727

ARTICLE
Improved image-based deformation measurement for
geotechnical applications
S.A. Stanier, J. Blaber, W.A. Take, and D.J. White

Abstract: This paper describes and benchmarks a new implementation of image-based deformation measurement for geotech-
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18

nical applications. The updated approach combines a range of advances in image analysis algorithms and techniques best suited
to geotechnical applications. Performance benchmarking of the new approach has used a series of artificial images subjected to
prescribed spatially varying displacement fields. An improvement by at least a factor of 10 in measurement precision is achieved
relative to the most commonly used particle image velocimetry (PIV) approach for all deformation modes, including rigid-body
displacements, rotations, and strains (compressive and shear). Lastly, an example analysis of a centrifuge model test is used to
demonstrate the capabilities of the new approach. The strain field generated by penetration of a flat footing and an entrapped
sand plug into an underlying clay layer is computed and compared for both the current and updated algorithms. This analysis
demonstrates that the enhanced measurement precision improves the clarity of the interpretation.

Key words: image analysis, model tests, particle image velocimetry, digital image correlation.

Résumé : Le présent article décrit et évalue la mise en œuvre récente d’une nouvelle méthode de mesure de déformation basée
sur des images et applicable au domaine de la l’ingénierie géotechnique. Cette nouvelle méthode combine une série
d’améliorations apportées aux algorithmes et techniques d’analyse d’images spécialement adaptés au domaine de l’ingénierie
géotechnique. L’évaluation des performances de cette nouvelle méthode a été réalisée à l’aide d’une série d’images artificielles
For personal use only.

soumises à des champs de déplacement définis et variant dans l’espace. Une amélioration de la précision des mesures d’un
facteur au moins égal à dix est obtenue si l’on utilise la nouvelle méthode plutôt que la méthode la plus souvent employée de la
vélocimétrie par images de particules (VIP), et ce pour tous les modes de déformation, y compris les déplacements, rotations et
déformations (par compression ou par cisaillement) de corps rigides. Enfin, on se sert d’un exemple d’analyse d’un essai sur
modèle en centrifugeuse pour montrer ce que permet de réaliser la nouvelle méthode. Le champ de déformation généré par la
pénétration d’un socle plat et d’un bouchon de sable piégé dans une couche d’argile sous-jacente est calculé et comparé aux
algorithmes nouveaux et actuels. Cette analyse montre que la plus grande précision des mesures permet d’améliorer la clarté de
l’interprétation. [Traduit par la Rédaction]

Mots-clés : analyse des images, essais sur modèle, vélocimétrie par images de particules, corrélation d’images numériques.

Introduction nical analyses use some form of cross-correlation to obtain integer

Measurement of soil deformations in geotechnical models us- pixel displacements followed by subpixel interpolation of the cor-
ing particle image velocimetry (PIV) techniques (Adrian 1991) — relation peak (e.g., GeoPIV (White et al. 2003), MatPIV (Sveen
also known as digital image correlation (DIC) (Sutton et al. 1983, and Cowen 2004), PIVlab (Thielicke and Stamhuis 2014), and
2000) — has become routine experimental practice in many OpenPIV (Taylor et al. 2010)). In these algorithms the subsets (or
geotechnical research laboratories over the past 15 years (White patches) are generally not allowed to deform (so-called zero-order
et al. 2001, 2003; Rechenmacher and Finno 2004; Iskander 2010; deformation).
Hall 2012; Take 2015). The capabilities of the algorithms and anal- More sophisticated PIV–DIC algorithms exist, and are intro-
ysis techniques that underpin PIV–DIC have also improved signif- duced later in this paper. They incorporate higher-order subset
icantly over the same timeframe. shape functions (typically first-order, thus allowing displacement
Figure 1a illustrates the typical setup of a geotechnical PIV–DIC gradients across the subsets), image intensity interpolation, and
analysis for a shallow footing experiment. A wide range of defor- deformation parameter optimization (e.g., Sutton et al. 2000; Pan
mation levels are present and accurate and precise displacement et al. 2006, 2012). Some of these advances have been incorporated
data are desired in both the near- and far-field regions of the in commercially-available PIV–DIC software (e.g., Vic-2D (Correlated
model. In a typical PIV–DIC analysis, a region of interest (RoI) is Solutions 2015); LaVision DaVis8 (LaVision 2015); MI-2D (MatchID
first defined within the initial (“reference”) image of the model 2015) although these are not freely available to the academic
and populated with a mesh of subsets (or “patches”) of user- community and the specific algorithms used are often propri-
defined size. The displacement of these subsets in subsequent etary. For applications in which small strains are of interest,
(“target”) images is found using one of the approaches illustrated such as geotechnical modeling and structural monitoring, photo-
in Fig. 1b. Most freely available PIV–DIC software used for geotech- grammetric corrections are often needed to obtain sufficiently accu-

Received 27 May 2015. Accepted 19 October 2015.

S.A. Stanier and D.J. White. Centre for Offshore Foundation Systems, The University of Western Australia, M053 Fairway, Crawley, WA 6009, Australia.
J. Blaber. Department of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA, USA.
W.A. Take. Department of Civil Engineering, Queen’s University, Kingston, ON, Canada.
Corresponding author: Sam Stanier (email: [email protected]).

Can. Geotech. J. 53: 727–739 (2016) dx.doi.org/10.1139/cgj-2015-0253 Published at www.nrcresearchpress.com/cgj on 30 October 2015.
 728 Can. Geotech. J. Vol. 53, 2016

Fig. 1. PIV–DIC analysis overview: (a) typical PIV–DIC scenario with associated causes of error; (b) general overview of the PIV–DIC method.
[Colour online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18
For personal use only.

rate PIV–DIC results and these correction routines are not usually precision can have on the interpretation of a classical geotechni-
integrated within either freeware or commercial programs. cal problem.
The purpose of this paper is to (i) describe the advantages that
the more sophisticated genera of PIV–DIC algorithms provide for Computational method
geotechnical applications involving small and large deformations Digital images captured during a geotechnical model test are
and (ii) quantify these advantages via benchmark cases using a
usually analysed in sequence, starting with an initial reference
freely available noncommercial algorithm that is well-suited to
image. If the reference image is retained as the initial image (the
the analysis of geotechnical model test images.
The specific software used for the benchmark cases is referred so-called “leapfrog” scheme; see Fig. 2a) then zero-order subsets
to as GeoPIV-RG and is an update of the GeoPIV program (which can suffer a loss of correlation in regions experiencing large de-
represents the typical algorithms currently used in research, and formations (Fig. 2d) due to a mismatch between the subset shape
is described by White et al. (2003)). A brief overview of the com- and the deformation being observed. Alternatively, if the refer-
putational approach is given first. The comparison is then per- ence image is updated after every computation (the so-called “se-
formed using artificial “soil-like” images subjected to various quential” scheme; see Fig. 2b) to minimize the distortion that
modes of deformation. Lastly, an example application is given would reduce the correlation in the target images, random walk
that illustrates the impact this improvement in measurement errors are accumulated (White et al. 2003) because the overall

Published by NRC Research Press

 Stanier et al. 729

Fig. 2. Overview of (a) “leapfrog”, (b) “sequential”, and (c) “preconditioned optimization” computation schemes alongside (d) schematic plots
of the expected evolution of correlation coefficient and random walk errors. [Colour online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18
For personal use only.

Published by NRC Research Press

 730 Can. Geotech. J. Vol. 53, 2016

displacement is being found as the sum of many small displace- Fig. 3. Flowchart for GeoPIV-RG computations.
ments, each of which have an associated error (Fig. 2d). Random
walk errors can become significant in regions of low deformation
when calculating strains from the derivatives of displacement.
The current version of GeoPIV uses a combination of these two
schemes to minimize accumulated random walk errors whilst
maintaining tolerable correlations. The number of increments to
be performed using the leapfrog scheme prior to updating the
reference image (the so-called leapfrog parameter) is manually
defined by the user by trial and error guided by the amount of
deformation occurring between sequential images. Also, the
“search zone” over which the correlation measure is computed
for each subset (szone) is specified by the user in GeoPIV. Unneces-
sarily large values of szone lead to computational inefficiency in
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18

regions of images experiencing small displacements, therefore

small values are preferred. However, the value specified must be
larger than the maximum displacement expected to occur be-
tween the reference and target images (see Figs. 2a, 2b). Therefore,
szone also cannot be predetermined and requires further trial-and-
error refinement to achieve the best balance between accuracy
and computational efficiency.
The new implementation presented in this paper avoids the
need for trial-and-error refinement of either the leapfrog value or
the search zone parameter by following the process illustrated in
Fig. 2c for each subset. The overarching framework controlling
the computation process is the reliability-guided (RG) method pro-
posed by Pan (2009), as implemented in MATLAB by Blaber et al.
(2015) (so the software is referred to as GeoPIV-RG). Each reference
subset is allowed to deform using a shape (or warp) function de-
For personal use only.

scribing first-order deformations in conjunction with image-intensity

interpolation techniques to improve the correlation between ref-
erence and target subsets via optimization (Schreier and Sutton
2002). After an initial “seed” subset has been analysed, subsequent
computations are preconditioned using the results from the pre-
viously computed neighbouring subset that has the highest cor-
relation (the so-called “preconditioned optimization” scheme; see
Fig. 2c). This approach leads to the definition of a search zone
being unnecessary whilst allowing the effective search zone to be
the whole image if necessary should the preconditioning process
not yield a close match to the optimized solution for any particu-
lar subset. The reference image is updated when the correlation
coefficient for either the seed or one of the subsets contravenes
user-defined thresholds, effectively optimizing the leapfrog pa-
rameter. The first-order subset shape function (which allows for
linear gradients of displacement across the subset) leads to im-
proved precision and reduced random walk errors (see Fig. 2d)
because correlation is better preserved allowing the reference
image to be updated less frequently for image sequences experi-
encing low deformations. The overall programmatic structure of
the implementation is summarized by the flowchart shown in
Fig. 3.

Seed computation
The computation process begins at a selected subset (circular in
shape in this instance) from part of the RoI that experiences min-
imal deformation, meaning that the correlation between the ref-
erence and target subset will be high and thus the chances of an
incorrectly computed seed occurring will be low. This point is
used as a seed from which the RG computation process propa-
gates. The displacement of this subset is computed following the
procedure outlined in Fig. 4a. Initially, the displacement of the
seed subset is estimated to the nearest integer pixel value using
normalized cross correlation (NCC) (Lewis 1995). Next the subset is
allowed to deform using a subset shape function (p) that describes a combination with bi-quintic B-spline interpolation of the de-
superposition of the first-order subset deformation modes illus- formed subset pixel intensities (e.g., Cheng et al. 2002) to adjust
trated in Fig. 5 (Pan et al. 2006). The inverse compositional Gauss– the subset shape function until the correlation between the refer-
Newton (IC–GN) method described by Pan et al. (2013) is used, in ence and target subsets is optimized. The exit criterion for the

Published by NRC Research Press

 Stanier et al. 731

Fig. 4. Flowcharts for (a) seed subset and (b) RG subset computation subroutines. IC-GN, inverse compositional Gauss–Newton method.
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18
For personal use only.

optimization is a user-defined maximum magnitude for the norm ters for the seed subset as an initial estimate (Zhou and Chen
of the subset shape function difference vector (|⌬p|max) between 2012). These four subsets are then placed in a queue of descending
successive computations (typically 1 × 10−5) and a maximum num- CCZNCC from which the subset with the optimal correlation coef-
ber of iterations (maxiter) per subset (typically 50). The zero- ficient is selected first. IC–GN and bi-quintic B-spline interpolation
normalized cross-correlation correlation coefficient (CCZNCC) is is once again used to optimize the deformation parameters for
used to indicate the degree of match where values of 1, 0, and –1 this subset, then it is removed from the queue. If not already in
indicate perfect, zero, and inverse correlation, respectively (Pan the queue, the CCZNCC is computed for the neighboring subsets
et al. 2010). The seed computation is deemed successful if the and those subsets are added to the queue. This process repeats,
CCZNCC value is greater than a user-defined limit, CCZNCC-seed-tol
calculating the displacement and deformation of all of the subsets
(typically 0.9).
across the entire RoI. The advantage of the RG framework over the
Reliability-guided computations usual processing of subsets across successive rows is three-fold.
Assuming the seed computation was successful, a high-density Firstly, the NCC, which is computationally expensive to deter-
grid of subsets is processed using the RG framework outlined in mine, is only computed for the seed subset and covers the entire
Fig. 4b. Firstly the CCZNCC of the four subsets surrounding the seed region of interest. Secondly, the subsets with higher correlation
are estimated using the displacement and deformation parame- coefficients are processed first, and the optimized deformation

Published by NRC Research Press

 732 Can. Geotech. J. Vol. 53, 2016

Fig. 5. Subset deformation modes considered by the first-order image can be periodically updated. In earlier versions of GeoPIV
shape function, p: (a, b) displacements and (c–f) displacement the updating interval was specified manually by the user and
gradients. Note: subsets are represented as squares for clarity, but refined through trial and error (White et al. 2003). An automatic
subset shape is arbitrary. reference image-updating scheme, similar to that proposed by
Pan et al. (2012), is used in GeoPIV-RG. After completion of the RG
computations for each target image, the CCZNCC for each subset is
compared with a second, slightly relaxed, user-defined limit (typ-
ically 0.75) denoted CCZNCC-min-tol. Using a relaxed tolerance on the
minimum permissible full field CCZNCC allows large deformations
to occur in certain regions of the model prior to reference image
updating. If the CCZNCC for any subset is less than CCZNCC-min-tol, then
the reference image is updated to the target image from the last
successful increment, otherwise the current reference image is
carried forward.
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18

The RG method is programmed to compute the displacement of

regularly gridded subsets. Consequently, an interpolation scheme
is required to compute the displacement of the user-defined sub-
set locations from the regularly gridded RG output. Bi-cubic spline
interpolation of the output from the RG process achieves this. To
safeguard the precision of the measurements during the interpo-
lation process, the subset spacing used in the RG process is re-
duced relative to that given in the user-defined mesh. For the
benchmarking analyses presented in this paper, halving the grid
spacing was sufficient to preserve accuracy of the measurements.
Due to the computational efficiency of the preconditioned IC–GN
optimization process, the computational cost of this four-fold in-
crease in RG computation grid density is minimal. The reference
image updating procedure ensures that the updating interval is
always optimized. The user can control the frequency of reference
For personal use only.

image updating indirectly by varying the correlation coefficient

tolerances, CCZNCC-seed-tol and CCZNCC-min-tol, with stricter values
resulting in more frequent reference image updating.

Performance comparison
Methodology
The performance of GeoPIV-RG is compared with GeoPIV (de-
scribed by White et al. 2003 and White et al. 2005), which is widely
used in geotechnical research and is typical of the many freely
available zero-order algorithms (e.g., MatPIV, PIVlab, and Open-
PIV). Therefore the following benchmarking analyses are gener-
ally indicative of the improvements in measurement precision
that can be attained by incorporating advances in PIV–DIC includ-
ing first-order subset shape functions, image intensity interpola-
tion, deformation parameter optimization, and automatic reference
image updating schemes similar to those described earlier.
Artificial images that represent geomaterials are preferred for
such benchmarking as they can be subjected to precisely pre-
scribed deformations and are unaffected by camera-induced lens
parameters are used to precondition the IC–GN optimization of distortions and camera-target movements (Lee et al. 2012). The
the neighbouring subsets. Thirdly, as a result of this approach, the images were generated in MATLAB by randomly projecting thou-
need to specify the expected maximum displacement within the sands of white dots onto a black background to subpixel posi-
displacement field is eliminated. tional resolution. Each white dot is defined by a Gaussian
brightness peak centered at a specified location. In this way the
Reference image updating location of the dot can be controlled precisely allowing smooth
Compared to conventional PIV–DIC applications, geotechnical spatially varying displacement fields to be prescribed. The refer-
model testing can involve tracking of a larger range of deforma- ence artificial image used in all of the artificial benchmarking
tions and highly circuitous displacement paths. For example, soil analyses presented herein is shown in Fig. 6a with the 1681 subset
often flows around penetrometers and deeply buried foundations. locations marked by yellow crosses in a centrally located zone, of
Unlike fluid mechanics studies, where constitutive relations are 400 × 400 pixels (px). This subset population is sufficiently large to
not a focus, geotechnical research is concerned with both instan- generate statistically valid measurements of the error in the
taneous flow fields and also the strain path histories, stress–strain image-based displacement measurements. For GeoPIV the side
behaviour, and the associated constitutive relations. Additional length of the square subsets, Ls, was 45 px, while for GeoPIV-RG
measures are therefore required to handle the resulting changes the subset diameter, Ds, was taken as 50 px giving comparable
in subset appearance, for example due to grain rearrangement, total pixels per subset (within ⬃3%).
because these lead to a reduction in the subset correlation and The theoretical and measured displacements at the 1681 subset
cause erroneous displacements to be estimated (known as “wild” locations were compared for four deformation modes: (i) rigid-
vectors). To counter this degradation of correlation, the reference body translation, (ii) rigid-body rotation, (iii) vertical strain, and

Published by NRC Research Press

 Stanier et al. 733

Fig. 6. Artificial images and imposed deformations: (a) example The precision error in the displacement measurements is quan-
image and subset locations; (b) rigid-body translation; (c) rigid-body tified by the standard error, ␳px, defined as the standard deviation
rotation; (d) vertical strain; (e) pure shear strain. [Colour online.] of the difference between the theoretical and calculated subset
displacements over the 1681 subsets. It is shown later that this
error increases as the deformation of the subset increases. To
convert these standard errors to a measure of the precision with
which deformations can be determined, the measurement errors
(␦) are defined as the error in the measured deformation, for a
given level of that deformation mode. An estimate of the random
error ␦ for each mode (rotation, vertical strain, and pure shear
strain) can then be defined as

(1) ␦␪ ⫽ tan⫺1 冉 兹2␳px

L
冊
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18

兹2␳px
(2) ␦␧y ⫽
L

兹2␳px
(3) ␦␧xy ⫽
L

where ␪, ␧y, and ␧xy denote rotation in degrees, vertical strain, and
pure shear strain, respectively, and L is a nominal gauge length in
pixels. Each random error ␦ is estimated for L = 25, 250, and
2500 px, which covers the range typically relevant. When viewing
a geotechnical model test, the varying deformation throughout
the image is of interest, so the relevant gauge length is compara-
ble to the spacing of the subsets, i.e., of the order of 25 px. Alter-
For personal use only.

natively, when viewing a geotechnical element test, at a stage

when the deformation is uniform, the gauge length might be
significantly larger, i.e., of the order of 2500 px. The ␦ estimates,
combined with a tolerable measurement error, allow the required
subset spacing to be identified. Alternatively, they show the image
scale (e.g., px/mm) required in a model test to detect a specific
level of deformation.

Results
Rigid-body translation
The standard errors, ␳px, in Fig. 7a are ⬃0.006 and ⬃0.001 px for
GeoPIV and GeoPIV-RG, respectively, indicating a modest im-
provement in precision for subpixel displacement measurement
for the new methods. Meanwhile, Fig. 7b presents the mean bias
error, ␮bias (Schreier et al. 2000), which is the mean discrepancy
between the actual and measured displacement, for the subpixel
rigid-body translation analysis. For GeoPIV, which uses NCC to
obtain the integer pixel displacements prior to subpixel interpo-
lation using bi-cubic splines, a significant periodic variation in
␮bias is evident for non-integer or non-half-integer displacements.
This behaviour is consistent with that reported by Amiot et al.
(2013) for PIV–DIC software incorporating bi-cubic interpolation.
In contrast, the bi-quintic B-spline interpolation process used in
the IC–GN optimization of GeoPIV-RG experiences mean bias er-
rors that never exceed 0.0005 px. This is comparable to the per-
formance reported by Lee et al. (2012) for NCC with bi-quintic
(iv) pure shear strain as illustrated in Figs. 6b–6e. For rigid-body
B-spline subpixel interpolation and consistent with the best per-
translation, displacement was applied in 0.025 px increments up
forming PIV–DIC software (with respect to mean bias errors) re-
to a maximum of 1 px. For the rotation, vertical strain, and pure
ported by Amiot et al. (2013) that also incorporated bi-quintic
shear strain analyses, the deformation magnitude imposed was
B-spline interpolation. Minimizing bias errors is particularly impor-
increased over 100 logarithmically spaced intervals up to a maxi- tant if strains are to be derived from the derivatives of displacement
mum of 90° of rotation and 50% strain. The total deformation fields as periodic bias can lead to erroneous localizations in strain
applied was chosen such that reference image updating was peri- fields.
odically required to validate the efficacy of the full computational
scheme. For the analyses performed using the current version of Rigid-body rotation
GeoPIV, the reference image was updated manually as infre- Figure 8a shows the evolution of ␳px with rigid-body rotation.
quently as possible to minimize the summation of random walk GeoPIV accumulates significant errors with increasing rotation
errors. because the ability of NCC to accurately track subset displace-

Published by NRC Research Press

 734 Can. Geotech. J. Vol. 53, 2016

Fig. 7. Rigid-body translation performance: (a) standard error, ␳px, and (b) mean bias error, ␮bias, for GeoPIV and GeoPIV-RG. [Colour online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18

Fig. 8. Rotation performance: (a) standard error for GeoPIV and GeoPIV-RG; (b) effect of gauge length on rotation error, ␦␪, for GeoPIV-RG.
[Colour online.]
For personal use only.

ments degrades progressively with rotation (Dutton et al. 2014), origin of rotation. However, the error magnitudes are clearly linked
causing accumulating drift in the displacements. The iterative to the rotation angle as the error magnitude tends to increase with
subset deformation optimization performed by GeoPIV-RG miti- rotation angle, as is confirmed for GeoPIV in Fig. 9c and GeoPIV-RG in
gates the degradation of correlation, resulting in a comparatively Fig. 9e where the evolution of errors is presented with respect to the
small precision error of ␳px < 1/1000th of a pixel, irrespective of the rotation angle imposed for subsets located at sections A–A, B–B,
rotation magnitude, whilst the mean errors are always less than and C–C.
1 × 10−4 px. As a result, the rotation error, ␦␪, is approximately For a basic PIV–DIC analysis where vector plots are used to
constant for rotation magnitudes greater than 1° (Fig. 8b). illustrate soil flow mechanisms, the poor performance of GeoPIV
The divergence in performance between GeoPIV and GeoPIV-RG for rotation is not necessarily problematic as the large magnitude
observed in Fig. 8 has a profound effect on the abilities of the of the vectors within the soil flow mechanism will mask the errors
respective algorithms to measure strain fields because the magni- induced by rotation. However, if strains are derived from the de-
tude of the error is random and not linked to the magnitude of the rivatives of the displacement components, the errors become very
displacement of the subset. To illustrate this, an additional anal- significant. For example, for a strain element with length L = 25 px
ysis was performed using the same analysis parameters and arti- positioned either side of subset A–A, similarly to eqs. (1)–(3), a
ficial images for a horizontal row of subsets spaced at 1 px generic estimate of the strain error (␦␧) can be taken as follows:
intervals at increasing distances from the origin of rotation (Lo) up
to a maximum of 200 px. For these subsets and for all rotation 2| ␦r |
increments, the rotation angle is the same but the displacement (4) ␦␧ ⫽
L
magnitude increases proportionally with the distance from the
origin of rotation, as illustrated in Fig. 9a. The resultant error
magnitudes (␦r) were calculated for rotation angles of 0.0°, 0.5°, For a rotation angle of 2°, the strain error for GeoPIV is of the
1.0°, 1.5°, and 2.0° and are plotted with respect to the distance order of ⬃1.2%, significantly limiting the ability to plot meaning-
from the origin of rotation in Figs. 9b and 9d for GeoPIV and ful strain fields. On the other hand, the strain errors for
GeoPIV-RG, respectively. GeoPIV-RG are less than ⬃0.012%, resulting in a two order of mag-
From these results it is clear that the error magnitudes and nitude improvement in strain measurement resolution as a result
directions are random and unrelated to the distance from the of the additional degrees of freedom provided by the first-order

Published by NRC Research Press

 Stanier et al. 735

Fig. 9. Randomness of rotation performance: (a) artificial image and a row of subset displacements (magnitudes amplified) illustrating the
rotation imposed; resultant error, ␦r, as a function of distance from the origin of rotation and rotation angle for (b, c) GeoPIV and (d, e) GeoPIV-RG.
Note: Vertical scales are different between (a, b) and (c, d) for clarity. [Colour online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18
For personal use only.

Published by NRC Research Press

 736 Can. Geotech. J. Vol. 53, 2016

Fig. 10. Vertical strain performance: (a) standard error for GeoPIV and GeoPIV-RG; (b) effect of gauge length on vertical strain error, ␦␧y, for
GeoPIV-RG. [Colour online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18

Fig. 11. Shear strain performance: (a) standard error for GeoPIV and GeoPIV-RG; (b) effect of gauge length on shear strain error, ␦␧xy, for
GeoPIV-RG. [Colour online.]
For personal use only.

subset shape function. The impact of this improvement is signif- strain of 0.1% (100 ␮3) the strain error is ⬃5 × 10−7 (point E in
icant for geotechnical research applications where more than Figs. 10b and 11b) or 0.05 ␮3 (equivalent to an SNR of ⬃2000).
simple instantaneous flow mechanisms are to be observed, such All of these artificial image analyses show that the new ap-
as when gross element distortions are to be monitored through proach is at least an order of magnitude more precise than the
large rotations. combination of NCC and bi-cubic spline interpolation method
employed by GeoPIV (White et al. 2003), for both small and large
Vertical and shear strain deformations.
Figures 10a and 11a illustrate that measurement precision dur-
ing deformation — either through vertical strain or shear strain — is Example application
similar to rotation (␳px ⬃1/1000 px) until ⬃0.01% or 1% strain,
beyond which the precision error rises approximately linearly To demonstrate the application of the new methodology, im-
with strain. ages from a model test performed in the drum centrifuge at The
When converted to ␦ estimates (Figs. 10b and 11b), these results University of Western Australia are used. The test involved similar
allow the practicality of detecting a given strain level to be as- techniques to the work reported by Hu et al. (2014), investigating
sessed. For example, if a model test with a varying deformation is punch-through of a 30 mm diameter flat footing (at 200g, so equiv-
being viewed, then a subset spacing of typically 25 px might apply. alent to 6 m diameter at prototype scale) on a 20 mm deep (4 m
In this case, if zones of the model have undergone strains of prototype) sand layer overlying clay. In the experiment, ⬃550 images
1% or 10% (points M1 and M2 in Figs. 10b and 11b), the resulting were recorded at a frequency of 5 Hz using the apparatus de-
strain errors would be ⬃0.025% and ⬃0.25%, respectively (equiva- scribed by Stanier and White (2013). Artificial texture was applied
lent to a signal-to-noise ratio, SNR, of ⬃40), which is likely to be to the exposed face of the model at the optimal artificial seeding
adequate for producing detailed and smooth deformation fields ratio (ASR) following the procedure proposed by Stanier and
(e.g., 20 contours of 0.05% or 0.5% up to the maximum of 1% or 10%, White (2013) to maximize the precision of the image-based defor-
respectively). Alternatively, if the application is an element test mation measurements.
with uniform deformation, so the gauge length is a larger propor- Analyses were performed on the series of images for the under-
tion of the image width — typically 2500 px in width — then at a lying clay layer only, using both GeoPIV and GeoPIV-RG with the

Published by NRC Research Press

 Stanier et al. 737

Table 1. Computation parameters for ex-

ample analysis.
Parameter Value
GeoPIV parameters
Ls (px) 45a
s (px) 25
szone (px) 15
Ref. image updating interval 10
GeoPIV-RG parameters
Ds (px) 50a
s (px) 25
maxiter 50
|⌬p|max 1×10−5
CCZNCC-seed-tol 0.9
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18

CCZNCC-min-tol 0.75

Fig. 12. Example application: flat footing on sand-over-clay; total maximum shear strain, ␰, at 1D penetration using (a) GeoPIV and (b) GeoPIV-RG.
Ref. image updating interval ⬃20
aSubset sizes chosen to have equivalent area

(within ⬃3%).

analysis settings summarized in Table 1. The time taken by each of

the algorithms to perform an analysis is dependent upon a num-
ber of factors, including the subset spacing, deformation magni-
tude, image texture quality, and available processing power. For
this particular analysis, GeoPIV-RG performed the computations
in ⬃20% of the time taken by GeoPIV. The total maximum shear
strain, ␰ (i.e., ⌬␧1 – ⌬␧2 summed through the deformation), was
calculated from the displacement fields following the large strain
procedure of White and Bolton (2004).
Figure 12 presents the distributions of ␰ after 1D of footing pen-
For personal use only.

etration. Significant noise is evident in the results from GeoPIV

from summed random walk errors and degradation of correlation
due to subset rotation and deformation. In contrast, the analysis
generated by GeoPIV-RG has lower noise, as is evident in the re-
gions experiencing small strains. Figure 13 shows a horizontal
cross section through both analyses at an initial normalized
depth, z/D, of 1.5, presented in terms of both the normalized dis-
placement magnitude and total maximum shear strain. The first-
order deformation algorithm of GeoPIV-RG results in smoother
spatial variation of both the displacements and strains across the
model, compared to the stepped cross section resulting from the
zero-order deformation algorithm of GeoPIV.
These enhancements result from the subset deformation opti-
mization capability of GeoPIV-RG as it preserves correlation and
precision in regions of large deformation. The reference image
also requires updating less frequently (see Table 1), which in turn
minimizes random walk errors. These advances create more precise
deformation measurements, which unlocks additional potential
applications. For example, more detailed verification of constitu-
tive models by extracting element-level responses within model
tests, and the quantification of geomaterial behaviour at both
smaller strains and higher levels of rotation and deformation
than was possible using previous image analysis methods.

Conclusions
This paper has shown that recent advances in PIV–DIC algo-
rithms coupled with photogrammetric correction routines allow
improved deformation measurements for geotechnical applications.
The algorithms have been incorporated in an update of a com-
monly used freeware PIV–DIC program. The prior version has
been used as a benchmark representing the approaches com-
monly used in geotechnical physical modelling. The benchmark-
ing used a series of artificial soil-like images subjected to prescribed
displacements and deformations. The advanced algorithms are
faster and more precise than the simpler zero-order PIV–DIC ap-
proach that is widely used and freely available to the research
community. Rigid-body displacements can be detected to a preci-
sion of ⬃0.001 px. There is a modest reduction in precision when

Published by NRC Research Press

 738 Can. Geotech. J. Vol. 53, 2016

Fig. 13. (a) Normalized displacement, w/D, and (b) total maximum shear strain, ␰, along the cross sections in Figs. 12a and 12b at an initial
normalized depth, z/D, of 1.5. [Colour online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18

the tracked soil is deforming. The effect of the gauge length (i.e., Correlated Solutions. 2015. The Vic-2D System: technology overview. Available
the separation of the measurement points) and the level of defor- at http://www.correlatedsolutions.com/vic-2d/.
Dutton, M., Take, W.A., and Hoult, N.A. 2014. Curvature monitoring of beams
mation or rotation on the precision of deformation measure- using digital image correlation. Journal of Bridge Engineering, 19(3): 05013001.
ments is quantified. For example, it is shown that when soil doi:10.1061/(ASCE)BE.1943-5592.0000538.
elements at close spacing experience rotation in a model test, the Hall, S. 2012. Digital image correlation in experimental geomechanics. ALERT
new implementation is approximately two orders of magnitude Geomaterials Doctoral Summer School 2012. ALERT Geomaterials.
pp. 69–102. ISBN: 978-3-00-039683-0.
more precise than the existing approach, resulting in significantly
For personal use only.

Hu, P., Stanier, S.A., Cassidy, M.J., and Wang, D. 2014. Predicting peak resistance
less noise in strain fields derived from the derivatives of the dis- of spudcan penetrating sand overlying clay. Journal of Geotechnical and
placements. At the other end of the scale, in an element test in Geoenvironmental Engineering, 140(2): 04013009. doi:10.1061/(ASCE)GT.1943-
which digital images are used to monitor the overall response of 5606.0001016.
the sample (i.e., the gauge length is significantly larger and taken Iskander, M. 2010. Modelling in transparent soils. Springer. ISBN-13:
978–3642025006.
here as 2500 px), the standard error is ⬃0.05 ␮3 at a strain of 0.1%. LaVision. 2015. Digital image correlation. Available at http://www.lavision.de/en/
An example analysis illustrates the value of the new approach techniques/digital-image-correlation.php.
by showing improved measurement of deformations during punch- Lee, C., Take, W.A., and Hoult, N.A. 2012. Optimum accuracy of two-dimensional
through of a flat footing on a sand-over-clay stratigraphy. These strain measurements using digital image correlation. Journal of Computing
in Civil Engineering, 26: 795–803. doi:10.1061/(ASCE)CP.1943-5487.0000182.
results demonstrate the benefits of the enhanced measurement Lewis, J.P. 1995. Fast normalized cross-correlation. Industrial Light and Magic.
precision provided by this software, which is freely available to [Self-published, expanded, and corrected version of: Lewis, J.P. 1995. Fast
the geotechnical research community. template matching. In Proceedings of Vision Interface 95, Canadian Image
Processing and Pattern Recognition Society, Quebec City, Que., 15–19 May
Acknowledgements and supplementary information 1995. pp. 120–123.]
MatchID. 2015. MI-2D: 2D and stereo digital image correlation platform. Avail-
The research presented here forms part of the activities of the able at http://www.matchidmbc.be/.
Centre for Offshore Foundation Systems (COFS), currently sup- Pan, B. 2009. Reliability-guided digital image correlation for image deformation
ported as a node of the Australian Research Council Centre of measurement. Applied Optics, 48(8): 1535–1542. doi:10.1364/AO.48.001535.
Excellence for Geotechnical Science and Engineering (grant PMID:19277087.
Pan, B., Xie, H., Xu, B., and Dai, F. 2006. Performance of sub-pixel registration
CE110001009) and through the Fugro Chair in Geotechnics, the algorithms in digital image correlation. Measurement Science and Technol-
Lloyd’s Register Foundation Chair and Centre of Excellence in ogy, 17: 1615–1621. doi:10.1088/0957-0233/17/6/045.
Offshore Foundations, and the Shell EMI Chair in Offshore Engi- Pan, B., Xie, H., and Wang, Z. 2010. Equivalence of digital image correlation
neering (held by the fourth author). The third author is supported criteria for pattern matching. Applied Optics, 49(28): 5501–5509. doi:10.1364/
AO.49.005501. PMID:20885489.
by the Natural Sciences and Engineering Research Council of Canada Pan, B., Wu, D., and Xia, Y. 2012. Incremental calculation for large deformation
(NSERC) and holds a Canada Research Chair at the GeoEngineering- measurement using reliability-guided digital image correlation. Optics and
Centre at Queen’s–RMC. The software code and example images Lasers in Engineering, 50: 586–592. doi:10.1016/j.optlaseng.2011.05.005.
relevant to this paper can be downloaded from http://www. Pan, B., Li, K., and Tong, W. 2013. Fast, robust and accurate digital image corre-
geopivrg.com. lation calculation without redundant computations. Experimental Mechan-
ics, 53: 1277–1289. doi:10.1007/s11340-013-9717-6.
Rechenmacher, A.L., and Finno, R.J. 2004. Digital image correlation to evaluate
References shear banding in dilative sands. Geotechnical Testing Journal, 27(1): 13–22.
Adrian, R.J. 1991. Particle-imaging techniques for experimental fluid mechanics. doi:10.1520/GTJ11263J.
Annual Review of Fluid Mechanics, 23: 261–304. doi:10.1146/annurev.fl.23. Schreier, H.W., and Sutton, M.A. 2002. Systematic errors in digital image corre-
010191.001401. lation due to undermatched subset shape functions. Experimental Mechan-
Amiot, F., Bornert, P., Dupre, J.-C., Fazzini, M., Orteu, J.-J., Poilâne, C., Robert, L., ics, 42(3): 303–310. doi:10.1007/BF02410987.
Rotinat, R., Toussaint, E., et al. 2013. Assessment of digital image correlation Schreier, H.W., Braasch, J.R., and Sutton, M.A. 2000. Systematic errors in digital
measurement accuracy in the ultimate error regime: main results of a col- image correlation caused by intensity interpolation. Optical Engineering,
laborative benchmark. Strain, 49: 483–496. doi:10.1111/str.12054. 39(11): 2915–2921. doi:10.1117/1.1314593.
Blaber, J., Adair, B., and Antoniou, A. 2015. Ncorr: open-source 2D digital image Stanier, S.A., and White, D.J. 2013. Improved image-based deformation measure-
correlation MATLAB software. Experimental Mechanics, 55(6): 1105–1122. ment in the centrifuge environment. Geotechnical Testing Journal, 36(6):
doi:10.1007/s11340-015-0009-1. 1–14. doi:10.1520/GTJ20130044.
Cheng, P., Sutton, M.A., Schreier, H.W., and McNeill, S.R. 2002. Full-field speckle Sutton, M.A., Wolters, W.J., Peters, W.H., Ranson, W.F., and McNeill, S.F. 1983.
pattern image correlation with B-spline deformation function. Experimental Determination of displacements using an improved digital correlation method.
Mechanics, 42(3): 344–352. doi:10.1007/BF02410992. Image and Vision Computing, 1(3): 133–139. doi:10.1016/0262-8856(83)90064-1.

Published by NRC Research Press

 Stanier et al. 739

Sutton, M.A., McNeill, S.R., Helm, J.D., and Chao, Y.J. 2000. Advances in two- d displacement
dimensional and three-dimensional computer vision. Photomechanics, Top- L gauge length
ics in Applied Physics, 77: 323–372. doi:10.1007/3-540-48800-6_10. Lo distance from the origin of rotation
Sveen, J.K., and Cowen, E.A. 2004. Quantitative imaging techniques and their Ls length of GeoPIV subset
application to wavy flow. In PIV and water waves. Edited by J. Grue, P.L.F. Liu,
and G.K. Pedersen. World Scientific.
maxiter maximum number of iterations per subset
Take, W.A. 2015. Thirty-Sixth Canadian Geotechnical Colloquium: Advances in NCC normalized cross correlation
visualization of geotechnical processes through digital image correlation. p subset deformation shape function
Canadian Geotechnical Journal, 52(9): 1199–1220. doi:10.1139/cgj-2014-0080. |⌬p|max maximum norm of the shape function difference
Taylor, Z.J., Gurka, R., Kopp, G.A., and Liberzon, A. 2010. Long-duration time- vector
resolved PIV to study unsteady aerodynamics. IEEE Transactions on Instru- r radius
mentation and Measurement, 59(12): 3262–3269. doi:10.1109/TIM.2010.2047149. SNR signal-to-noise-ratio
Thielicke, W., and Stamhuis, E.J. 2014. PIVlab – towards user-friendly, affordable s subset spacing
and accurate digital particle image velocimetry in MATLAB. Journal of Open
Research Software, 2(1): e30. doi:10.5334/jors.bl.
szone search zone parameter
White, D.J., and Bolton, M.D. 2004. Displacement and strain paths during plane- u horizontal displacement
strain model pile installation in sand. Géotechnique, 54(6): 375–397. doi:10. V vertical load
1680/geot.2004.54.6.375. v vertical displacement
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 186.106.47.150 on 02/28/18

White, D.J., Take, W.A., Bolton, M.D., and Munachen, S.E. 2001. A deformation w foundation displacement
measuring system for geotechnical testing based on digital imaging, close- x horizontal position
range photogrammetry, and PIV image analysis. In Proceedings of the 15th xf final horizontal position
International Conference on Soil Mechanics and Geotechnical Engineering, xi initial horizontal position
Istanbul, Turkey. Balkema, Rotterdam. pp. 539–542.
y vertical position
White, D.J., Take, W., and Bolton, M. 2003. Soil deformation measurement using
Particle Image Velocimetry (PIV) and photogrammetry. Géotechnique, 53(7): yf final vertical position
619–631. doi:10.1680/geot.2003.53.7.619. yi initial vertical position
White, D.J., Take, W.A., and Bolton, M.D. 2005. Discussion of “Accuracy of Digital z depth
Image Correlation for measuring deformations in transparent media” by ␦ measurement error
Samer Sadek, Magued G. Iskander, and Jinyuan Liu. Journal of Computing in ␦r resultant error
Civil Engineering, 19(2): 217–219. doi:10.1061/(ASCE)0887-3801(2005)19:2(217). ␦␧ strain error
Zhou, Y., and Chen, Y.Q. 2012. Propagation function for accurate initialization ␦␧xy shear strain error
and efficiency enhancement of digital image correlation. Optics and Lasers in
Engineering, 50(12): 1789–1797. doi:10.1016/j.optlaseng.2012.06.017.
␦␧y vertical strain error
␦␪ rotation error
␧1 major principal strain
List of symbols ␧2 minor principal strain
For personal use only.

␧xy pure shear strain

ASR artificial seeding ratio ␧y vertical strain
CCZNCC zero normalized cross-correlation correlation coef- ␪ rotation angle
ficient ␮bias mean bias error
CCZNCC-min-tol full field correlation coefficient tolerance ␰ total maximum shear strain
CCZNCC-seed seed correlation coefficient ␳px standard error of displacement measurement in
CCZNCC-seed-tol seed correlation coefficient tolerance pixels
D diameter
Ds diameter of GeoPIV-RG subset

Published by NRC Research Press