You are on page 1of 9

Error estimation in measuring strain elds with DIC on planar sheet metal

specimens with a non-perpendicular camera alignment


P. Lava
a,
, S. Coppieters
a
, Y. Wang
b
, P. Van Houtte
b
, D. Debruyne
a,b
a
Catholic University College Ghent, Association K.U. Leuven, Department of Mechanical Engineering, Gebroeders Desmetstraat 1, B-9000 Gent, Belgium
b
Department MTM, Katholieke Universiteit Leuven, Kasteelpark Arenberg 44, B-3001 Leuven (Heverlee), Belgium
a r t i c l e i n f o
Article history:
Received 8 July 2010
Received in revised form
30 August 2010
Accepted 30 August 2010
Available online 16 September 2010
Keywords:
Digital image correlation
Stereo vision
Strain calculations
Systematic errors
Finite element simulations
a b s t r a c t
The determination of strain elds based on displacement components obtained via 2D-DIC is subject to
several errors that originate from various sources. In this contribution, we study the impact of a non-
perpendicular camera alignment to a planar sheet metal specimens surface. The errors are estimated in
a numerical experiment. To this purpose, deformed images that were obtained by imposing nite
element (FE) displacement elds on an undeformed image are numerically rotated for various Euler
angles. It is shown that a 3D-DIC stereo conguration induces a substantial compensation for the
introduced image-plane displacement gradients. However, higher strain accuracy and precision are
obtained up to the level of a perfect perpendicular alignment in a proposed rectied 2D-DIC setup.
This compensating technique gains benet from both 2D-DIC (single camera view, basic amount of
correlation runs, no cross-camera matching nor triangulation) and 3D-DIC (oblique angle compensa-
tion). Our conclusions are validated in a real experiment on SS304.
& 2010 Elsevier Ltd. All rights reserved.
1. Introduction
Digital image correlation (DIC), being a member of the class of
optical full-eld measurement methods, has undergone contin-
uous modication and improvement since the early eighties. As a
result, DIC experiments are currently extensively applied to study
e.g. the deformation characteristics of a wide range of materials in
tensile experiments on planar specimens [18]. One of the major
drawbacks of the technique, however, is the absence of a clear
error estimation on the obtained experimental results.
Recently, elaborate studies have been dedicated to the
validation of single camera DIC (designated 2D-DIC) displacement
and strain elds [921]. In theory, however, 2D-DIC is restricted
to planar specimens subject to in-plane deformations. In addition,
the single recording camera is assumed to be perpendicularly
aligned with the specimens surface. If not, relative out-of-plane
motion of the specimen with respect to the imaging system will
introduce image-plane displacement gradients, resulting in
corrupted displacement and strain eld data [22]. One possible
solution to accommodate for this kind of experimental noise is a
stereovision 3D-DIC system setup, employing two or more
cameras to record the specimen from various points of view.
The stereo system, however, involves multiple correlation runs,
including cross-camera subset matching. In addition, a stereo
system urges the input of calibration data. Accordingly, at the
eventual benet of avoiding errors due to deviations from
planarity, one introduces (other) extra error sources in the
obtained displacement and strain elds.
In the past, the effect of out-of-plane motion on 2D and 3D DIC
measurements was the subject of investigation in Refs. [14,22]. It
was shown that a single 2D imaging system is sensitive to out-of-
plane translation (DZ=Z, where Z is the distance from the object to
the pin hole and DZ the out-of-plane translation displacement)
and out-of-plane rotation (function of both rotation angle and
image distance Z). Accordingly, to minimize the effect of out-of-
plane motion on 2D-DIC measurements one can increase the
image distance or alternatively use a telecentric lens. In addition,
it was shown that a stereovision system measures all displace-
ment components with relatively high accuracy.
The above-mentioned papers, however, focussed on rigid body
deformations of the specimen. In Refs. [16,23], errors in 2D-DIC
displacement and strain elds were investigated in regions with
large heterogeneous deformations. It was shown that conscious
choices in the interpolation order, shape function, subset size,
strain-window size and strain-window interpolation order should
be made according to the spatial variation of the strain elds. In
view of the multiple correlation runs and cross-camera subset
matching in a 3D-DIC system, the accumulating effect of these
additional error sources may extinguish the benets of the out-of-
plane displacement compensation.
In this paper, we estimate the errors involved in a 3D-DIC stereo
setup using a similar procedure as outlined in Refs. [16,23], focussing
on a realistic uni-axial tensile test on a perforated tensile sheet metal
specimen subject to large in-plane plastic deformation. A reference
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/optlaseng
Optics and Lasers in Engineering
0143-8166/$ - see front matter & 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.optlaseng.2010.08.017

Corresponding author. Tel.: +3292658722; fax: +3292658648.


E-mail address: pascal.lava@kahosl.be (P. Lava).
Optics and Lasers in Engineering 49 (2011) 5765
image is numerically deformed by imposing nite element (FE)
displacement elds corresponding to various load steps. These are
obtained by the commercial software package Abaqus [24]. Next, the
reference image and numerically deformed images corresponding to
large plastic strains are numerically rotated for various Euler angles.
The unrotated images are considered as taken by camera 1, the
rotated as captured by camera 2. An identical rotation procedure is
adopted to a calibration pattern, allowing the determination of the
specic camera parameters. Accordingly, we can validate our strain
predictions obtained in a simulated perpendicular (camera 1), non-
perpendicular (camera 2) and stereo setup on a planar specimen at
high plastic deformation by comparing them to the well-known FE
strain elds at the Gauss points. All DIC results are computed with
our recently developed inhouse DIC platform MatchID [25].
It is important to remark that our validation procedure is
numerical by nature. As such, we are not sensitive to differences
in lighting conditions and, moreover, we can avoid out-of-plane
rotations/translations in the perpendicular setup due to (a)
Poissons effect (b) small amounts of specimen bending (c) local
necking during the loading process and (d) deviations from ideal
grip constraints. We are well aware that in a realistic experiment
these effects are hard to avoid, but can be minimized when
appropriate experimental conditions are fullled.
Finally, we propose a compensating technique which gains benet
from both 2D-DIC (single camera view, basic amount of correlation
runs, no cross-camera subset matching and triangulation) and 3D-DIC
(out-of-plane displacement compensation) implementations. In this
so-called rectied 2D-DIC setup we rst rectify the non-perpendi-
cular images with the obtained camera parameters, simulating a
perpendicular alignment in a single camera setup. Next, this new
procedure is veried by a real experiment on SS304.
The outline of this article is as follows. In Section 2 we present
the formalism for the numerical deformation and out-of-plane
rotation of the reference images. In Section 3 the 3D-DIC system is
described, whereas in Section 4 attention is paid to the procedure
of rectication. Our results, both numerical and experimental are
included in Section 5. We conclude in Section 6.
2. Generation of synthetic images
In order to construct our numerical images, a similar
procedure is adopted as in Refs. [16,23]. First, the deformed
images are obtained by imposing nite element (FE) displacement
elds on an undeformed image yielding plastic deformation of the
specimen. In particular, Abaqus [24] was used to simulate an uni-
axial tensile test on a perforated tensile specimen. The specimen
has a thickness of 0.8mm in the simulation and assumes a von
Mises yield criterion in combination with an isotropic hardening
behaviour that was derived from tests on SS 304 sheet metal:
s 15330:033e
0:48
. The mesh that was generated to perform
the calculations is shown in the centre panel of Fig. 1, with the
arrow denoting the direction of the applied force. The FE
calculation yields well-known displacement elds at the nodal
points and strain values at the Gauss points of the nite elements.
The average generated elements have a size of approximately
33 pixels. Next, the nodal displacements are used to numeri-
cally deform a random speckle pattern displayed in the left panel
of Fig. 1, representing an undeformed state. The displacement
elds at integer pixel locations are determined through bicubic
interpolation within the corresponding nite element. The right
panel of this gure shows the resulting numerically deformed
image corresponding to a maximum equivalent plastic strain
of 0.21.
By construction, the reference image and the numerically
deformed images correspond to a camera point of view with
sensor plane parallel to the object plane. In order to investigate
the impact of a non-perpendicular alignment of the camera to the
object plane, we need to numerically rotate the reference and
deformed images. A 3D rotation around the origin yields a
rotation matrix
R
cosccosf coscsinfsinysinccosy coscsinfcosysincsiny
sinccosf sincsinfsinycosccosy sincsinfcosycoscsiny
sinf cosfsiny cosfcosy
_

_
_

_,
1
where y,f,c represent the Euler angles around the X-, Y- and
Z-axis, respectively (see Fig. 2). The rotated image coordinates can
then be obtained as
x
rot
im
y
rot
im
_ _

R
11
x
im
R
12
y
im
R
31
x
im
R
32
y
im
R
21
x
im
R
22
y
im
R
31
x
im
R
32
y
im
_

_
_

x
im
y
im
_ _

q
r
_ _
: 2
A similar procedure can now be adopted as for the numerical
deformation case. Indeed, as depicted in Eq. (2) the rotated image
coordinates (x
im
rot
,y
im
rot
) yield well-known displacement elds q and r
with respect to the reference coordinates. Again, the resulting
greyscale values at integer pixel locations are determined through
bicubic interpolation. Since the spatial variation of strain in one
element is linear, one can expect that the additional errors due to
this interpolation procedure are minor. In Fig. 2 the rotated
reference pattern is shown corresponding to a rotation around the
X- and Y-axis.
Fig. 1. Reference speckle pattern for the uni-axial tensile test with a subset of
1919 pixels (left), mesh pattern with the applied force (middle) and speckle
pattern with nite element deformation (right panel) corresponding to e
pl
eq
0:21.
Fig. 2. Euler reference rotation frame (left), rotated reference speckle pattern
corresponding to fy (middle) and yx (right) angle of 301.
P. Lava et al. / Optics and Lasers in Engineering 49 (2011) 5765 58
3. 3D digital image correlation
3.1. Formalism
In this section we give a brief overview of 3D-DIC principles
relevant to our discussion. A more elaborate discussion can be
found in e.g. Refs. [26,27]. Stereo-vision DIC involves a funda-
mental extension of 2D-DIC correlation concepts with stereovi-
sion principles. Indeed, after synchronization of the image
acquisition two cameras capture N pairs of images simultaneously
during the loading process. Next, the measurement of 3D
displacements proceeds as depicted in Fig. 3.
(1) A subset (x
0
, y
0
) is selected in the reference image of camera 1.
Next, 2D-DIC cross-camera correlation yields a displacement
vector (q,r) and matching position in the reference image of
camera 2. Thanks to a geometric constraint inherent to any
calibrated stereo imaging system, one can reduce this 2D
search to a 1D search along the corresponding epipolar line,
simplifying the search routine and saving computer time. This
process is repeated for a dense map of subsets in the reference
image. The determined set of corresponding points and the
input of stereo camera parameters results in the initial object
shape with 3D coordinates (X, Y, Z) via a classical triangulation
method [28], a process usually referred to as reconstruction.
(2) The same set of subsets in the reference image of camera 1 is
used to perform a standard 2D-DIC correlation to locate the
matching positions (x
0
+u
0
, y
0
+v
0
) in the deformed image of
camera 1.
(3) The same set of subsets in the reference image of camera 1 is
used to perform a cross-camera correlation to identify the
matching positions (x
0
+u
1
, y
0
+v
1
) in the deformed image of
camera 2.
(4) Again, the obtained set of corresponding points for the
deformed images and the input of stereo camera parameters
results in the deformed 3D conguration Xu,Yu,Zu.
(5) The 3D displacement eld (U, V, W) corresponding to the
specic loading step can now be directly obtained by
subtracting Xu,Yu,Zu and (X, Y, Z).
The above-mentioned steps 25 are repeated for each set of
deformed images, yielding a complete track of the deformation
process in three dimensions.
An alternative common adopted approach is to perform only
one cross-camera matching run (step 1) and to update the
reference images. This procedure, however, involves cumulative
errors that are not benecial to the displacement uncertainty [29].
The DIC sub-pixel displacement elds contain all the informa-
tion needed to calculate the in-plane normal strain and shear
components e
xx
, e
yy
and e
xy
. This calculation involves the
determination of the so-called deformation gradient matrix F,
describing the relative spatial position of two neighbouring
particles after deformation in terms of their relative material
position before deformation. Expressing the deformation gradient
matrix in terms of the horizontal (x-direction) and vertical
(y-direction) displacement components u and v yields
F
1
@u
@x
@u
@y
@u
@z
@v
@x
1
@v
@y
@v
@z
@w
@x
@w
@y
1
@w
@z
_

_
_

_
3
The determination of the strain tensor in e.g. the logarithmic
EulerAlmansi convention e
lnEA
is now straightforward:
e
ln EA
lnV ln

FF
T
_ _ _
, 4
with V the so-called left stretch tensor obtained via the Cauchy
theorem of polar decomposition. The use of this tensor is common
in metal plasticity since tension, compression and torsion results
nearly coincide when logarithmic strain is plotted against the true
stress [24].
As explained in Ref. [16], in order to improve the strain
calculation with increased accuracy, one rst smooths the estimated
displacement elds before starting the differentiation process by
nding an analytical expression of a surface, which, in a least
squares sense, approximates the experimental displacement values
in a selected region. This is a so-called strainwindow containing
NN nearest neighbours for each component of displacement.
3.2. Stereovision calibration
As discussed above, stereovision systems urge the input of
camera calibration parameters. These are determined via a bundle
adjustment technique as outlined in Refs. [26,30], using various
images of a translated and rotated planar regular grid pattern.
Within a LevenbergMarquardt optimization routine, the differ-
ences between the measured sensor coordinates and the model
predictions of these locations are minimized. As such, one arrives
at the following set of intrinsic and extrinsic parameters:
(c
x
,c
y
)[pixels]: image plane centre location for each camera;
(f
x
,f
y
) [pixels]: focal lengths in horizontal and vertical directions
for each camera;
f
s
[pixels]: accounting for skewing of sensor array;
k
1
: radial distortion coefcient;
(T
x
,T
y
,T
z
): translation components to transform origin of
camera 1 to origin of camera 2; and
y,f,c: relative orientation of camera 2 with respect
to camera 1, representing a rotation around the X-, Y- and
Z-axis, respectively.
Determining the camera parameters, however, is a numerical
cumbersome task, with various error sources. Notwithstanding, an
accurate camera parameter determination is of crucial importance
in viewof precise 3D-DIC displacement elds. It is not the purpose of
this contribution, however, to perform a scrutinized estimation of
errors involved with the calibration process.
Fig. 3. Schematic view of the process of 3D-DIC.
P. Lava et al. / Optics and Lasers in Engineering 49 (2011) 5765 59
3.3. Error estimation in cross-camera matching
In Section 5 the overall errors for the 3D-DIC system will be
studied. Before embarking on this, we should rst clearly
understand the roots of these errors. Indeed, as can be inferred
from Fig. 3, a 3D-DIC stereo system involves various correlation
runs and reconstruction, each of them with their specic errors
contributing to the overall accuracy and precision of the
measurement. Errors due to reconstruction in image stereo-
correlation were the subject of investigation in Ref. [31]. Error
estimation (process 2) in regular single camera 2D-DIC was
already the subject of our investigation in Refs. [16,23]. It was
shown that the impact of the subset size on the strain precision
persists, even when a smoothing procedure as outlined above is
implemented. Cross-camera 2D-DIC matching (process 1), how-
ever, needs some further investigation. Due to the numerical
rotation process of the images, we are provided with well-known
displacement values (q,r) in Eq. (2), allowing us to estimate the
errors of the displacement results obtained by 2D-DIC calcula-
tions. In this section we are only interested in errors due to the
cross-camera matching. As such, a full 2D search across the entire
image is performed instead of a 1D search along the correspond-
ing epipolar line, since this latter would introduce errors due to
the calibration procedure. In the overall 3D DIC mechanism,
however, this epipolar constraint is incorporated.
A similar procedure is followed as outlined in Ref. [16] in
which we calculate the mean Dq
av
and the standard deviation s of
the measured differences Dq q
imp
q
DIC
. The following standard
settings are adopted: r
ASSD
, afne transformation, bicubic inter-
polation and a subset size of 2121. The errors on the vertical
displacement components r and for rotations around the X-axis
are not shown since they are of the same order, yielding similar
conclusions concerning the impact of the parameter under
investigation.
First, in Fig. 4, we study the impact of the adopted DIC
interpolation order on the cross-camera matching. As stated in
Ref. [11], an accurate choice of this quantity is of the utmost
importance in the DIC algorithm since it introduces a large
reduction in the bias of displacement elds when going from
bilinear to bicubic interpolation order. As can be inferred, this
conclusion is conrmed for stereo angles ranging from 5 up to 401,
with the bicubic and bicubic spline routines yielding the optimum
results.
Next, in Fig. 5 the mean error and corresponding standard
deviations are displayed for different dimensions of the subset as
a function of the rotation angle with respect to the Y-axis. As can
be inferred, the accuracy and precision are relatively high when
small subset sizes in combination with stereo angles below 301
are considered. Larger subset sizes and stereo angles involve
questionable matching results. This is in agreement with the
results of Ref. [26]. The choice for a small subset size, however, is a
venture into dangerous territory since its lower limit is estab-
lished by the granularity of the speckle pattern. As such, one can
wonder if it is not possible to keep the subset size xed to a
reasonable large value combined with an alternate shape
function. In Ref. [26], one considers a homographic shape
function, which reads as
mx,y,s
nx,y,s
_ _

s
0
xs
1
ys
2
s
6
xs
7
y1
s
3
xs
4
ys
5
s
6
xs
7
y1
_

_
_

_
: 5
This a more natural approach to describe cross-camera displace-
ments in view of the perspective projection of planar objects. In
Eq. (5), (x,y) and m,n represent the pixel positions in the image
0
0.02
0.04
0.06
0.08
0.1

q
a
v
BC
BCSPLINE
BL
(deg)

0
0.025
0.05
0.075
0.1
0 5 10 15 20 25 30 35 40
Fig. 4. Impact of the interpolation order on the mean systematic errors (upper
panels) and standard deviation (lower panels) for cross-camera horizontal
displacements as a function of out-of-plane rotation around the Y-axis. The solid,
dashed and dot-dashed lines adopt a bicubic (BC), bicubic spline (BCSPLINE) and
bilinear (BL) interpolation routine, respectively.
0
0.02
0.04
0.06
0.08
0.1

q
a
v
11
21
31
41
(deg)

0
0.025
0.05
0.075
0.1
0 5 10 15 20 25 30 35 40
Fig. 5. Impact of the subset size on the mean systematic errors (upper panels) and
standard deviation (lower panels) for cross-camera horizontal displacements as a
function of out-of-plane rotation around the Y-axis. The solid, dashed, dot-dashed,
and dotted lines represent a subset size of 11 11, 2121, 3131 and 4141,
respectively.
0
0.025
0.05
0.075
0.1

q
a
v
Affine
Irregular
Quadratic
(deg)

0
0.025
0.05
0.075
0.1
0 5 10 15 20 25 30 35 40
Fig. 6. Impact of the subset shape function on the mean systematic errors (upper
panels) and standard deviation (lower panels) for cross-camera horizontal
displacements as a function of out-of-plane rotation around the Y-axis. The solid,
dashed and dot-dashed lines represent the results of afne, irregular and quadratic
transformations, respectively.
P. Lava et al. / Optics and Lasers in Engineering 49 (2011) 5765 60
-0.001
0
0.001
0.002
0.003
0.004
0
0.005
0.01
0.015
-0.001
0
0.001
0.002
0.003
0
0.005
0.01
0.015
0.02
0
0.005
0.01
0.015

x
x

y
y
2D Perp
2D NonPerp
3D Stereo
2D Rect

x
y
(deg)
|Mean|
0
0.005
0.01
10 0 20 30 40 10 0 20 30 40
Fig. 7. Impact of an oblique angle observation on the mean error and corresponding standard deviations for various Euler pitch angles at a maximum equivalent plastic
strain of e
pl
eq
0:21. The solid, dashed, dot-dashed and dotted lines correspond to a 2D perpendicular setup, a 2D non-perpendicular setup, a 3D stereo setup and a 2D
rectied setup, respectively.
0
0.0025
0.005
0.0075
0.01
0
0.002
0.004
0.006
0.008
0.01
0
0.0025
0.005
0.0075
0
0.005
0.01
0.015
0
0.0025
0.005
0.0075
3D STEREO

x
x

y
y

x
y
X (mm)
|Mean|
0
0.002
0.004
0.006
0.008
0
0.0025
0.005
0.0075
0.01
0
0.002
0.004
0.006
0.008
0.01
0
0.0025
0.005
0.0075
0
0.005
0.01
0.015
0
0.0025
0.005
0.0075
2D RECT
= 5
= 15
= 30
= 40
X (mm)
|Mean|
0
0.002
0.004
0.006
0.008
-20

x
x

y
y

x
y
-10 0 10 20 -20 -10 0 10
-20 -10 0 10 20 -20 -10 0 10
Fig. 8. Mean error and standard deviations for a stereo 3D-DIC (a) and 2D-DIC rectied (b) setup at a maximum equivalent plastic strain of e
pl
eq
0:21 as a function of the
distance to the rotation axis. The solid, dashed, dot-dashed and dotted lines correspond to various Euler pitch angles.
P. Lava et al. / Optics and Lasers in Engineering 49 (2011) 5765 61
frame of camera 1 and 2, respectively, whereas s is the parameter
vector to be identied. A homographic shape function was tested
in our approach, but in accordance with Ref. [26], convergence
problems appeared during the matching process. A valid alter-
native is to use rectied images for which the epipolar lines are
horizontal and aligned. Unfortunately, this involves additional
interpolation, at the expect of additional positional errors. In this
contribution, we chose to alter the order of the subset shape
function. In Fig. 6, the impact of an irregular and quadratic shape
function on the mean errors and standard deviations is shown.
These transformation orders were dened in Ref. [16]. As can be
inferred, the introduction of quadratic terms substantially
improves the cross-camera matching process. This is not surpris-
ing since a rst-order Taylor expansion of the above-mentioned
more natural homography produces quadratic terms. Moreover,
the quadratic shape function does not suffer from convergence
problems.
To conclude, in this section it was shown that 3D-DIC is much
more complex compared to standard 2D-DIC. Accordingly, one
can already expect that the overall errors for the 3D-DIC system
will be larger. Before we embark on this, however, we rst
concentrate on a technique that might be a viable alternative to
standard 2D-DIC when the perpendicular condition is not fullled.
4. Image rectication
It is clear that a 3D-DIC stereosystem as depicted in Section 3.1
suffers from multiple error sources due to the accumulation of
various 2D-DIC correlation runs (cross-camera and deformation
tracking) and the camera calibration procedure. In Section 5 the
overall errors for the 3D-DIC system will be studied. First,
however, attention is paid to a technique that gains benet from
both 2D-DIC (single camera view, basic amount of correlation
runs and no cross-camera subset matching) and 3D-DIC (out-of-
plane displacement compensation) implementations. This image
rectication procedure transforms the non-perpendicular
images to a camera point of view perpendicular aligned to the
specimen plane and includes the following steps:
(1) As for the 3D-DIC stereo system a clear prerequisite for the
rectication process is a precise determination of the camera
parameters. As described in Section 3.2, we use a regular grid
pattern translated and rotated into various positions. For the rst
calibration image, we put the regular planar grid pattern parallel
to the specimens surface and at the same position of the
specimen. As such, we can consider the world reference frame as
the frame parallel to the object plane frame. Accordingly, we
arrive at six extrinsic parameters (R, T) describing the relative
position of the camera frame to the (parallel) world reference
frame and the intrinsic parameters f
x
,f
y
,f
s
,c
x
,c
y
,k
1
.
(2) First, the real distorted images are transformed into ideal
distortion-free images.
(3) Next, the four corners of the specimens undistorted images
are converted from sensor (x
s
u
, y
s
u
) to world coordinates (X
W
,
Y
W
, Z
W
) on the basis of the following relationships (Z
W
0):
x
u
s
y
u
s
_ _

c
x
f
x
R
11
X
W
R
12
Y
W
T
x
R
31
X
W
R
32
Y
W
T
z
f
s
R
21
X
W
R
22
Y
W
T
y
R
31
X
W
R
32
Y
W
T
z
c
y
f
y
R
21
X
W
R
22
Y
W
T
y
R
31
X
W
R
32
Y
W
T
z
_

_
_

_
:
6
(4) Finally, for all other points the rectication is implemented
backwards, i.e. starting from the new image plane applying
the inverse transformations of the four corner points. As such
the pixel values in the new image plane can be computed as a
bicubic interpolation of the integer pixel values in the old
image plane.
We remark that the rectied image is in general not contained
in the same region of the image plane as the original image.
Accordingly, one has to adjust the focal lengths of the camera for
the rectication process.
5. Results
5.1. Numerical
Now that we are provided with various DIC frameworks to
measure displacement and strain elds on planar specimens
observed from an oblique angle, we can embark on the validation
procedure via the numerical experiment as described in Section 2.
Since camera 1 and 2 correspond to the unrotated and rotated
images, respectively, the following numerical achievements are
obtained: 2D-DIC perpendicular setup via camera 1, a 2D-DIC
oblique angle setup via camera 2, a 3D-DIC stereo setup by
combining camera 1 and 2, and a 2D-DIC rectied setup by
rectifying the images of camera 2. As stated above, the stereo
system and the rectication procedure require the input of
camera parameters, both intrinsic and extrinsic. To this purpose,
we perform an identical numerical rotation procedure as
described in Section 2 on a calibration image of a regular grid
0
0.0025
0.005
0.0075
0.01
0
0.002
0.004
0.006
0.008
-0.08
0
0.0025
0.005
0.0075
0.01

x
x
= 30
2D Perp
2D Rect
3D

0
0.002
0.004
0.006
0.008
-0.07 -0.06 -0.05 -0.04 -0.03 -0.02 -0.01 0
-0.08 -0.07 -0.06 -0.05 -0.04 -0.03 -0.02 -0.01 0

x
x

xx

xx

= 5
Fig. 9. Mean error and standard deviation for the transverse strain component at
an Euler pitch angle of f5
3
(a) and f30
3
(b). Results are binned over
equidistant strain intervals. The circles, triangles and squares correspond to a
perpendicular 2D-DIC, a 3D-DIC and a rectied 2D-DIC setup, respectively.
P. Lava et al. / Optics and Lasers in Engineering 49 (2011) 5765 62
pattern. Next, the ctive camera parameters are estimated in a
bundle-adjustment calibration procedure, with the Euler angles
xed to the imposed values. As such, the object distance for all
forthcoming numerical results corresponds to Z 612mm. All
results presented below adopt the following standard settings:
r
ASSD
, afne transformation, bicubic interpolation, a subset size of
2121, a step size of 2 and a strain window of 10 10 with
bilinear interpolation.
Fig. 7 displays the impact of an oblique angle observation on
the mean error and standard deviation of the measured
differences De e
imp
e
DIC
for various Y-axis rotation angles at a
maximum equivalent plastic strain of e
pl
eq
0:21. The errors for
rotations around the X-axis are of the same order, yielding similar
conclusions concerning the framework under consideration. We
emphasize that our main focus is on sheet metal applications
where thickness reduction is of minor importance compared to
the object distance. As can be inferred, the errors related to the
2D-DIC setup subject to oblique angle observation gain in
importance when larger Euler angles are probed. The stereo
3D-DIC accuracy and precision are much better and nearly
independent of the rotation angle since this framework largely
compensates the involved oblique angle. The improvements,
however, do not bring the errors down to the level of the
perpendicular 2D-DIC setup. By rectifying the images captured by
the non-perpendicular camera a priori to a 2D-DIC correlation
run, a substantial higher accuracy and precision is achieved,
nearly reproducing the results of the perpendicular setup. Since
both methods are provided with the same set of camera
calibration parameters, these differences can be mainly attributed
to an accumulation of error sources inherent in the 3D-DIC
system, as e.g. cross-camera matching, camera calibration, multi-
ple correlation runs and triangulation.
The accuracy and precision results presented in Fig. 7 are
obtained by including all measured points in the entire image
with an equal weight. The impact of the oblique angle, however,
largely depends on the distance to the rotation axis. As such, in
Fig. 8 we present the mean error and standard deviations for a
stereo 3D-DIC and 2D-DIC rectied setup at a maximum
equivalent plastic strain of e
pl
eq
0:21 as a function of the distance
to the rotation axis, with Euler pitch angles ranging from 51 up to
401. As can be inferred, the angle variation has nearly no impact
on the accuracy and precision of both methods, nor for points
close or at a larger distance to the rotation axis. This conrms that
3D-DIC and the rectied 2D-DIC accurately account for oblique
angle observations. The overall error for 3D-DIC, however,
overshoots the ones of the rectication method for the entire
distance range. As already stated, these deviations are due to the
complexity of the 3D-DIC system.
The principle goal of this paper is to study the impact of an
oblique angle observation when the specimen is subject to
substantial plastic deformation. Since the results presented above
reveal that both rectied 2D-DIC and 3D-DIC accurately account
for various out-of-plane rotations, we will further on concentrate
on two xed Euler pitch angles: a realistic f5
3
and 301, which is
often recognized as the optimum stereo angle. In order to be able
to present the strain accuracy and precision as a function of the
strain value, the following binning procedure is adopted. First,
the maximum and minimum strain values are determined. Next,
0
0.005
0.01
0.015

x
y
= 5

xy

0
0.002
0.004
0.006
0.008
-0.02
0
0.005
0.01
0.015
2D Perp
2D Rect
3D
= 30
0
0.002
0.004
0.006
0.008
-0.015 -0.01 -0.005 0 0.005 0.01 0.015 0.02
-0.02 -0.015 -0.01 -0.005 0 0.005 0.01 0.015 0.02

x
y

xy
Fig. 11. Same as in Fig. 9 but for the shear strain component.
0
0.005
0.01
0.015
0.02
= 5
0
0.005
0.01
0.015
0
0.005
0.01
0.015
0.02

y
y2D Perp
2D Rect
3D

0
0.005
0.01
0.015
0

y
y

= 30
0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2

yy

yy
Fig. 10. Same as in Fig. 9 but for the longitudinal strain component.
P. Lava et al. / Optics and Lasers in Engineering 49 (2011) 5765 63
we subdivide the strain values into equidistant intervals. Finally,
we calculate the mean and standard deviation of each strain
interval.
In Figs. 911 we present the results for the transverse e
xx
, the
longitudinal e
yy
and the shear e
xy
strain components obtained
with the 2D-DIC, the rectied 2D-DIC and the 3D-DIC model. All
three models have the following expected common behaviour: a
decreasing accuracy and precision when higher strain values are
probed. On the other hand, at a certain lower strain threshold of
e 0:01, the uncertainties start to gain in importance. Indeed, it is
well-known that measuring small strain values with DIC is a
cumbersome task. It is clear that for all three strain components
the 2D perpendicular setup yields the utmost accurate results,
closely followed by the predictions of the rectied 2D-DIC setup.
The 3D-DIC model yields comparable results at modest strain
values. The increasing error behaviour, however, is much more
outspoken compared to the 2D-DIC alternatives. This can again be
mainly attributed to the cumulative effect of multiple correlation
runs. As can be inferred, the errors of 3D-DIC are more or less a
factor of three larger compared to the ones of a 2D-DIC setup,
exactly the ratio of correlation runs appearing in 3D-DIC versus
2D-DIC. These conclusions were conrmed by the commercial
software Vic3d [32], yielding similar 3D-DIC results. Finally, it is
important to remark that the strain values are consistent. Indeed,
equal strain values in e
xx
,e
yy
,e
xy
yield almost equal accuracy and
precision values. As such, the strain errors are inherent to the
strain value, and not to the strain component considered.
Fig. 12. Oblique angle view of the uni-axial tensile test on a SS304 sheet metal specimen with a central hole (a). Panels (b), (c) and (d) display the exx,eyy and exy strain
component, respectively, obtained with rectied images in a 2D-DIC framework.
Table 1
Calibration parameters for the stereovision system.
Parameter Camera 1 Camera 2 Transformation
c
x
(pixels) 666.02709 675.9223 y (deg) 0.52678
c
y
(pixels) 584.9601 616.7829 f (deg) 17.2755
f
x
(pixels) 3438.1036 3430.8922 c (deg) 2.2669
f
y
(pixels) 3438.1036 3430.8922 T
x
(mm) 109.8739
f
s
(pixels) 1.3712 1.36264 T
y
(mm) 0.13709
k
1
0.017 0.012 T
z
(mm) 38.9439
P. Lava et al. / Optics and Lasers in Engineering 49 (2011) 5765 64
5.2. Experimental
In this last part, a real experimental case is studied on stainless
steel SS304. An uni-axial tensile test is performed on sheet
material with a thickness of 0.8mm. A perforated tensile speci-
men is loaded up to 15.9kN, inducing strain values well beyond
yielding. The experiment is tracked by two CCD cameras (Retiga
1300) with a resolution of 12801024 pixels. The cameras are
positioned so that camera 1 has an approximate perpendicular
alignment to the specimens surface and camera 2 has an angled
view, as shown in Fig. 12. As such, we can evaluate the
experiment with a 2D-DIC perpendicular, 2D-DIC rectied an
3D-DIC method. The camera parameters of the experiment are
displayed in Table 1. These were obtained by capturing 23 rotated
and translated positions of the grid pattern. The rst grid pattern
is put parallel to and at the same position of the specimens
surface, in view of retrieving the extrinsic parameters needed for
the 2D-DIC rectied method as described in Section 4. The overall
calibration is accurate, with a standard deviation of residuals in
the pixel positions for all images of 0.027, indicating the
calibration is adequate for stereovision measurements. In Fig. 12
the strain components are displayed at a maximum equivalent
plastic strain of 0.15 obtained via a 2D-DIC rectication method.
Uncertainties in the material model parameters make it cumber-
some to consider the FE values as baseline results. As such, we
take the 2D-DIC perpendicular strain values of camera 1 as a
reference. The deviations of the rectied 2D-DIC and the 3D-DIC
setup are displayed in Table 2. As can be inferred, the above-
mentioned numerical conclusions are conrmed by the experi-
ment, with the 2D-DIC rectied calculations yielding a higher
precision and accuracy for all three strain components. This
indicates that the 2D-DIC rectied setup can be a valid alternative
in the accurate determination of strain elds on planar sheet
metal specimens with an oblique angle observation.
6. Conclusions
In this contribution we studied the impact of a non-
perpendicular alignment of the camera with a planar sheet metal
specimen subject to large plastic deformation. Errors of 2D-DIC
and 3D-DIC are estimated in both a numerical and real experi-
ment. It is shown that the 3D-DIC stereo setup largely compen-
sates oblique angle observation. The accuracy and precision,
however, are not of the same order as the ones of a perpendicular
2D-DIC setup. As such, in measuring strain elds on planar
specimen major attention should be paid to the perpendicular
alignment of the camera to the specimens surface. If this
condition is hard to fulll, a rectied 2D-DIC method could be
a valid alternative. This compensating technique gains benet
from both 2D-DIC (single camera view, basic amount of correla-
tion runs and no cross-camera subset matching) and 3D-DIC (out-
of-plane displacement compensation).
References
[1] Cooreman S, Lecompte D, Sol H, Vantomme J, Debruyne D. Exp Mech
2008;48:42133.
[2] Lecompte D, Cooreman S, Coppieters S, Vantomme J, Sol H, Debruyne D. Eur
Comput Mech 2009;18:393418.
[3] Lecompte D, Smits A, Sol H, Vantomme J, Van Hemelrijck D. Int J Solids Struct
2007;44:164356.
[4] Kajberg J, Lindkvist G. Int J Solids Struct 2004;41:343959.
[5] Rossi M, Broggiato GB, Papalini S. Meccanica 2008;43:18599.
[6] Van Paepegem W, Shulev AA, Roussev IR, De Pauw S, Degrieck J, Sainov VC.
Opt Lasers Eng 2009;47:3907.
[7] Willems A, Lomov SV, Verpoest I, Vandepitte D. Opt Lasers Eng 2009;47:
34351.
[8] Ivanov D, Ivanov S, Lomov S, Verpoest I. Opt Lasers Eng 2009;47:36070.
[9] Bruck HA, McNeill SR, Sutton MA, Peters WH. Exp Mech 1989;29:2617.
[10] Vendroux G, Knauss WG. Exp Mech 1998;38:8692.
[11] Schreier HW, Braasch JR, Sutton MA. Opt Eng 2000;39:291521.
[12] Schreier HW, Sutton MA. Exp Mech 2002;42:30310.
[13] Pan B, Xie H-m, Xu B-q, Dai F-l. Meas Sci Technol 2006;17:161521.
[14] Haddadi H, Belhabib S. Opt Lasers Eng 2008;46:18596.
[15] Pan B, Xian K-m, Xie H-m, Asundi A. Meas Sci Technol 2009;20:062001.
[16] Lava P, Cooreman S, Coppieters S, De Strycker M, Debruyne D. Opt Lasers Eng
2009;47:74753.
[17] Bornert M, et al. Exp Mech 2009;49:35370.
[18] Triconnet K, Derrien K, Hild F, Baptiste D. Opt Lasers Eng 2009;47:72837.
[19] Cofaru C, Philips W, Van Paepegem W. Meas Sci Technol 2010;21:055102.
[20] Reu PL, Miller TJ. SEM annual conference, Indianapolis, 2010.
[21] Fazzini M, Mistou S, Dalverny O, Robert L. Opt Lasers Eng 2010;48:3359.
[22] Sutton MA, Yan JH, Tiwari V, Schreier HW, Orteu JJ. Opt Lasers Eng
2008;46:74657.
[23] Lava P, Cooreman S, Debruyne D. Opt Lasers Eng 2010;48:45768.
[24] Abaqus. Theory manual, Version 6.7, 2007.
[25] /http://www.matchid.orgS.
[26] Sutton MA, Orteu JJ, Schreier HW. Image correlation for shape, motion and
deformation measurements. New York: Springer; 2009.
[27] Orteu JJ. Opt Lasers Eng 2009;47:28291.
[28] Hartley R, Sturm P. Comput Vision Image Understanding 1997;68:14657.
[29] Hild F, Raka B, Baudequin M, Roux S, Cantelaube F. Appl Opt 2002;41:
681528.
[30] Triggs B, Mclauchlan P, Hartley R, Fitzgibbon A. Bundle adjustmenta
modern synthesis. In: Triggs B, Zisserman A, Szeliski R, editors. Vision
algorithms. Berlin: Springer; 2000.
[31] Fazzini M, Mistou S, Dalverny O. In: 14th international conference on
experimental mechanics, vol. 6, 2010. id.31009.
[32] Vic3D digital image correlation program. Correlated Solutions, Inc. /http://
www.correlatedsolutions.comS.
Table 2
Mean error and standard deviations for the three strain components in a rectied 2D-DIC and 3D-DIC setup.
Model jDexxj sexx jDeyyj seyy jDexyj sexy
Rectied 2D 0.000186 0.00089 0.000162 0.000918 3.5638E05 0.000999
Stereo 3D 0.000212 0.00166 0.000221 0.00185 0.00265 0.00588
The results of a perpendicular 2D-DIC measurement are considered as baseline ones.
P. Lava et al. / Optics and Lasers in Engineering 49 (2011) 5765 65

You might also like