is the residual vector of image point obser-
T
using Eq. (1), i.e., the model will become zoom-dependent if all vation ð x y Þ ; Dt, DP, and Dc denote the correction terms of the EO
coefficients of independents containing f s are set to zero, and the parameters, space point coordinates, and model coefficients,
model becomes focus-dependent if all coefficients of independents respectively. Ax;y , Bx;y , and Cx;y denote the first-order partial deriva-
containing zs or f l are set to zero. tives of the error equations related to the aforementioned correc-

(a) Nikon D3200 (b) Nikon D90 (c) Canon EOS 5D Mark II
Fig. 3. Scatter plots of the Exif focal length and the zoom setting.
 S. Zheng et al. / ISPRS Journal of Photogrammetry and Remote Sensing 102 (2015) 62–72 65

Table 1
Calibrated results of adjacent zoom settings.

Zoom setting fExif (mm) Result ID x0 (mm) y0 (mm) c (mm) k1 k2 p1 p2

Nikon D3200 2.9 52 1 0.1335 0.2826 51.4128 1.58E04 8.55E08 4.77E05 9.04E05
3 52 2 0.1338 0.2907 52.9122 1.58E04 8.55E08 4.69E05 9.10E05
Canon EOS 5D Mark II 0.5 28 3 0.0076 0.1689 27.5544 1.46E04 1.86E07 1.24E05 1.36E05
0.7 28 4 0.0148 0.1419 28.7373 1.30E04 1.68E07 3.21E05 2.04E05

Table 2
Accuracy evaluation of adjacent zoom settings.

Zoom setting fExif (mm) Evaluate accuracy with Ha (mm) RMSb of image residuals (pixel) RMSEc of 3D coordinates (mm)
Nikon D3200 2.9 52 Result 1 9060 0.38 0.624 (1:14500)
Result 2 9060 5.22 9.721 (1:900)
3 52 Result 2 8870 0.38 0.754 (1: 11800)
Result 1 8870 5.22 11.711 (1:800)
Canon EOS 5D Mark II 0.5 28 Result 3 4190 0.19 0.352 (1:11900)
Result 4 4190 12.51 12.457 (1:300)
0.7 28 Result 4 3950 0.24 0.433 (1:9100)
Result 3 3950 13.44 12.251 (1:300)
a
H: the average perpendicular distance from camera stations of each setting to the central plane of the 3D test field.
b
RMS: root mean square.
c
RMSE: root mean square error.

Table 3
Information of the images and stations.

Nikon D3200 Nikon D90 Canon EOS 5D Mark II

zs ðf Exif Þ f s ðf d Þ Stations Images zs ðf Exif Þ f s ðf d Þ Stations Images zs ðf Exif Þ f s ðf d Þ Stations Images

0.5 (22 mm) 0.8 (Infinite) 2 2 0.5 (22 mm) 1 (Infinite) 2 2 0 (24 mm) 6 (4.3 m) 2 2
0.5 (22 mm) 1.1 (4.22 m) 2 2 0.5 (22 mm) 1.25 (4.22 m) 2 2 0 (24 mm) 5 (3.0 m) 2 2
0.5 (22 mm) 1.4 (2.82 m) 2 2 0.5 (22 mm) 1.5 (2.82 m) 2 2 0 (24 mm) 4 (1.9 m) 2 2
0.5 (22 mm) 1.7 (2.11 m) 2 2 0.5 (22 mm) 1.75 (2.11 m) 2 2 0 (24 mm) 3 (1.3 m) 2 2
0.5 (22 mm) 2 (1.68 m) 2 2 0.5 (22 mm) 2 (1.68 m) 2 2 0 (24 mm) 2 (0.9 m) 2 2
1 (26 mm) 0.8 (Infinite) 3 3 1 (26 mm) 1 (Infinite) 2 2 1 (32 mm) 6 (4.3 m) 3 3
1 (26 mm) 1.1 (4.22 m) 3 3 1 (26 mm) 1.25 (4.22 m) 2 2 1 (32 mm) 5 (3.0 m) 3 3
1 (26 mm) 1.4 (2.82 m) 3 3 1 (26 mm) 1.5 (2.82 m) 2 2 1 (32 mm) 4 (1.9 m) 3 3
1 (26 mm) 1.7 (2.11 m) 3 3 1 (26 mm) 1.75 (2.11 m) 2 2 1 (32 mm) 3 (1.3 m) 3 3
1 (26 mm) 2 (1.68 m) 3 3 1 (26 mm) 2 (1.68 m) 2 2 1 (32 mm) 2 (0.9 m) 3 3
2 (38 mm) 0.8 (Infinite) 4 4 2 (38 mm) 1 (Infinite) 2 2 2 (42 mm) 6 (4.3 m) 2 4
2 (38 mm) 1.1 (4.22 m) 4 4 2 (38 mm) 1.25 (4.22 m) 2 2 2 (42 mm) 5 (3.0 m) 2 4
2 (38 mm) 1.4 (2.82 m) 4 4 2 (38 mm) 1.5 (2.82 m) 2 2 2 (42 mm) 4 (1.9 m) 2 4
2 (38 mm) 1.7 (2.11 m) 4 4 2 (38 mm) 1.75 (2.11 m) 2 2 2 (42 mm) 3 (1.3 m) 2 4
3 (52 mm) 0.8 (Infinite) 3 10 3 (52 mm) 1 (Infinite) 2 4 3 (51 mm) 6 (4.3 m) 3 6
3 (52 mm) 1.1 (4.22 m) 3 11 3 (52 mm) 1.25 (4.22 m) 2 4 3 (51 mm) 5 (3.0 m) 4 4
3 (52 mm) 1.4 (2.82 m) 3 14 3 (52 mm) 1.5 (2.82 m) 2 4 3 (51 mm) 4 (1.9 m) 3 6
4 (80 mm) 0.8 (Infinite) 3 13 4 (80 mm) 1 (Infinite) 3 6 4.2 (68 mm) 6 (4.3 m) 3 6
4 (80 mm) 1.1 (4.22 m) 3 17 4 (80 mm) 1.25 (4.22 m) 2 8 4.2 (68 mm) 5 (3.0 m) 4 8

Fig. 4. 3D test field and camera station geometry for Nikon D3200.
66 S. Zheng et al. / ISPRS Journal of Photogrammetry and Remote Sensing 102 (2015) 62–72

(a) Nikon D3200 (b) Nikon D90 (c) Canon EOS 5D Mark II

Fig. 5. Variations of intrinsic parameters.

 S. Zheng et al. / ISPRS Journal of Photogrammetry and Remote Sensing 102 (2015) 62–72
tion terms. Lx;y is the constant item calculated by the approximate
values of unknowns. Moreover, a control point is considered a true
value. Thus, the corresponding partial derivatives in matrix Bx;y are
null. However, counterparts are valid for a checkpoint.
2.4. Bundle adjustment model and the solution
As discussed previously, the EO parameters, unknown space
points, and model coefficients should satisfy Eq. (4). The corre-
sponding error equation can be written as follows:
V ¼ ADt þ BDP þ CDc  L W; ð5Þ
where V is the residual vector of observations consisting of Vx;y in
Eq. (4). A, B, and C are the partial derivative matrixes, which con-
sists of Ax;y , Bx;y , and Cx;y , respectively, in Eq. (4). L is the vector of
constant items, and W is the weight matrix.
Compared with the conventional self-calibration bundle
adjustment for fixed lens camera (Triggs et al., 2000), in our
self-calibration the model coefficients of intrinsic parameters are
introduced as additional parameters instead of intrinsic parame-
ters themselves. The normal equation with respect to the error
equation (Eq. (5)) has a similar structure to that of the common
self-calibration bundle adjustment and the final solution is
achieved by an iterative least squares estimation process.
Given that the fundamental model is nonlinear and the solution
process is iterative, we need to estimate the initial values for all
unknowns. The initial model coefficients of principal point offsets
and lens distortion parameters are set to zero. According to the
empirical model of the principal distance in Eq. (1) and treating
the focal length in the exchangeable image file (Exif) as c, the initial
fci g ði ¼ 0; 1;    ; 9Þ are acquired through linear least squares esti-
mation. For initial EO parameters of each image, three dimensional
DLT (Abdel-Aziz, 1971) is firstly performed. Then the EO parame-
ters are extracted from 11 coefficients of DLT (Heikkila and
Silven, 1997), wherein the principal point offsets are fixed as zero
and the principal distance as the focal length in the Exif. Thereafter,
iterative space resection is further performed to find more accurate
EO parameters.
3. Experimental description and results
3.1. Zoom lens devices and 3D test field
Three digital single-lens reflex cameras fitted with different
zoom lenses are used in our experiments. The first camera is a
Nikon D3200 with a Nikkor 18 mm to 105 mm f/3.5 G to 5.6 G
ED VR zoom lens. The second camera is a Nikon D90 with another
Nikkor 18 mm to 105 mm f/3.5 G to 5.6 G ED VR zoom lens. The
third camera is a Canon EOS 5D Mark II with a Canon EF 24 mm
to 70 mm f/2.8 L USM zoom lens. These cameras all have
complementary metal–oxide–semiconductor sensors and are of
low-, medium-, and high-level consumer grade classifications,
respectively. Among the three cameras, Nikon D3200 has the high-
est full resolution of 6016  4000 and the smallest pixel size of
3.85 lm  3.85 lm. The two parameters for Nikon D90 are
4288  2848 and 5.5 lm  5.5 lm and for Canon EOS 5D Mark II
are 5616  3744 and 6.4 lm  6.4 lm.
The zoom lens calibration in this study is completed in a 3D test
field, with a depth of approximately 2.1 m, width of 6.2 m, and
height of 3.0 m. A total of 204 circular control points, with diame-
ters of 25 mm, were obtained in the test field. Positioning precision
is approximately 0.12 mm.
67
3.2. Zoom and focus recording
In previous research, the nominal focal length in the Exif file
(denoted as f Exif ) is commonly used for modeling the variation of
intrinsic parameters because of its convenience (Fraser et al.,
2012; Nakano and Chikatsu, 2009; Wu et al., 2013). But few studies
have been done on the recording accuracy of f Exif . To verify the reli-
ability of f Exif and its effect on zoom lens calibration, we conducted
the following tests. First, a thin and soft ruler (shown in Fig. 2) with
a minimum scale of 1 mm was pasted on the zoom ring of each
camera to record the zoom setting (denoted as zs ). Then, cameras
were set up to M mode, locked at manual focusing, and all settings
were unchanged. Thereafter, sequential images were taken by
turning the zoom ring at an interval of 1 mm from the smallest
end to the largest end of zooming by simultaneously recording
the zoom settings through the ruler. The scatter graphs of manu-
ally recorded zs versus automatically recorded f Exif for the three
devices are shown in Fig. 3. Based on their correspondence rela-
tionship, the regression model of the focal length (denoted as f l )
with zs , displayed in Eq. (6) was established.
8 2 3
< 16:53 þ 12:21zs  2:26zs þ 0:91zs ; Nikon D3200
>
f l ¼ 16:85 þ 11:49zs  2:15zs þ 0:83z3s ; Nikon D90
2
>
:
23:74 þ 7:72zs þ 0:26z2s þ 0:95z3s ; Canon EOS 5D Mark II
ð6Þ
From Fig. 3, we observe that some images have different zs but
share the same f Exif . Therefore, we assume the correctness of f Exif
and further conduct another test on Nikon D3200 and Canon EOS
5D Mark II. We complete two individual self-calibrations at two
adjacent zs , which have the same f Exif . The calibrated results are
shown in Table 1. Thereafter, we further evaluate the accuracy of
each self-calibration by using intrinsic parameters that are cali-
brated by itself and by the adjacent group. The detailed evaluation
procedure is described in Section 4.1 and the evaluated accuracies
are shown in Table 2.
For adjacent zs in Table 1, sharing a common f Exif , their intrinsic
parameters are different. Shown in Table 2, the accuracies evalu-
ated with each other’s intrinsic parameters show a serious
degradation even with the relative accuracy below 1:1000.
Conceivably, if f Exif is used to interpolate intrinsic parameters,
two adjacent zs having a common f Exif will share same intrinsic
parameters, which is probable to cause serious accuracy degrada-
tion at these settings.
Therefore, we do not use f Exif to model or interpolate intrinsic
parameters during the calibration of our zoom lens devices but
use artificial rulers (shown in Fig. 2) to record zs for the sake of
recording accuracy. Moreover, the focus distance (denoted as f d )
or focus setting (denoted as f s ) value is not written in the Exif for
Canon EOS 5D Mark II. Although the Canon zoom lens marks sev-
eral f d positions on the lens barrel, the interval is rather wide so
that accurate f d is hard to be estimated. Thus, we also use artificial
rulers to record the f s for the sake of a unified procedure.
3.3. Image acquisition
To acquire sharp images of the test field at different zoom and
focus settings, the DOF must be considered. Three adjustable pho-
tography parameters, i.e., focal length, focus distance, and aper-
ture, affect the DOF of an image. As mentioned previously, the
aperture is a negligible factor in modeling the variations of intrin-
sic parameters. Therefore, during the process of taking images,
aperture, exposure time, and ISO are adjusted properly to capture
sharp and bright images. Since the use of rulers for focus recording,
68 S. Zheng et al. / ISPRS Journal of Photogrammetry and Remote Sensing 102 (2015) 62–72

Table 4
Accuracies of ZLC I and ZLC II compared with monofocal self-calibration.

Device zs ðf Exif Þ f s ðf d Þ H Self-calibration ZLC I ZLC II

(mm)
RMS of image RMSE of 3D RMS of image RMSE of 3D RMS of image RMSE of 3D
residuals (pixel) coordinates (mm) residuals (pixel) coordinates (mm) residuals (pixel) coordinates (mm)
Nikon D3200 0.75 0.8 5410 0.23 0.434 (1:12500) 0.83 1.002 (1:5400) 0.87 1.137 (1:4800)
(24 mm) (Infinite)
0.75 2.0 5960 0.21 0.494 (1:12100) 0.89 1.394 (1:4300) 0.57 0.844 (1:7100)
(24 mm) (1.68 m)
1.5 1.7 6580 0.28 0.506 (1:13000) 0.63 1.416 (1:4600) 1.25 2.832 (1:2300)
(32 mm) (2.11 m)
2.5 1.4 8890 0.33 0.677 (1:13100) 0.5 0.964 (1:9200) 0.38 0.77 (1:11600)
(45 mm) (2.82 m)
3.5 0.8 9000 0.38 0.538 (1:16700) 0.64 0.595 (1:15100) 2.29 2.436 (1:3700)
(66 mm) (Infinite)
3.5 1.1 8890 0.47 0.501 (1:17700) 1.01 1.155 (1:7700) 2.33 2.999 (1:3000)
(66 mm) (4.22 m)
Nikon D90 0.75 1.0 5000 0.2 0.327 (1:15300) 0.63 1.168 (1:4300) 1.09 1.935 (1:2600)
(24 mm) (Infinite)
0.75 2.0 5070 0.22 0.489 (1:10400) 0.45 0.715 (1:7100) 1.02 1.904 (1:2700)
(24 mm) (1.68 m)
1.5 1.0 6030 0.25 0.325 (1:18500) 0.83 1.517 (1:4000) 1.24 2.217 (1:2700)
(32 mm) (Infinite)
2.5 1.25 9000 0.29 0.478 (1:18800) 0.35 0.676 (1:13300) 0.74 1.77 (1:5100)
(45 mm) (4.22 m)
2.5 1.5 8980 0.3 0.564 (1:15900) 0.35 0.665 (1:13500) 0.65 1.518 (1:5900)
(45 mm) (2.82 m)
3.5 1.0 9080 0.38 0.625 (1:14500) 0.81 1.144 (1:7900) 1.26 2.07 (1:4400)
(66 mm) (Infinite)
Canon EOS 0.5 4 4270 0.18 0.36 (1:11800) 0.63 0.923 (1:4600) 1.62 2.163 (1:2000)
5D Mark II (28 mm) (1.9 m)
0.5 2 4340 0.21 0.378 (1:11500) 0.54 0.895 (1:5800) 1.27 2.064 (1:2100)
(28 mm) (0.9 m)
1.5 3 5500 0.25 0.436 (1:12600) 0.68 0.872 (1:6300) 0.88 1.316 (1:4200)
(35 mm) (1.3 m)
2.5 6 6290 0.22 0.349 (1:18000) 0.78 1.013 (1:6200) 1.35 1.964 (1:3200)
(46 mm) (4.3 m)
2.5 5 5720 0.21 0.294 (1:19500) 0.74 0.781 (1:7300) 1.26 1.694 (1:3400)
(46 mm) (3.0 m)
3.5 6 7120 0.25 0.306 (1:23200) 0.55 0.581 (1:12300) 0.77 0.819 (1:8700)
(58 mm) (4.3 m)

the autofocus function is restricted because we find that the same significantly with increasing focal length. Fig. 4 shows the 3D test
focus ruler value will denote different focus distances before and field and camera station geometry for Nikon D3200.
after an autofocus operation. In our experiment, all images are
taken in the M mode, at manual focus setting, and under full
3.5. Calibration results
resolution.

To verify the proposed method, zoom lens calibration for the

3.4. Image geometry
Three cameras fitted with different zoom lenses are calibrated
in our experiments. For each device, a total of 19 combinations
of zoom and focus settings are selected for self-calibration. Table 3
shows the detailed information of the images and stations of our
zoom lens calibration experiments.
In Table 3, zs and f s are dimensional numbers, which are
recorded manually by reading the artificial rulers. f Exif is the nom-
inal focal length that is automatically written in the Exif file of each
image. f d is the focus distance. For Nikon D3200 and Nikon D90,
the focus distance is automatically written in the Exif, while it does
not exist in the Exif for Canon EOS 5D Mark II, so we estimate it
from the lens manually.
At large focal length and small focus distance, sharp images
cannot be acquired because of the depth of the field. As a conse-
quence, a smaller number of configurations are used at larger focal
lengths. And the FOV will decrease with increasing focal length,
thus, a small number of control points can be contained in an
image. To maintain approximately balanced control conditions at
each configuration, the number of images taken increases
three devices was conducted according to the procedure illustrated
in Fig. 1, which is denoted as ZLC I. Besides, we also performed
another zoom lens calibration (denoted as ZLC II) based on the
strategy that conducting monofocal self-calibrations and then
using the calibrated intrinsic parameters to solve the model coeffi-
cients. The intrinsic parameters consisting of x0 , y0 , c, k1 , k2 , p1 , and
p2 were introduced into each monofocal self-calibration of ZLC II.
Subsequently, model coefficients of ZLC II were solved through
least square estimation, wherein the choice of the order of each
parameter model was same as ZLC I. Moreover, both ZLC I and
ZLC II were implemented by using identical calibration data, i.e.,
image coordinates, 3D space coordinates, and zoom and focus
recording data.
Some current zoom/focus-dependent calibration methods fix
the principal point offsets during zooming and focusing operations
(Li and Lavest, 1996; Wu et al., 2013) and deem that the discrepan-
cies of the principal point offsets can be compensated by correlated
parameters, such as decentering distortion parameters and EO
parameters. To verify the applicability and flexibility of the sum-
marized empirical zoom/focus models in Eq. (1), only the constant
terms of the principal point offset models for Canon EOS 5D Mark II
 S. Zheng et al. / ISPRS Journal of Photogrammetry and Remote Sensing 102 (2015) 62–72 69

were adopted in ZLC I, i.e., the coefficients of other terms in the Although the three zoom lens devices are of distinct consumer
principal point offset models are set to zero. Accordingly, in ZLC grades and have different photographic configurations, there is not
II for Canon EOS 5D Mark II, we first performed monofocal self-cal- a notable difference in accuracy among them. Besides, we can also
ibrations and then fixed the principal point offsets as the averages observe that monofocal self-calibration outperforms zoom lens
to conduct self-calibrations again for other parameters (c, k1 , k2 , p1 , calibrations. Therefore, there is necessity to further explore the
and p2 ). Moreover, the principal distance model of Nikon D3200 is factor causing accuracy loss during zoom lens calibration.
specified as second-order polynomials, whereas the counterparts
of other cameras are specified as third-order polynomials. The
4.2. The quantitative analysis of zoom lens instability
model coefficients and theoretical accuracies of both zoom lens
calibrations for all three devices are shown in Table A.1, Appendix
Most studies have focused on the variations of intrinsic param-
A.
eters with zoom and focus operations. However, only a few of
To exhibit the variation trends of intrinsic parameters for the
these studies have investigated instability problems in detail. For
three devices intuitively, Fig. 5 shows the comparison between
example, (Shortis et al., 2006) studied the effect of gravity as cam-
the intrinsic parameters computed with the model coefficients of
eras rolled and (Läbe and Förstner, 2004) studied geometric stabil-
ZLC I (in blue1 color) and the intrinsic parameters solved by mono-
ity over time. To further explain the reason why zoom lens
focal self-calibrations of ZLC II (in red color).
calibration can never have the same accuracy as monofocal self-
Intrinsic parameter variations are consistent with previous
calibration and to effectively use zoom lenses for photogrammetric
research (Alvarez et al., 2012; Fraser and Al-Ajlouni, 2006; Kim
applications, we need to ascertain zoom lens instability further and
et al., 2010; Li and Lavest, 1996; Willson, 1994; Wu et al., 2013)
examine its influence on accuracy.
(Fig. 5). The principal distance and radial distortion parameters
In our zoom lens calibration, two compelling factors may cause
show similar variation trends in different devices and can be fitted
accuracy degradation. First, suppose that we turn the zoom ring to
well, whereas the principal point offsets and decentering distortion
the same zoom setting twice, trivial offsets of the internal compo-
parameters do not show the same trends. Although the principal
nents will inevitably occur in the zoom lens. Second, given that
point offsets and decentering distortion parameters cannot be fit-
the zoom and focus settings are subject to the judgment of the
ted well, their discrepancies can be compensated for by each other
recorders, reading errors are also unavoidable. To analyze the joint
and the EO parameters because of their correlations, thus leading
influence of the factors, an additional experiment is further con-
to slight accuracy degradation, which is also reported by (Tang
ducted on Nikon D3200 and Canon EOS 5D Mark II. The additional
and Fritsch, 2013; Wu et al., 2013). Moreover, we observe that
experiment is discussed in detail in the following paragraphs.
the zoom setting is the dominant factor affecting intrinsic param-
As shown in Fig. 6, two circular targets with space distance
eters. However, the influence of focusing increases with increasing
D ¼ 1:95 m are pasted on a flat wall. The camera is stationary on
focal length, which is also reported by (Fraser and Al-Ajlouni, 2006;
a tripod, and the optical axis of the lens is approximately perpen-
Wu et al., 2013).
dicular to the wall. The distance between the camera and wall is
denoted as L (11 m). The camera is set to M mode and manual
4. Accuracy evaluation and discussion focus. The experimental group of images is obtained as follows:
(a) rotate the zoom ring to a specific zoom setting and take a
4.1. Accuracy evaluation photo; (b) rotate the zoom ring to different degrees and back to
the same position and then take a photo again; (c) repeat the oper-
To test the performance of the proposed method in terms of ation five times. A control group of images is also obtained by tak-
accuracy, a total of 18 (6 for each device) bundle adjustments at ing five sequential photos without rotating the zoom ring, namely,
settings that were not selected for aforementioned zoom lens cal- by pressing the shutter for five times. Finally, each image distance
ibrations were performed by using 124 of the 204 3D control (denoted as d) between two targets is measured for further
points as checkpoints. Within bundle adjustments, intrinsic analysis.
parameters calibrated by monofocal self-calibration and interpo- For convenience, we assume that the influence of lens distortion
lated with model coefficients of ZLC I and ZLC II were considered on d is trivial and negligible. On the basis of the principle of similar
known values. Thereafter, the RMSE against the true 3D coordi- triangles, c=L ¼ d=D. Thereafter, mc ¼ md L=D can be derived, where
nates of all checkpoints, as the absolute accuracy, is considered md and mc represent the standard deviations of the image length
the accuracy indicator. Besides, the RMS of the image projection and principal distance, respectively. mc also denotes the principal
residuals is also recorded. The accuracies evaluated by using the distance uncertainty.
ZLC I and ZLC II against monofocal self-calibration are shown in To analyze the effect of principal distance uncertainty quantita-
Table 4. tively, we first conduct several monofocal self-calibrations at typi-
Table 4 shows that the relative accuracies of ZLC I range from cal zoom settings for Nikon D3200 and Canon EOS 5D Mark II.
1:4000 to 1:15,100, which is better than that of ZLC II (1:2000 to Thereafter, we add small increments, i.e., mc , mc , 2mc , and
1:11,600). In experiments, ZLC I was implemented on the basis of
the developed fundamental mathematical model with model coef-
ficients as the additional parameters and all unknowns were solved
in a single bundle adjustment. Thus, the solution of model coeffi-
cients achieved by ZLC I is a global optimum. While for ZLC II,
monofocal calibrations and the subsequent modeling intrinsic
parameters mean separate local optimizations. Compared with
ZLC I, ZLC II needs to solve extra intrinsic parameters at the stage
of monofocal calibrations, which means more potential accuracy
loss in practice.

1
For interpretation of color in Fig. 5, the reader is referred to the web version of
this article. Fig. 6. Geometric schematic graph.
70 S. Zheng et al. / ISPRS Journal of Photogrammetry and Remote Sensing 102 (2015) 62–72

Table 5
The principal distance uncertainty.

Device Experimental group Control group

zs ðf Exif Þ md (mm) mc (mm) zs ðf Exif Þ md (mm) mc (mm)

Canon EOS 5D Mark II 1 (32 mm) 0.0231 0.1302 1 (32 mm) 0.0003 0.0017
3 (51 mm) 0.0465 0.2621 3 (51 mm) 0.0006 0.0038
Nikon D3200 2 (38 mm) 0.0196 0.1108 2 (38 mm) 0.0003 0.0019
4 (80 mm) 0.0866 0.4883 4 (80 mm) 0.0019 0.01061

Table 6
The influence of principal distance uncertainty on measurement accuracy.

zs ðf Exif Þ f s ðf d Þ Principal distance H (mm) RMS of image residuals (pixel) RMSE of 3D coordinates (mm)

Nikon D3200 2 (38 mm) 1.1 (4.22 m) c 8710 0.26 0.542 (1:16100)
2 (38 mm) 1.1 (4.22 m) c  mc 8710 0.5 0.892 (1:9800)
2 (38 mm) 1.1 (4.22 m) c + 2 mc 8710 1.1 1.653 (1:5300)
4 (80 mm) 0.8 (Infinite) c 8860 0.57 0.433 (1:20500)
4 (80 mm) 0.8 (Infinite) c + mc 8860 1.2 1.038 (1:8500)
4 (80 mm) 0.8 (Infinite) c  2 mc 8860 2.78 3.099 (1:2900)
Canon EOS 5D Mark II 1 (32 mm) 6 (4.3 m) c 5120 0.22 0.441 (1:11600)
1 (32 mm) 6 (4.3 m) c  mc 5120 0.52 1.075 (1:4800)
1 (32 mm) 6 (4.3 m) c + 2 mc 5120 0.96 2.534 (1:2000)
3 (51 mm) 6 (4.3 m) c 7070 0.27 0.404 (1:17500)
3 (51 mm) 6 (4.3 m) c + mc 7070 0.66 1.332 (1:5300)
3 (51 mm) 6 (4.3 m) c  2 mc 7070 1.28 2.945 (1:2400)

2mc , to the principal distance and then use bundle adjustment with
fixed intrinsic parameters to evaluate the accuracies, similar to
that discussed in Section 4.1. The results are shown in detail in
Table 6.
As shown in Tables 5 and 6, the mc or twofold mc offsets of the
principal distance will lead to noticeable accuracy degradation.
This result also suggests that slight zoom lens instability and
reading errors can cause high accuracy degradation. This high
accuracy degradation is also one reason why zoom lenses can
never perform as well as fixed lenses in terms of photogrammet-
ric accuracy. Specifically, if the devices and recording approach
are given, zoom lens instability and recording errors will be
unavoidable. Thus, zoom lens calibration will not achieve the
same accuracy as monofocal lens calibration will preclude high
accuracy photogrammetry with the use of consumer grade
cameras with zoom lenses. Unless improved devices and more
accurate recording approaches are used, these problems can
never be tackled effectively.
that based on the strategy by conducting monofocal calibra-
tions followed by modeling.
(3) We found that the repeatable recording of focal length in the
Exif at adjacent zoom settings when using our Nikon D3200
and Canon EOS 5D Mark II and demonstrated the possibility
in serious accuracy degradation by using the Exif focal
length to model and interpolate intrinsic parameters.
Despite this finding, it is recommended that practitioners
make use of the nominal value in the Exif to conduct zoom
lens calibration due to its easy access to record and the
record precision certainly depends on the behavior of the
lens.
(4) Regardless of how accurately the models fit variations of
intrinsic parameters, lens instability and reading error that
use artificial rulers are unavoidable factors that lead to
noticeable accuracy degradation. Thus, zoom lens calibration
can never achieve the same accuracy as monofocal self-
calibration.

5. Conclusions
In this study, we present a novel method for zoom lens calibra-
tion in a 3D test field. We reach the following conclusions on the
basis of the experimental results and related analysis:
(1) The zoom lens calibration strategy, which incorporates the
empirical zoom/focus models of intrinsic parameters into
traditional collinearity equations to construct collinearity
equations with zoom- and focus-related intrinsic parame-
ters, is proven feasible.
(2) In the proposed method, all images with different zoom and
focus settings can be processed in a single bundle adjust-
ment and model coefficients of intrinsic parameters are
solved simultaneously. Thus, the proposed method success-
fully achieves a global optimal model coefficients for zoom
lens calibration. In our experiments, the evaluated accura-
cies of the zoom lens calibration based on the proposed
strategy range from 1:4000 to 1:15,100, which outperform
On the basis of these conclusions, our future work will focus on
two issues, i.e. implementing the proposed method using a planar
board with coded targets and exploring more accurate and efficient
automatic way to record zoom and focus setting.
Acknowledgments
This research is supported by the Chinese National Natural
Science Foundation Programs (Nos. 41301518 and 41171357),
the Wuhan Science and Technology Project Plan (No.
2013060501010149), and the Fundamental Research Funds for
the Central Universities (No. 2014213020201). We are very grate-
ful to the anonymous reviewers and members of the editorial
board for their suggestions of improving this article.
Appendix A. Model coefficients and their theoretical accuracies
See Table A.1.
Table A1
Model coefficients and their theoretical accuracies.

Intrinsic Parameter Independent Item ZLC I ZLC II

Nikon D3200 Nikon D90 Cannon EOS 5D Mark II Nikon D3200 Nikon D90 Cannon EOS 5D Mark II
Coefficient r Coefficient r Coefficient r Coefficient r Coefficient r Coefficient r
x0 Constant 0.1312 7.69E03 0.0616 0.0143 0.0215 7.85E04 0.0248 0.2193 0.093 0.4303 0.0765 0.0678
zs 0.1438 3.41E03 0.1078 5.42E03 0 0 0.0763 0.0858 0.0783 0.125 0 0
z2s 0.0318 5.86E04 6.14E03 9.15E04 0 0 0.0261 0.0118 0.0664 0.0144 0 0
fs 0.0409 0.0106 0.0592 0.0184 0 0 0.0353 0.3003 0.06 0.556 0 0
z s fs 0.0258 1.93E03 5.48E03 3.10E03 0 0 0.0119 0.0465 0.0236 0.0683 0 0
2
fs 0.0192 3.47E03 0.0235 5.67E03 0 0 0.0158 0.0988 0.0273 0.1742 0 0

y0 Constant 0.0467 6.33E03 0.344 0.0102 0.1685 4.51E04 0.1183 0.1101 0.3801 0.1887 0.1825 0.0343
zs 0.0491 3.81E03 0.0517 5.97E03 0 0 0.0588 0.0431 0.0102 0.055 0 0

S. Zheng et al. / ISPRS Journal of Photogrammetry and Remote Sensing 102 (2015) 62–72
z2s 5.30E03 6.71E04 4.31E03 1.00E03 0 0 0.0407 5.90E03 9.60E03 0.0061 0 0
fs 0.0159 8.56E03 8.85E03 0.0129 0 0 0.03264 0.1508 0.0156 0.244 0 0
z s fs 0.0155 2.06E03 1.24E03 3.19E03 0 0 7.56E03 0.0233 7.60E03 0.0301 0 0
2
fs 8.49E04 2.84E03 4.86E03 3.94E03 0 0 0.0114 0.0496 2.50E03 0.0761 0 0

c Constant 19.2422 4.83E03 18.069 0.0432 24.6109 9.09E03 19.8555 2.655 17.8902 2.174 24.1406 1.8419
zs 3.7684 2.97E03 6.9743 0.0346 4.7168 4.20E03 2.8735 1.0387 7.1564 3.276 4.9448 0.79
fs 0.0189 6.54E03 0.3714 0.0841 0.1223 7.69E03 0.1684 3.6368 3.81E03 2.095 0.47 1.5913
z2s 2.6119 5.21E04 1.0016 0.0105 1.9645 1.72E03 2.887 0.1428 0.409 0.662 1.7624 0.277
z s fs 0.2853 1.53E03 1.0692 0.04 0.0174 2.28E03 0.3124 0.5631 0.4939 3.7851 0.0136 0.4419
2
fs 8.35E03 2.20E03 0.0502 0.0547 0.031 2.04E03 0.104 1.1968 0 0 0.1257 0.4297
z2s fs 0 0 0.0555 5.27E03 0.0207 3.38E04 0 0 0.246 0.3369 0.0195 0.0553
zs f s
2 0 0 0.3119 0.0116 2.91E03 3.03E04 0 0 0.0083 1.0808 3.02E03 0.0584
z3s 0 0 0.2584 1.13E03 0.1416 2.36E04 0 0 0.4101 0.0574 0.108 0.0315
3
fs 0 0 0.0933 0.0118 2.42E03 1.71E04 0 0 0.0659 0.2651 9.82E03 0.0365

k1 Constant 3.54E04 2.27E06 2.73E05 4.35E06 1.46E04 1.84E06 1.89E05 1.30E05 4.50E05 2.90E05 1.69E04 8.00E06
1/fl 2.83E02 2.41E04 0.0142 3.80E04 3.60E03 2.23E04 0.0166 1.40E03 0.0224 2.72E03 7.73E03 2.38E04
1/fl 2 1.177 8.79E03 0.4154 0.0133 0.2035 8.37E03 0.5038 0.0493 0.7268 0.0991 0 0
1/f l
3 18.5138 0.0996 0.396 0.1443 2.4012 9.91E02 0.7354 0.5447 2.8818 1.1175 0.8656 0.0712
fs 5.58E05 2.20E06 3.21E05 4.21E06 4.38E06 2.58E07 8.33E06 1.10E05 4.23E05 3.10E05 1.31E06 3.00E06
2 2.43E05 8.32E07 1.55E05 1.41E06 4.18E07 2.98E08 6.71E06 4.00E06 1.79E05 1.10E05 6.07E08 3.76E07
fs
k2 Constant 1.59E06 1.29E08 3.46E07 2.61E08 1.93E07 4.37E09 2.69E07 4.14E08 5.80E07 8.97E08 1.33E07 1.91E08
1/fl 2.10E04 1.39E06 5.10E05 2.28E06 3.44E05 5.35E07 3.91E05 4.00E06 7.18E05 8.00E06 7.58E06 5.42E07
1/f l
3 7.52E03 5.09E05 2.32E03 7.92E05 1.60E03 2.03E05 1.81E03 1.60E04 3.07E03 3.06E04 0 0
1/f l
3 9.32E02 5.80E04 0.0203 8.62E04 0.0184 2.41E04 0.0138 1.77E03 0.0281 3.46E03 1.10E03 1.62E04
fs 2.55E07 1.29E08 3.82E08 2.56E08 1.18E08 6.32E10 8.21E09 3.56E08 1.14E07 9.61E08 2.62E09 7.27E09
2
fs 1.02E07 4.88E09 2.07E08 8.59E09 1.50E09 7.25E11 2.03E09 1.31E08 4.15E08 3.29E08 3.03E10 8.56E10

p1 Constant 1.98E05 5.08E06 6.82E05 9.50E06 2.32E05 7.42E07 8.25E06 4.00E05 6.09E05 1.18E04 7.20E05 2.10E05
zs 1.86E05 1.54E06 4.16E05 2.39E06 7.08E06 2.85E07 3.61E05 1.60E05 2.22E05 3.40E05 1.29E05 8.00E06
z2s 4.63E06 1.81E07 8.48E06 2.41E07 1.21E06 3.75E08 1.04E05 2.00E06 5.66E06 4.00E06 1.07E06 1.00E06
fs 4.28E05 6.86E06 1.22E05 1.22E05 1.00E06 3.40E07 1.02E05 5.50E05 3.90E06 1.53E04 8.03E06 1.10E05
z s fs 5.79E07 8.52E07 9.91E06 1.36E06 6.77E07 5.30E08 2.90E06 9.00E06 2.66E06 1.90E05 4.72E07 2.00E06
2
fs 1.39E05 2.23E06 9.35E06 3.76E06 1.79E07 4.41E08 3.88E06 1.80E05 1.47E06 4.80E05 8.88E07 1.00E06

p2 Constant 1.58E05 4.05E06 4.73E05 6.42E06 1.75E05 8.75E07 3.79E05 3.30E05 7.64E05 4.00E05 3.47E05 1.60E05
zs 6.18E05 1.46E06 1.91E05 2.14E06 3.71E06 3.47E07 3.17E05 1.30E05 3.40E06 1.20E05 2.83E06 6.00E06
z2s 1.42E05 1.98E07 6.26E06 2.66E07 2.63E06 4.67E08 4.27E06 2.00E06 1.30E06 1.00E06 2.71E06 9.83E07
fs 1.37E05 5.57E06 3.05E06 8.26E06 2.39E06 4.68E07 1.99E05 4.60E05 1.33E05 5.10E05 4.96E06 9.00E06
z s fs 8.74E06 8.21E07 4.35E06 1.20E06 2.32E07 7.42E08 4.44E06 7.00E06 2.06E07 6.00E06 6.29E08 1.00E06
2
fs 2.26E06 1.83E06 4.59E06 2.56E06 3.50E07 6.08E08 6.71E06 1.50E05 5.63E06 1.60E05 5.71E07 1.00E06

71
⁄r denotes the theoretical accuracies of model coefficients, which can be estimated by using the inverse coefficient matrix of the normal equations and unit weighted RMS.
72 S. Zheng et al. / ISPRS Journal of Photogrammetry and Remote Sensing 102 (2015) 62–72

References Sanz-Ablanedo, E., Chandler, J.H., Wackrow, R., 2012. Parameterising internal
camera geometry with focusing distance. Photogramm. Rec. 27, 210–226.
Sarkis, M., Senft, C.T., Diepold, K., 2007. Modeling the variation of the intrinsic
Abdel-Aziz, Y., 1971. Direct linear transformation from comparator coordinates in
parameters of an automatic zoom camera system using moving least-squares,
close-range photogrammetry. In: ASP Symposium on Close-Range
Automation Science and Engineering, 2007. CASE 2007. In: IEEE International
Photogrammetry in Illinois, 1971.
Conference on. IEEE, pp. 560–565.
Agapito, L., Hayman, E., Reid, I., 2001. Self-calibration of rotating and zooming
Sarkis, M., Senft, C.T., Diepold, K., 2009. Calibrating an automatic zoom camera with
cameras. Int. J. Comput. Vis. 45, 107–127.
moving least squares. IEEE Trans. Automat. Sci. Eng. 6, 492–503.
Ahmed, M.T., Farag, A.A., 2000. A neural optimization framework for zoom lens
Schmidt, W., Rudolph, M., Papenfuss, A., Friedrich, M., Mohlenbrink, C.,
camera calibration. Comput. Vis. Pattern Recogn., 1403–1409
Kaltenhauser, S., Furstenau, N., 2009. Remote airport traffic control center
Alvarez, L., Gómez, L., Henríquez, P., 2012. Zoom dependent lens distortion
with augmented vision video panorama. In: Digital Avionics Systems
mathematical models. J. Math. Imaging Vis. 44, 480–490.
Conference, 2009. DASC’09. IEEE/AIAA 28th. IEEE, pp. 4. E. 2-1-4. E. 2–15.
Atienza, R., Zelinsky, A., 2001. A practical zoom camera calibration technique: an
Shortis, M., Bellman, C., Robson, S., Johnston, G., Johnson, G., 2006. Stability of zoom
application on active vision for human-robot interaction. Proc. Aust. Conf.
and fixed lenses used with digital SLR cameras. In: Proceedings of the ISPRS
Robotics Automat., 85–90
Commission V Symposium of Image Engineering and Vision Metrology,
Brown, D.C., 1971. Close-range camera calibration. Photogramm. Eng. 37, 855–866.
Citeseer, pp. 285–290.
Burner, A.W., 1995. Zoom lens calibration for wind tunnel measurements. Video
Song, K.-T., Tai, J.-C., 2006. Dynamic calibration of pan–tilt–zoom cameras for traffic
Metrics IV SOIE, 19–33.
monitoring. IEEE Trans. Syst. Man Cybern. 36, 1091–1103.
Chen, Y.-S., Shih, S.-W., Hung, Y.-P., Fuh, C.-S., 2001. Simple and efficient method of
Tang, R.F., Fritsch, D., 2013. Correlation analysis of camera self-calibration in close
calibrating a motorized zoom lens. Image Vis. Comput. 19, 1099–1110.
range photogrammetry. Photogramm. Rec. 28, 86–95.
Chendeb, S., Fawaz, M., Guitteny, V., 2013. Calibration of a moving zoom-lens
Tang, R.F., Fritsch, D., Cramer, M., 2012. New rigorous and flexible Fourier self-
camera for augmented reality applications, industrial electronics (ISIE), 2013.
calibration models for airborne camera calibration. ISPRS J. Photogramm.
IEEE Int. Symp., 1–5
Remote Sens. 71, 76–85.
Chentt, Y.-S., Hung, Y.-P., Fuh, C.-S., Shih, S.-W., 2000. Camera calibration with a
Tao, X., Janabi-Sharifi, F., Cho, H., 2009. An active zooming strategy for variable field
motorized zoom lens. Int. Conf. Pattern Recogn., 4495–4498
of view and depth of field in vision-based microassembly. IEEE Trans. Automat.
Fraser, C., Al-Ajlouni, S., 2006. Zoom-dependent camera calibration in digital close-
Sci. Eng. 6, 504–513.
range photogrammetry. Photogramm. Eng. Remote Sens. 72, 1017–1026.
Tarabanis, K., Tsai, R.Y., Goodman, D.S., 1992. Modeling of a computer-controlled
Fraser, C., Cronk, S., Stamatopoulos, C., 2012. Implementation of zoom-dependent
zoom lens. IEEE Int. Conf. Robotics Automat., 1545–1551
camera calibration in close-range photogrammetry. ISPRS-Int. Arch.
Tarhan, M., Altug, E., 2011. A catadioptric and pan-tilt-zoom camera pair object
Photogramm., Remote Sens. Spatial Inf. Sci. 1, 15–19.
tracking system for UAVs. J. Intell. Robotic. Syst. 61, 119–134.
Fraser, C.S., 1997. Digital camera self-calibration. ISPRS J. Photogramm. Remote
Tatzgern, M., Grasset, R., Kalkofen, D., Schmalstieg, D., 2014. Transitional
Sens. 52, 149–159.
Augmented reality navigation for live captured scenes, virtual reality (VR),
Heikkila, J., Silven, O., 1997. A four-step camera calibration procedure with implicit
2014 iEEE. IEEE, 21–26.
image correction. Comput. Vis. Pattern Recogn., 1106–1112
Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W., 2000. Bundle
Hosoda, K., Moriyama, H., Asada, M., 1995. Visual servoing utilizing zoom
Adjustment—A Modern Synthesis, Vision Algorithms: Theory and Practice.
mechanism. Int. Conf. Robotics Automat., 178–183
Springer, pp. 298–372.
Kim, D., Oh, J., Sohn, K., Shin, H., 2010. Automatic radial distortion correction in
Tsai, R.Y., 1987. A versatile camera calibration technique for high-accuracy 3D
zoom lens video camera. J. Electron. Imaging 19, 043010-043010-043017.
machine vision metrology using off-the-shelf TV cameras and lenses. Robotics
Kim, S.W., Yun, K., Yi, K.M., Kim, S.J., Choi, J.Y., 2013. Detection of moving objects
Automat., IEEE J. 3, 323–344.
with a moving camera using non-panoramic background model. Mach. Vis.
Utsumi, Y., Sommerlade, E., Bellotto, N., Reid, I., 2012. Cognitive active vision for
Appl. 24, 1015–1028.
human identification, Robotics and Automation (ICRA), 2012 IEEE International
Läbe, T., Förstner, W., 2004. Geometric stability of low-cost digital consumer
Conference on. IEEE, pp. 1238–1245.
cameras. In: Proceedings of the 20th ISPRS Congress, Istanbul, Turkey, pp. 528–
Wakutsu, R., Chikatsu, H., 2011. Practical calibration for consumer grade digital
535.
camera with integrated high zooming lens, SPIE Optical Metrology.
Li, M., Lavest, J.-M., 1996. Some aspects of zoom lens camera calibration. IEEE Trans.
International Society for Optics and Photonics, pp. 80850T–80850T-80811.
Pattern Anal. Mach. Intell. 18, 1105–1110.
Warren, M., Mejias, L., Yang, X., Arain, B., Gonzalez, F., Upcroft, B., 2013. Enabling
Lichti, D.D., Kim, C., Jamtsho, S., 2010. An integrated bundle adjustment approach to
Aircraft Emergency Landings Using Active Visual Site Detection. Springer Tracts
range camera geometric self-calibration. ISPRS J. Photogramm. Remote Sens. 65,
in Advanced Robotics Springer International Publishing, pp. 167–181.
360–368.
Wiley, A.G., Wong, K.W., 1995. Geometric calibration of zoom lenses for computer
Lichti, D.D., Qi, X.J., Ahmed, T., 2012. Range camera self-calibration with scattering
vision metrology. Photogramm. Eng. Remote Sens. 61, 69–74.
compensation. ISPRS J. Photogramm. Remote Sens. 74, 101–109.
Willson, R.G., 1994. Modeling and calibration of automated zoom lenses, photonics
Nakano, K., Chikatsu, H., 2009. Exif dependent camera calibration in close range
for industrial applications. Int. Soc. Opt. Photon., 170–186
digital photogrammetry. J. Jpn. Soc. Photogramm. Remote Sens. 48, 180–187.
Wu, B., Hu, H., Zhu, Q., Zhang, Y., 2013. A flexible method for zoom lens calibration
Parian, J.A., Gruen, A., 2010. Sensor modeling, self-calibration and accuracy testing
and modeling using a planar checkerboard. Photogramm. Eng. Remote Sens. 79,
of panoramic cameras and laser scanners. ISPRS J. Photogramm. Remote Sens.
555–571.
65, 60–76.
Zhang, Y., Hu, K., Huang, R., 2012. Bundle adjustment with additional constraints
Remondino, F., Fraser, C., 2006. Digital camera calibration methods: considerations
applied to imagery of the Dunhuang wall paintings. ISPRS J. Photogramm.
and comparisons. Int. Arch. Photogramm., Remote Sens. Spatial Inf. Sci. 36, 266–
Remote Sens. 72, 113–120.
272.
Salvi, J., Armangué, X., Batlle, J., 2002. A comparative review of camera calibrating
methods with accuracy evaluation. Pattern Recogn. 35, 1617–1635.