2003 - Soil Deformation Measurement Using Particle Image Velocimetry PIV by WHITE

8/11/2019 2003 - Soil Deformation Measurement Using Particle Image Velocimetry PIV by WHITE

1/13

White, D. J., Take, W. A. & Bolton, M. D. (2003). Geotechnique 53, No. 7, 619631

619

Soil deformation measurement using particle image velocimetry (PIV)and photogrammetry

D . J . W H I T E, W . A . T A K E a n d M . D . B O L T O N

A deformation measurement system based on particleimage velocimetry (PIV) and close-range photogram-metry has been developed for use in geotechnical testing.In this paper, the theory underlying this system is de-scribed, and the performance is validated. Digital photo-graphy is used to capture images of planar soildeformation. Using PIV, the movement of a fine mesh ofsoil patches is measured to a high precision. Since PIVoperates on the image texture, intrusive target markersneed not be installed in the observed soil. The resultingdisplacement vectors are converted from image space toobject space using a photogrammetric transformation. Aseries of validation experiments are reported. These

demonstrate that the precision, accuracy and resolutionof the system are an order of magnitude higher thanprevious image-based deformation methods, and are com-parable to local instrumentation used in element testing.This performance is achieved concurrent with an orderof magnitude increase in the number of measurementpoints that can be fitted in an image. The performance ofthe system is illustrated with two example applications.

KEYWORDS: centrifuge modelling; deformation; ground move-ments; laboratory equipment; laboratory tests; model tests

Nous avons developpe un systeme de mesure des defor-mations base sur la velocimetrie dimage de particules(PIV) et la photogrammetrie rapprochee, pour lutilisa-tion dans les essais geotechniques. Dans cet article, nousdecrivons la theorie a la base du systeme et nous validonssa performance. Nous utilisons la photographie numeri-que pour saisir les images de deformation planaires dusol. En utilisant la PIV, nous mesurons avec precision lemouvement de morceaux de sol a grains fins. Comme laPIV sappuie sur la texture de limage, il nest pasnecessaire dinstaller des marqueurs de cible intrusifsdans le sol observe. Nous convertissons les vecteurs dedeplacement qui en resultent despace-image en espace-

objet en utilisant une transformation photogrammetrique.Nous donnons les resultats dune serie dessais de valida-tion. Ceux-ci demontrent que la precision, lexactitude etla resolution du systeme sont dun ordre de grandeurplus eleve quavec les methodes precedentes de deforma-tion basees sur limage et sont comparables a linstru-mentation locale utilisee dans les essais elementaires.Nous obtenons cette performance en meme temps quuneaugmentation de lordre de grandeur dans le nombre depoints de mesure qui peuvent etre places dans une image.Nous illustrons la performance du systeme avec deuxexemples dapplications.

INTRODUCTIONDeformation measurement in geotechnical testingThe assessment of soil behaviour in element tests or physicalmodels at strain levels representative of prototype situationsrequires the detection of pre-failure strains. Modern tech-niques of local instrumentation have increased the precisionof element test strain measurement into the small-strain(,0001%) range (Scholey et al., 1995). However, measure-ment techniques for the construction of planar deformationfields in geotechnical model tests remain less precise.

Various image-based techniques have been used tomeasure planar deformation fields in geotechnical elementand model tests. These include X-ray (Roscoe et al.,

1963; Phillips, 1991) and stereo-photogrammetric methods(Butterfield et al., 1970; Andrawes & Butterfield, 1973).More recently, the use of computer-based image processingtechniques has led to the development of automatic targettracking systems (Taylor et al., 1998; Paikowsky & Xi,2000).

These experimental techniques rely on the presence ofartificial targetsusually lead shot or coloured beadswith-in the deforming soil to provide reference points that aretracked as the element test or modelling event proceeds. Thereliance on target markers has a number of drawbacks. A

dense grid of markers can influence the behaviour. A widelyspaced grid of markers provides sparse data. Markerscan become obscured by soil during the course of anexperiment.

Performance: accuracy, precision and resolutionThe performance of a measurement system can be as-

sessed by considering the errors associated with accuracy,precision and resolution. Accuracy is defined as the systema-tic difference between a measured quantity and the truevalue. Precision is defined as the random difference betweenmultiple measurements of the same quantity. Resolution is

defined as the smallest interval that can be present in areading.

The precision of a deformation measurement system basedon the tracking of target markers depends on the methodused to identify the location of the target within the photo-graph or digital image. The accuracy of the system dependson the process used to convert from the measured locationwithin the image (image-space coordinates) to the locationwithin the element or model test (object-space coordinates).Both precision and accuracy can be expressed in a non-dimensional fashion as a fraction of the width of the field ofview (FOV).

By directly measuring the position of target markers onX-ray film, Phillips (1991) achieved a measurement preci-

sion of 1/10 000th of the FOV. Butterfield et al. (1970) andAndrawes &Butterfield (1973) achieved a comparable preci-sion (1/11 000th of the FOV) using a stereo-photogrammetricmethod to measure photographic film. An advantage of the

Manuscript received 25 November 2002; revised manuscriptaccepted 30 April 2003.

Discussion on this paper closes 1 March 2004, for further detailssee p. ii. The Schofield Centre, Cambridge University Engineering Depart-ment, UK.


2/13

stereo-photogrammetric method over conventional target-tracking techniques is that embedded markers are notrequired if soil grains can be distinguished. No assessmentof accuracy was made in either case. Instead, it was assumedthat image-space measurements could be scaled directly intoobject-space coordinates. In other words, it was assumed thatthe image scaledefined as the ratio between an object-space length and its length within the image or radiograph

was constant across the field of view.Taylor et al. (1998) avoided the assumption of constantimage scale by using close-range photogrammetry to recon-struct the object-space coordinates of target markers fromthe image-space measurements. An accuracy of 1/5600th ofthe FOV was achieved, when tracking target markers usingvideo photography. Paikowsky & Xi (2000) used a second-degree polynomial fit to reconstruct object-space coordinates,and achieved an accuracy of 1/1266th of the FOV.

This paper describes the use of particle image velocimetry(PIV) and close-range photogrammetry to measure deforma-tion in geotechnical models and element tests. The resultingperformance is shown to exceed conventional image-basedmeasurement techniques and offers precision and accuracythat are comparable to local instrumentation.

DIGITAL PHOTOGRAPHYThe measurement technique described in this paper oper-

ates by processing digital images, which can be captureddirectly from digital cameras or grabbed from an analoguevideo signal. In digitised form, a monochrome image is amatrix containing the intensity (brightness) recorded at eachpixel on the cameras charge coupled device (CCD). Thisintensity matrix is defined as I(U), where U (u, v) is thepixel coordinate. Colour images consist of three intensitymatrices, one for each colour channel.

Most geotechnical physical model and element tests takeplace sufficiently slowly to be captured by digital still

cameras, which typically have a maximum frame rate of01 Hz. These cameras offer better resolution and improved

picture quality compared with video. The analogue transferof video signals leads to image deterioration, and storage ontape creates line jitter. In contrast, digital still images under-go analogue digital conversion within the camera, prevent-ing the addition of noise during transfer and storage stages.The images analysed in this paper were captured using aKodak DC280 digital still camera, which has an imageresolution of 1760 3 1168 pixels.

PARTICLE IMAGE VELOCIMETRY (PIV)Theory

PIV is a velocity-measuring technique that was originallydeveloped in the field of experimental fluid mechanics(Adrian, 1991). The technique was originally implementedusing double-flash photography of a seeded flow. The result-ing photographs contain image pairs of each seed particle.For PIV analysis, the photograph is divided into a grid oftest patches. The displacement vector of each patch duringthe interval between the flashes is found by locating thepeak of the autocorrelation function of each patch. The peakin the autocorrelation function indicates that the two imagesof each seeding particle captured during the flashes areoverlying each other. The correlation offset is equal to thedisplacement vector.

A modified approach has been used to implement PIV ingeotechnical testing. Whereas fluid requires seeding withparticles to create features upon which image processing canoperate, natural sand has its own texture in the form ofdifferent-coloured grains and the light and shadow formedbetween adjacent grains when illuminated. Texture can beadded to an exposed plane of clay by the addition ofcoloured flock material or dyed sand.

The image processing conducted during PIV to measurethe displacement between a pair of digital images is shownschematically in Figs1 and2.The first image is divided intoa grid of test patches. Each test patch, Itest(U), consists of a

sample of the image matrix, I(U), of size L 3 L pixels.To find the displacement of the test patch between images

Image 1 (tt1)

Test patch, Itest(U)

u

Mesh of test

patches, for

which

displacement

vectors are

evaluated

L

Itest(U)

Isearch(U s)

s

smax

smax

Image 2 (tt2) Search patch,

Isearch(U)

V

Fig. 1. Image manipulation during PIV analysis

620 WHITE, TAKE AND BOLTON


3/13

1 and 2, a search patch Isearch(U + s) is extracted from thesecond image. This search patch extends beyond the testpatch by a distance smax, in the u and v directions, defining

the zone in which the test patch is to be searched for (Fig.1). The cross-correlation of Itest(U) and Isearch(U + s) isevaluated, and normalised by the square root of the sum ofthe squared values of Isearch(U + s) over the range of U

occupied by the test patch. The resulting normalised correla-tion plane Rn(s) indicates the degree of match between thetest and search patch over the offset range in the domain of

s (Fig. 2(a)). To reduce the computational requirement, thecorrelation operations are conducted in the frequency do-main by taking the fast Fourier transform (FFT) of eachpatch and following the convolution theorem.

1

0.98

0.96

0.94

0.92

0.930

2010

010

2030 30

2010

010

2030

Rn(s)

susv

(a) Enlargement of

correlation peak

1.01

1

0.99

0.98

0.97

0.96

0.95

0.942

1

0

1

2

1.5

0.5

0.5

1.52

1.51

0.50

0.51

1.52

Rn(s)

susv

(b)

1.01

1

0.99

0.98

0.97

0.96

0.95

0.942

1

0

1

2

1.5

0.5

0.5

1.52

1.5 1

0.50

0.51

1.52

Rn(s)

susv

(c)

Maximum

Displacement vector,

speakto sub-pixel

precision

Fig. 2. Evaluation of displacement vector from correlation plane, Rn(s): (a)correlation function Rn(s); (b) highest correlation peak (integer pixel); (c) sub-pixel interpolation using cubic fit over 61 pixel of integer correlation

SOIL DEFORMATION MEASUREMENT 621


4/13

The highest peak in the normalised correlation plane,Rn(s), indicates the displacement vector of the test patch,speak (Fig. 2(b)). The correlation plane is evaluated at singlepixel intervals. By fitting a bicubic interpolation to theregion close to the integer peak, the displacement vector isestablished to sub-pixel resolution (Fig. 2(c)). Generally, thebicubic interpolation function is evaluated at 1/200th pixelintervals, yielding a system resolution of 0005 pixels. To

improve the resolution, a smaller interval can be selected,with a corresponding increase in computational burden.However, a resolution greater than 0005 pixels is unneces-sary since in most cases the errors associated with accuracyand precision are larger.

The flowchart for PIV analysis of an image series isshown in Fig. 3. The procedure described above for evaluat-ing a single displacement vector is repeated for the entiregrid of test patches, producing the displacement fieldbetween the image pair. The analysis is continued by substi-tuting the second image (t t2) with a subsequent image(t t3). The analysis is repeated by comparing the first andthird images. As the analysis continues, distortion of theviewed soil or changes in lighting may reduce the sharpnessof the correlation peak that identifies the displaced locationof a given PIV patch. If the correct correlation peak isdrowned by random noise on the correlation plane anincorrect, or wild, vector will be recorded. To eliminatethis possibility, the initial image, in which the test patchesare established, is updated at intervals during the analysis.

Assessment of PIV performanceA series of experiments was conducted to assess the

precision of PIV. These experiments involved the controlledrigid-body movement of a planar body below a fixed camera.In each case, a mesh of test patches was established on theplanar body, and an image pair were compared. The randomvariation in the recorded displacement vectors indicates the

precision of the measurement technique; ideally all thevectors would be identical during a rigid body translation.

The series of experiments investigated the following influ-ences on PIV precision:

(a) test patch size(b) soil type/appearance(c) movement distance: whole or fraction of a pixel.

The first experiment, A, used an artificial image of soil,consisting of a matrix of randomly generated pixel intensi-ties, in the range 0 to 255. This random image allows thesoil movement to be exactly controlled, and makes itpossible to remove any image changes due to lighting orcamera shake. Fig. 4(a) shows a sample of this randomimage, with an enlargement of a 10 3 10 pixel test patch.

A series of seven PIV analyses were conducted, in whichthe entire 500 3 500 pixel random image was covered insquare test patches with side length, L, of 6, 8, 10, 16, 24,32 and 50 pixels respectively. The analysis was conductedby comparing the random image with itselfthat is, over arigid body displacement increment of zero pixels. Thescatter of displacement measurements recorded during com-

parison of a random image with itself is identical to thatwhich would be recorded had a new displaced image beencreated by cropping one side of the original image. Thecorrelation plane would be identical, but displaced by anamount equal to the artificial image shift.

Therefore the results of this initial experiment, althoughconducted over an image displacement of zero pixels, corre-spond to any integer pixel rigid-body translation of anartificial soil image, in which the brightness of each pixel israndom, and independent of its neighbours.

Each PIV analysis revealed a scatter of values, distributedclose to zero. Larger PIV patches produced less scatter, andtherefore improved precision. By plotting the standard devia-tion in measured displacement against test patch size, theinfluence of patch size on measurement precision can beseen (Fig. 5). Curve A shows the results from the seven PIV

Select test patch from

mesh in image 1

Evaluate cross-correlation of test

and search patch using FFT

Repeat for other test

patches in image 1

Repeat for subsequent

images in series

Use bicubicinterpolation to findsub-pixel location of

correlation peak

Normalise by cross-

correlation of search

patch with mask

Fig. 3. Flowchart of PIV analysis

0

100

200

0 100 2000

100

200

0 100 2000

100

200

0 100 2000

100

200

0 100 200

0

10

0 100

10

0 100

10

0 100

10

0 10

(a) (b) (c) (d)

Fig. 4. Images used in PIV validation experiments (axes in pixels): (a) random image; (b) random image,2 3 2 grains; (c) Dogs Bay sand; (d) kaolin clay with artificial texture



5/13

analyses of the random image, revealing a precision betterthan 001 pixels for patches greater than 8 3 8 pixels insize. The largest PIV patches, of size 50 3 50 pixels, weredistributed around zero with a standard error of 00007pixels. In order for these analyses to be a test of themeasurement precision rather than the measurement resolu-tion, the sub-pixel correlation peak was evaluated at intervalsof 1/10 000th of a pixel, rather than 1/200th.

These precision errors are associated with asymmetry ofthe correlation peak. In the case of this random image, theheight of the correlation peak prior to interpolation, ex-pressed mathematically as the value of Rn(s) for s 0, isunity, as an identical match for the test patch exists in thesearch patch. In the case of this entirely random image, thevalues of Rn(s) one pixel to either side of the peak lie inthe sea of random noise in which no positive correlationexists between the test patch and the underlying region ofthe search patch. If the values of Rn(s) on either side of thepeak were identical, the interpolated correlation peak wouldbe symmetrical and the sub-pixel peak would lie at theinteger value. In contrast, if these values differ, the inter-polated peak is asymmetric, and the small random errors

shown in curve A, and associated with measurement preci-sion, are created.

The discussion above demonstrates that a very high meas-urement precision can be obtained using a random soilimage, as a very sharp correlation peak is created. However,real images of soil are not entirely random. The brightnessesof neighbouring pixels are not independent, as they maydepict parts of the same grain. Expressed mathematically,the spatial brightness frequency of a soil image is lower.Neighbouring pixels are of similar colour.

A second experiment was conducted to examine the influ-ence of spatial brightness frequency. The random image usedin experiment A was artificially enlarged such that blocks of4 (2 3 2) pixels were of identical brightness. The resulting

image, of size 1000 3 1000 pixels, was analysed in the samefashion as described previously, using test patch sizes vary-ing from 6 3 6 to 50 3 50 pixels. A section of the image,and a typical 10 3 10 test patch, are shown in Fig. 4(b).

The results, shown as curve B in Fig. 5, revealed areduced precision, compared with the entirely random im-age. The precision errors associated with each patch size areincreased by a factor of 3, with 103 10 pixel patchesexhibiting a scatter of 00183 pixels. This reduced precisionis once again associated with the shape of the sub-pixelcorrelation peak. Although a perfect match exists at s 0,the adjacent values of Rn(s) are no longer in the random sea

of the correlation plane; some positive correlation exists atan offset of one pixel as the grains are now of size 2 3 2pixels. The resulting interpolated correlation peaks are lesssharp, are asymmetric, and lead to reduced precision.

Experiment C used a real image of Dogs Bay sand. Thisimage was taken at a nominal image scale of01 mm/pixel,giving a FOV of 176 mm3 112 mm (Fig. 4(c)). This FOVwas chosen as it is comparable to the model testing forwhich this system was developed. The D50 size of the sandwas found to be 044 mm by sieving.

In order to continue to isolate the influence of image typefrom possible precision errors due to non-rigid body motion,lighting changes or camera shake, experiment C comparedthe image of Dogs Bay sand with itself, for various patch

sizes (curve C in Fig. 5). A further reduction in precisioncompared with the random 2 3 2 grain image is evident.

Experiment D introduced the influence of random changesin appearance associated with taking multiple photos of atranslating object. These changes in appearance arise due tolighting changes and random variations in CCD sensitivity.The tray of sand was translated below the fixed camerausing a micrometer. A series of images was captured, atdisplacement increments corresponding to both integer andfractions of a pixel. Experiment D compared two images ofDogs Bay sand between which the tray of sand was trans-lated by (nominally) one pixel. Fig. 6(a) shows the entireimage, including the translating tray and micrometer, withthe mesh of 323 32 pixel test patches superimposed. The

measured displacement field is shown in Fig. 6(b).Curve D shows the standard error in recorded displace-

ment for each patch size. The mean value of recordeddisplacement was 10017 pixels; there was a discrepancy of

0.2

0.18

0.16

0.14

0.12

0.1

0.08

0.06

0.04

0.02

05 10 15 20 25 30 35 40 45 50

Patch size, L: pixels

Standarderror,Ppixel:pixels

UB, upper bound, equation (1)

Key

A. Random image

B. Random image, 22 grains

C. DBS, artificial movement

D. DBS, 1 pixel real movement

E. DBS, 0.365 pixel real movement

F. Textured clay, artificial movement

Z

Fig. 5. PIV precision against patch size



6/13

03 m between the desired one pixel movement andthe actual movement that was manually applied via themicrometer.

In experiment D, a reduction in precision compared withexperiment C was recorded. The random errors associatedwith multiple images increased the standard error for10 3 10 pixel PIV patches to 00460 pixels.

Experiments A D considered displacement incrementsequal (or very close) to an integer value. Experiment Ecompared two images between which the tray of sand wastranslated by 0365 pixels, to examine the precision asso-ciated with displacements of a fraction of a pixel. Curve E

shows the recorded errors, which indicate a further reductionin precision; the standard error for 103 10 pixel PIVpatches is 00642 pixels. These results indicate that precisionis reduced when interpolating far from an integer value;

greater reliance is placed on the sub-pixel bicubic interpola-tion.

Experiment F examined the possibility of imparting artifi-cial texture to clay, which is otherwise not suitable for imageanalysis, being of uniform brightness. Fig. 4(d) shows animage of kaolin clay dusted with fine dyed sand. This imagewas compared with itself, in the manner of experiment C forDogs Bay sand. The resulting precision (curve F) is compar-able to experiment C, indicating that the analysis techniqueis as applicable to artificially textured clay as to naturalsand.

This series of validation experiments reveals that the

precision of PIV is a strong function of patch size, L, and isalso influenced by image content. The empirically derivedcurve UB in Fig. 5is an upper bound on the precision error,rpixel, and is given by equation (1). This equation allows a

Micrometer Translating tray

Mesh of 3232 pixel

test patches

(a)

750

700

650

600

550700 750 800 850 900 950 1000 1050 1100 1150 1200 1250

Mean vector magnitude1.0017 pixels, standard deviation0.0100 pixels (point Z in Fig. 5)

(b)

Axes in pixels, vectors scaled20

Fig. 6. Precision validation experiment D: (a) translating tray containing Dogs Bay sand; (b) measureddisplacement vectors



7/13

conservative estimate of the random errors present in subse-quent PIV data to be made.

rpixel 0:6

L

150 000

L8 (1)

By dividing this error, rpixel, by the image width in pixels,the precision can be expressed as a fraction of the FOV.

The selection of an optimum patch size in PIV analysis

requires two conflicting interests to be balanced. Largerpatches offer improved precision (Fig. 5, equation (1)),whereas smaller patches allow the image to contain a greaternumber of measurement points, revealing detail in areas ofhigh strain gradient.

CLOSE-RANGE PHOTOGRAMMETRYTheory

Having obtained image-space deformation data using PIV,these measurements must be converted into object-spacecoordinates. The coordinate systems for image and objectspace are defined in Fig. 7. The conversion from pixelcoordinates into object-space position is represented by thetransformation U) X.

Previous image-based deformation systems used in geo-technics, with the notable exceptions of those described byTaylor et al. (1998) and Paikowsky & Xi (2000), have usedthe assumption that a single scale factor can be used toconvert between pixel coordinates and object-space position.This assumption is valid only if the camera behaves accord-ing to the pinhole model, the object plane is exactly parallelto the CCD plane, and the pixels are square. These assump-tions are never all valid in the case of short focal lengthdigital photography. The sources of image distortion, whichcause a spatial variation in image scale, are as follows.

Non-coplanarity of the CCD and object planes. If thenormals to the CCD and object planes are not parallel, a

distorted view of the object plane will be projected onto theCCD. When assembling an experiment, an approximatealignment can be achieved. However, compared with thesmall errors associated with PIV precision, a slightmisalignment of the camera can create a relatively large lossof accuracy. For example, if a camera with a 208 wide FOV ismisaligned by 18, the camera-to-object distance at the edge ofthe image will be reduced by 03%. In other words, a rigidbody motion corresponding to 1000 pixels at the imagecentre will be seen as 1003 and 997 pixels at the edges ofthe image.

In the case of centrifuge modelling, even if the cameracan be posed correctly prior to the experiment, the increasedself-weight of the camera and supporting frame may cause achange in pose. Take (2003) observed a 38 apparent rotationdue to distortion of the lens system of a digital still camerawhen tested at an acceleration of 100g.

Radial and tangential lens distortion. The pinhole camera

model, signified in Fig.7 by a bundle of rays passing straightthrough a single point to form a perspective projection of theobject on the CCD, is an approximation. Radial lensdistortion causes light rays to be deflected radially from thenormal to the lens(Slama, 1980), and is commonly known as

fisheye. A second error arises as the centres of curvature ofthe lens surfaces through which the light is refracted are notalways collinear. This is particularly the case for camerascontaining multiple lenses, such as compact digital cameras.This second error creates decentring distortion, which has aradial and tangential component. The tangential component isknown as barrelling.

Refraction through a viewing window. An additional

source of image scale variation arises when the object liesbehind a viewing window. An apparent variation in objectsize arises, depending on the thickness and refractive index ofthe window, and the inclination of the incident rays.

CCD pixel non-squareness. A final source of image scalevariation occurs because CCD pixels are not perfectly square.Although this is a small correction, it can be incorporated asa linear scaling factor in the transformation from pixelcoordinates to image-space coordinates. The pixel aspectratio, , is defined as the height of a pixel divided by thewidth: it is typically in the range 1 0004 (Heikkila &Silven, 1997;Ahmad & Chandler, 1999), and is considered tobe constant over the CCD.

U) XtransformationThe principles of close-range photogrammetry have been

used to develop a more robust transformation U) X. Eachof the sources of image distortion described above is ex-plicitly modelled, yielding a 14-parameter transformationfollowing Heikkila & Silven (1997), which has been ex-tended to include distortion due to observation through aviewing window. The 14 parameters are as follows:

(a) camera pose: six parameters to describe the translation

Object space coordinates

(X,Y)

Image space pixel

coordinates (on CCD),U(u,v) Image space reference frame,x

x

yz

Object space reference frame,X

X

Y

Z

y

x

z

Z

Y

X

Fig. 7. Object-space and image-space coordinate systems



8/13

and rotation between the image-space, x, and object-space, X, coordinate systems.

(b) focal length, f(c) principal point, (u0, v0): the pixel coordinates of the

intersection of the optical axis and the CCD plane(d) radial lens distortion (two parameters, k1, k2)(e) tangential lens distortion (two parameters, p1, p2)(f) CCD pixel squareness ().

In addition to these parameters, the refractive index andthickness of any viewing window(s) are included in thetransformation. Snells law is used to model refraction.

In order to convert a set of image-space coordinatesdeduced using PIV into object-space coordinates, the 14transformation parameters are required. These are found bylocating the image-space position of a set of referencetargets whose object-space coordinates are known. From thisset of known U and X coordinates, the transformationparameters can be optimised.

Transformation parameters that do not vary over time, orwhen the camera is moved between viewing positions, canbe found during a single calibration experiment, separatefrom the experiment in which deformation measurements areto be gathered. Taylor et al. (1998) observed that theintrinsic camera parameters (relating to principal point, focallength, and radial and tangential lens distortion) of analogueminiature video cameras do not vary even during centrifugetesting. However, the digital still cameras used in thisresearch suffer significant internal distortion under centrifugetesting. Therefore it proved necessary to deduce all 14transformation parameters on an image-by-image basis.

The procedure for calibrating a set of image-space coordi-nates found from PIV is as follows.

Reference target identification. The image-space coordi-nates of a small number of reference targets, whose object-space coordinates (relative only to each other) have been

deduced prior to the experiment, are identified. Thisidentification is carried out by multiple-threshold centroiding(Take, 2003), yielding a precision better than 01 pixels for atypical 20 3 20 pixel reference target. Having identified thelocation of these targets in the initial image, their subsequentmovement is found using PIV.

Transformation optimisation. As the refraction and lensdistortion models contain non-linear components, the U) Xtransformation must be optimised iteratively. A first estimateof the linear parameters is made using a direct lineartransformation (Abdel-Aziz & Karara, 1971). These para-meters are used as initial values for a LevenbergMarquardtnon-linear optimisation, using the solution method describedby Heikkila & Silven (1996). A correction vector masking

the influence of refraction is then applied to the first non-linear solution. Improved transformation parameters arefound, and the process is repeated until steady values arereached.

Conversion of PIV measurements from image space (U) toobject space (X). The deduced transformation parameters areused to convert the image-space coordinates into object-spacecoordinates. After removing the non-linear lens distortion, alinear camera model is used to establish the object-spacemeasurement coordinates in the absence of refraction. Finally,having established the pose of the camera, a correction vectorto account for refraction is applied.

Validation of photogrammetric calibration procedureThe photogrammetric calibration procedure governs the

accuracy of this measurement systemthat is, the systematic

difference between the measured and true coordinates. Asimage-based measurement systems report an initial positionof an object, and its displacement vector between an imagepair, two types of accuracy can be defined. Positionalaccuracy is the difference between the reported initial posi-tion of an object and its true value. Movement accuracy isthe difference between the reported displacement vector ofan object and its true value.

A series of validation experiments were conducted toassess both the positional accuracy and the movement accu-racy of this system. Positional accuracy was assessed bycomparing the systematic difference between the true coordi-nates of a grid of reference targets, and the coordinatescalculated from reconstruction of the measured image-spacevalues. A certified photogrammetric reference field consist-ing of three regular grids of dots provided a reference fieldof known object-space coordinates (Fig. 8).

The position of each dot within the reference field waslocated in image space by multiple-threshold centroiding(Take, 2003). Knowing the image- and object-space loca-tions of each dot, the optimal photogrammetric transforma-tion parameters were found. Using these parameters, theobject-space coordinates of each dot were estimated fromthe image-space values, and the discrepancies between theseand the true values were found.

Three different images were reconstructed by this method:

Image A: A view of the photogrammetric target with thecamera axis nominally normal to the target plane(Fig. 8).

Image B: A view of the photogrammetric target with thecamera axis nominally normal to the target plane,seen through a 72 mm layer of Perspex.

Image C: A view of the photogrammetric target with thecamera inclined at 208 from normal to the targetplane.

In addition, each image was reconstructed using a constantscale factor, as a simple alternative to photogrammetriccorrection. This is the conventional calibration method thathas been used by previous image-based deformation meth-ods, with the exception of Taylor et al . (1998) andPaikowsky & Xi (2000). The constant scale factor waschosen such that the mean discrepancy between the meas-ured and actual coordinates was zero, and the standarddeviation was minimised.

Table 1 shows the standard deviation of the discrepancyvectors between the measured and actual coordinates. Thesevalues demonstrate that the use of photogrammetric calibra-tion improves positional accuracy compared with the use ofa constant image scale factor. The transformation U ) Xfor image A is illustrated in Fig. 9by the spatial variation in

image scale in the X and Y directions. This curved surfacedemonstrates that the apparently regular grid of dots inimage A (Fig. 8) contains systematic distortion, which canbe captured by the photogrammetric transformation. Thebowl-shaped curvature in Fig. 9 arises from radial lensdistortion; the asymmetry indicates non-coplanarity of theimage and object planes. This non-coplanarity is presentdespite careful attempts to position the camera axis ortho-gonal to the target plane.

If the discrepancy vectors for close range photogrammetrysummarised in Table 1 are plotted on the original image,their spatial distribution is random, rather than systematic. Asystematic distribution of discrepancy vectors would implythat the transformation function U ) X, which is smooth

and continuous and is illustrated by the curved image scalesurface in Fig. 9, is unable to shape itself appropriately. Thiswould arise if the optical behaviour of the camera was notcaptured correctly in the U) X transformation.



9/13

Array of 480 3 mm diameter dots, located at 6 mm centres

Y

X

Fig. 8. Image A: photogrammetric target

Table 1. Standard deviation (m) between actual and reconstructed location of reference dots

Reconstruction method Image A Image B Image C

Direction X Y X Y X YClose-range photogrammetry 74 86 88 90 94 73Linear scaling 251 332 194 249 2464 1288

7.46

7.44

7.42

7.4

7.38

7.36

7.34

7.32

7.3

140120

10080

6040

20

0 50

050

100

150

Y-coordinate: mmX-coordinate: mm

Image scale inX-direction: pixels/mm

7.46

7.44

7.42

7.4

7.38

7.36

7.34

7.32

7.3

140120

10080

6040

20

0 50

0

50

100

150

Y-coordinate: mmX-coordinate: mm

Image scale in Y-direction: pixels/mm

Fig. 9. Image A: spatial variation in image scale



10/13

Linear scaling, for example, could be depicted by ahorizontal surface in Fig. 9, leading to a systematic spatialdistribution in discrepancy between the assumed image scaleand the actual image scale, and hence a systematic spatialdistribution in positional error.

The random spatial distribution of the discrepancy vectorsindicates that the discrepancy errors do not arise from thephotogrammetric calibration routine. Instead, the errors lie in

the underlying data of the image- and object-space locationsof the reference dots. Therefore the positional accuracy ofthis system is not limited by the calibration routine, but byeither the multiple-threshold centroiding method used tolocate the reference targets or the accuracy of the photo-grammetric reference field itself.

The movement accuracy of the system is defined as thedifference between the true and measured values of anobject-space displacement vector of a single patch foundfrom comparison of two images. This discrepancy arisesfrom the error in the PIV image-space displacement meas-urement (estimated from equation (1)) and the change in theU ) X transformation created by the error in the measuredmovement of the reference markers (which are also foundusing PIV, and can be estimated from equation (1)). Thesetwo additive errors can be summed geometrically. In thecase of L 32, which corresponds to the patch size used totrack the 3 mm dots in Fig. 8, these errors sum to 00265pixels, or 1/66374th of the FOV.

This definition of movement accuracy is supported by afurther validation exercise. The photogrammetric referencefield shown in Fig. 8 was translated over a distance of 6pixels below a stationary camera. A subset of 20 dots fromthe array of 480 3 mm diameter dots was selected torepresent reference markers in a typical experiment. Thesereference markers are used to define the object-space refer-ence frame. As the photogrammetric target was translated,not distorted, the movement of the 460 remaining dotsrelative to the 20 reference dots should be equal to zero.

The movement of each of the 480 dots was tracked usingPIV patches of size 323 32 pixels, yielding initial anddisplaced image-space coordinates for each dot. The subsetof 20 dots was used to evaluate the photogrammetric trans-formation for the initial and displaced images, and thesetransformation parameters were used to evaluate the object-space displacement vectors of the remaining dots. The meandiscrepancy between the true displacement of zero and themeasured values was found to be 2 m. Noting that the fieldof view was 240 mm, this error corresponds to a normalised

movement accuracy better than 1/100 000th. This accuracyexceeds the value predicted by equation (1), confirming thatthis offers a conservative estimate of measurement error.

The observation that movement accuracy (that is, theaccuracy by which displacement vectors are evaluated) canexceed positional accuracy can be illustrated as follows. Theactual and measured locations of a single stationary refer-ence target and a nearby soil element in two images are

shown schematically in Fig. 10. As discussed earlier, theerrors associated with reference target location using multi-ple threshold centroiding are larger than the random errorsin displacement measurements made using PIV. Fig. 10shows a discrepancy between the actual and measured initialcoordinates of the reference marker, introduced by thecentroiding procedure. This discrepancy vector is carriedthrough to the transformation U) X, leading to a similarpositionalerror in locating the nearby soil element.

However, the error in the measured movement of the soilelement is smaller. The movements of the reference markerand the soil element between the initial and displacedimages are measured using PIV. The discrepancy vectorintroduced when locating the reference marker is carriedthrough unchanged, contributing zero error over the move-ment step. Therefore the error in the movement vector arisesonly from PIV measurements of the soil and referencemarker, which can be summed geometrically. The initialerror introduced while locating the reference target does notinfluence the measured displacement vector.

System performanceThe overall performance of this measurement system can

be quantified by

(a) precision(b) accuracy (positional and movement)(c) the size of the array of measurement points that can be

contained within a single view.The precision of this system derives from the PIV technique,is a strong function of PIV patch size, L, and can beconservatively estimated from equation (1).

The positional accuracy of the system depends on themethod used to establish the image-space location of thereference targets. Multiple-threshold centroiding allows theimage-space location of a typical reference target (303 30pixels) to be established to an accuracy better than 01pixels. For the 2-megapixel digital still camera used in this

Grid of PIV patches

Grid of reference markers

Soil element

Stationary

reference target

Initial image, tt1

Displaced image, tt2

Actual position

Measured position

Measured displacement

vector Xmeasured

Movement errorxmeasuredXactual

Actual displacement

vector Xactual

Positional error

Displaced location

introduces error arising

from PIV

Initial location is subject

to error arising from

centroiding

Fig. 10. Positional and movement accuracy



11/13

250000

200000

150000

100000

50000

00 10000 20000 30000 40000 50000 60000Measurement point array size

1/Measurementprecision(expressedasafactionoftheFOV)

Greater precision

CCD size (106pixels)

2

3

4

5

6 More measurement points

Patch size, L

6 6

10 10

24 24

50 50

Performance achievableusing film/video/target markers

Fig. 11. Precision against measurement array size

Fig. 12. Displacement field around plane strain displacement pile in sand: (a) mesh of PIV patches; (b) displacement vector field; (c)magnitude of displacement vectors

0

200

400

600

800

1000

0 400 800 1200 1600

Horizontal pixel coordinate,v

Verticalp

ixelcoordinate,u

(a)

120

100

80

60

40

20

00 50 100 150 200

Horizontal coordinate: mm

(b)

Verticalcoordinate:mm

1 mm displacement 5 mm displacement

120

100

80

60

40

20

00 50 100 150 200

Horizontal coordinate: mm

(c)

Verticalcoord

inate:mm

0.025

0.025

0.025

0.025

0.025

0.05

0.05

0.05

0.05

0.075

0.075

0.1

0.1

0.1

0.1

0.1

1.5

1.41.3

1.21.1

0.90.8

0.7 0.6

0.4

0.4

0.4

0.5

0.3

0.250.2

0.15



12/13

research (17603 1168 pixel resolution), this accuracy corre-sponds to 1/17 600th of the FOV.

The movement accuracy of the system derives from theerrors of the PIV operations used to (a) track the soil inimage space, and (b) track the reference targets from whichthe image-space to object-space transformation parametersare derived. In both cases the errors are a strong function ofthe PIV patch size (equation (1)). Geometric summation

of the errors yields an estimate of the movement accuracy.This estimate agrees closely with the results of a validationexercise.

The number of measurement points that can be estab-lished in an image is a function of the PIV patch size.By using smaller PIV patches, the measurement arraysize can be increased, at a cost of reduced precision. Fig.11 shows the relationship between precision and measure-ment array size for different resolutions of camera CCD,based on the empirical estimate of precision provided byequation (1).

EXAMPLE APPLICATIONSThe performance of this system is illustrated with two

example applications. Fig. 12(a) shows a view through thewindow of a plane-strain calibration chamber used to simu-late the installation of a displacement pile (White, 2002). Amesh of 3201 PIV patches overlies the soil below the tip ofthe 322 mm wide model pile. Visible in the image are the14 reference targets used to optimise the photogrammetrictransformation parameters. This image is compared with a

subsequent image captured after 15 mm of pile movement.PIV analysis of the two images followed by photogrammetriccalibration reveals the displacement field shown in Fig.12(b).

To reveal the zone of deformation, contours of displace-ment vector magnitude are shown in Fig. 12(c). The finemesh of PIV patches is able to capture the high velocitygradient close to the pile tip, while also revealing the very

small movements in the far field.In the second example application, the imaging system isused to observe the onset of progressive failure in a modelclay embankment (Take, 2003). Of the three cameras usedto record the displacement fields, only the camera focusedon the crest of the embankment is presented in Fig. 13(a).This image, taken moments before gravity turn-on, is differ-ent from those captured at elevated acceleration levels. Inthe centrifuge environment, the camera lens acts as acantilever, which deflects according to its increased self-weight. Thus the downwards rotation of the camera lenswith respect to the object plane results in an apparent largelyupwards displacement of the PIV patches, in this caseapproaching 100 pixels for the gravity turn-on to 60g (Fig.13(b)). PIV analysis of the 20 reference targets provides thebasis to extract the soil movement from the camera move-ment and distortion by tracking the displacement vectors ofthe stationary targets (Fig. 13(c)). Finally, photogrammetriccalibration reveals the hidden soil movement during consoli-dation (Fig. 13(d)) necessary to describe the initial state ofthe model, which had been dwarfed by the gravity-inducedchanges in the calibration parameters.

100

200

300

400

500

600

700

800

900

1000

1100

200 400 600 800 1000 1200 1400 1600

Horizontal pixel coordinate,u

(a)

Verticalpixelcoordinate,v

0

200

400

600

800

1000

200 400 600 800 1000 1200 1400 16000

Horizontal coordinate; mm

(b)


100 pixels

0

200

400

600

800

1000

200 400 600 800 1000 1200 1400 16000


(c)


100 pixels

280

260

240

220

200

180

160

140

0 50 100 150 200


(d)

Verticalcoordinate:mm

10 mm

Fig. 13. Determination of displacement vectors during gravity turn-on to 60g using close-range photogrammetry: (a) mesh of PIVpatches at crest of model embankment; (b) observed image-space vector field during gravity turn-on to 60g; (c) observed image-spacedisplacements of stationary reference markers; (d) actual object-space consolidation settlement



13/13

CONCLUSIONSA new system for measuring deformation in geotechnical

tests based on digital photography, particle image veloci-metry (PIV) and close-range photogrammetry is presented.Compared with previous image-based methods of displace-ment measurement, this system offers an order-of-magnitudeincrease in accuracy, precision, and measurement array size.This performance is achieved using a relatively inexpensive

(,

300) 2-megapixel digital still camera. In the future, theavailability of higher-resolution CCDs at lowercostwill leadto corresponding increases in performance (Fig. 11).

REFERENCESAbdel-Aziz, Y. I. & Karara, H. M. (1971). Direct linear transfor-

mation into object-space coordinates in close-range photo-grammetry. Proceedings of the symposium on close-range

photogrammetry, Urbana, Vol. 1, p. 118.Adrian, R. J. (1991). Particle imaging techniques for experimental

fluid mechanics. Ann. Rev. Fluid Mech. 23, 261304.Ahmad, A. & Chandler, J. H. (1999). Photogrammetric capabilities

of the Kodak DC40, DCS420 and DCS460 digital cameras.Photogramm. Rec. 16 , No. 94, 601615.

Andrawes, K. Z. & Butterfield, R. (1973). The measurement ofplanar displacements of sand grains. Geotechnique 23, No. 5,571576.

Butterfield, R., Harkness, R. M. & Andrawes, K. Z. (1970). Astereo-photogrammetric technique for measuring displacement

fields. Geotechnique 20, No. 3, 308314.Heikkila, J. & Silven, O. (1996). Calibration procedure for short

focal length off-the-shelf CCD cameras. Proc. 13th Int. Conf.Pattern Recognition, Vienna, 166170.

Heikkila, J. & Silven, O. (1997). A four-step camera calibrationprocedure with implicit image correction. Proc. IEEE ComputerSoc. Conf. Computer Vision and Pattern Recognition(CVPR97), San Juan, 11061112.

Paikowsky, S. G. & Xi, F. (2000). Particle motion tracking utilizing

a high-resolution digital CCD camera. ASTM Geotech. Test. J.23, No. 1, 123134.

Phillips, R. (1991). Film measurement machine user manual.Cambridge University Technical Report, CUED/D-Soils/TR246.

Roscoe, K. H., Arthur, J. R. F & James, R. G. (1963). Thedetermination of strains in soils by an X-ray method. Civ. Engng

Public Works Rev. 58, 873876, 10091012.Scholey, G. K., Frost, J. D., Lo Presti, D. C. F. & Jamiolkowski, M.

(1995). Review of instrumentation for measuring small strainsduring triaxial testing of soil specimens. ASTM Geotech. Test. J.18, No. 2, 137156.

Slama, C. C. (ed.) (1980). Manual of photogrammetry, 4th edn.American Society of Photogrammetry, VA, USA.

Take, W. A. (2003). The influence of seasonal moisture cycles onclay slopes. PhD dissertation, University of Cambridge.

Taylor, R. N., Grant, R. J., Robson, S. & Kuwano, J. (1998). Animage analysis system for determining plane and 3-D displace-ments in soil models. Proc. Centrifuge 98, 7378. Tokyo.

White, D. J. (2002). An investigation into the behaviour of pressed-in piles. PhD dissertation, University of Cambridge.


2003 - Soil Deformation Measurement Using Particle Image Velocimetry PIV by WHITE

Documents

Transcript of 2003 - Soil Deformation Measurement Using Particle Image Velocimetry PIV by WHITE