Home

Search

Collections

Journals

About

Contact us

My IOPscience

Three-dimensional particle image velocimetry: experimental error analysis of a digital angular

stereoscopic system

This content has been downloaded from IOPscience. Please scroll down to see the full text.
1997 Meas. Sci. Technol. 8 1455
(http://iopscience.iop.org/0957-0233/8/12/009)
View the table of contents for this issue, or go to the journal homepage for more

Download details:
IP Address: 140.254.87.149
This content was downloaded on 20/09/2014 at 12:55

Please note that terms and conditions apply.

Meas. Sci. Technol. 8 (1997) 14551464. Printed in the UK

PII: S0957-0233(97)86173-5

Three-dimensional particle image

velocimetry: experimental error
analysis of a digital angular
stereoscopic system
N J Lawson and J Wu
The GK Williams CRC, Department of Chemical Engineering,
The University of Melbourne, Parkville, Victoria 3052, Australia
CSIRO, Division of Building, Construction and Engineering, Graham Road,
Highett, Victoria 3190, Australia
Received 22 July 1997, accepted for publication 25 September 1997
Abstract. Experimental error analysis of a digital angular stereoscopic PIV system
is presented. The paper firstly describes an experimental rig which includes the
design of a novel PIV test block for in situ calibration. This allowed the user to set
up a static seeded flow volume which was translated in and out of plane to record
PIV images using two megapixel CCD cameras positioned for angular stereoscopic
viewing. PIV data were collected for a range of camera angles up to 45 and for a
range of flow displacements and processed by cross correlation into a set of
two-dimensional calibration and flow displacement vectors. These 2D data were
then processed into three-dimensional data by the use of geometric and bicubic
spline interpolation algorithms and an error analysis performed on the predicted
displacements. Results from this analysis have shown optimum system
performance will be obtained by using camera angles of between 20 and 30 and f
numbers of f 16 and higher. The results have also shown a theoretical prediction of
system performance derived in previous work, which considers the ratio of out of
plane to in plane errors, matches to within 8 and 18% of the experimental system
performance.

1. Introduction
The measurement of fluid flows in engineering has
continually generated innovative techniques varying from
a simple Pitot tube [1] to complex laser techniques [2].
Particle image velocimetry (PIV) is a recent but well
established, non-intrusive technique for the measurement
of mean and instantaneous fluid velocity from a single
plane of interest [3, 4]. PIV in its simplest form involves
placing a pulsed laser sheet into the measured fluid flow
plane and recording the position of seeding particles by
the use of double-exposure photography. This results in a
transparency consisting of a large number of particle image
pairs whose average displacement over each small region
of the photograph is a measure of the mean velocity at
the corresponding point in the fluid. From the developed
transparency, a two-dimensional (2D) velocity map is then
extracted by using optical or digital spatial correlation
analysis to determine the particle displacement within each
small region of the flow [5, 6].
The application of PIV to increasingly complex
fluid flows [7, 8] has resulted in the demand for threedimensional (3D) measurements and the development
c 1997 IOP Publishing Ltd
0957-0233/97/121455+10$19.50

of 3D PIV techniques. A recent review of 3D PIV

highlighted the different methods that had been successfully
demonstrated at that time [9].
These were mainly
holographic PIV (HPIV) [10, 11] and stereoscopic PIV
methods [1216]. More recent methods have included
dual-plane PIV [17, 18] and off axis stereoscopic PIV
[19]. If these techniques are considered in terms of their
imaging and data processing systems, HPIV is the most
demanding in both cases requiring complex optics and
advanced data processing techniques. In contrast dualplane PIV and stereoscopic PIV only require conventional
cameras and light sheet arrangements although intensive
image processing is necessary.
Any PIV system will have a level of accuracy
which will determine the validity of its application.
Previous authors have investigated in detail the error
characteristics of a simple 2D PIV system [2022]. To
date, however, limited work has been carried out on the
error characteristics of 3D PIV systems and has mainly
been restricted to stereoscopic configurations [12, 13, 19].
This previous work has recently been extended by the
present authors to give a set of theoretical design curves for
stereoscopic systems [23]. These characteristics allowed
1455

N J Lawson and J Wu

Figure 1. Diagram of experimental stereoscopic imaging system.

the overall performance of an angular or translational

stereoscopic system to be analysed for a range of camera
angles and separations.
In what follows, experimental work is presented which
aims to validate the theory outlined in the previous work
[23]. The analysis is based on a digital angular stereoscopic
system which is used to investigate the system performance
over a range of camera angles and fluid displacements.
These experimental results are then directly compared to
the theoretical predictions and any discrepancies discussed.
2. Experimental technique
The following section describes the stereoscopic imaging
system and data acquisition methods used to obtain the
experimental PIV data. The system was based on a
direct calibration method by using a design of a PIV test
block developed for the project. This method assumed
acceptable linearity between the object and image planes
by using an appropriate object distance and resulted in
a much simplified image processing routine which could
use conventional correlation and interpolation routines
for the calculations. This section also describes the
testing procedures adopted and analysis of the error ratio
characteristics from experimental error data.
2.1. Imaging system
Figure 1 illustrates the stereoscopic imaging system used
to record the PIV data. In the imaging plane two Kodak
1456

MegaPlus CCD cameras were mounted onto separate

rotation stages at an angle to the z axis where 1 =
and 2 = +. These stages were then both mounted onto
two x translation platforms which were secured onto one
end of an optical breadboard at an appropriate separation.
This system allowed fine adjustment of the camera positions
to achieve camera angles up to 45 , camera separations
in x up to 600 mm and object distances up to 800 mm.
Resolution of the translation stages was 10 m and the
rotation stages 1/10 of a degree. Both cameras had a
resolution of 12801024 pixels with 8-bit greyscale output
and an array size of 10 mm 8 mm. All the PIV images
were recorded using 60 mm Nikon Micro-Nikor lenses. A
lens f number of f 16 was used for camera angles up to
= 30 and f 22 for any camera angles above this. These
f numbers were based on recommendations for depth of
field by Adrian [3].
The object plane consisted of an xz translation stage
centred about the origin O as shown in figure 1 and a light
sheet system connected to a 4 W argon ion light source
through a fibre-optic delivery. On top of the translation
stages a fixed bracket was used to mount the calibration
grid or PIV test block and provided exact alignment with
the laser sheet for each set of measurements. The light sheet
system was coupled to the fibre optic through a mounting
block which contained a collimation lens and glass rod.
This allowed adjustment of the light sheet divergence and
the light sheet thickness, w, which was eventually set to
2.5 mm for all imaging. The object distance, d0 , as defined
in figure 1 was set to 380 mm for all camera angles. Both

3D PIV

2.2. PIV test blocks

(a)

(b)
Figure 2. (a) Left-hand CCD image of calibration grid.
(b) Right-hand CCD image of calibration grid ( = 45 ).

Previous PIV test methods have used glass plates

coated with seeding and adhesive with conventional flood
illumination for imaging [13, 21].
In the following,
however, an alternative test method is presented based
on a design of PIV test block. This method allows
a PIV light source to be used and provides realistic
seeding characteristics such as those found in the measured
fluid. The technique also allows the recording of a
combination of in and out of plane data and can be
adjusted to provide a wide range of particle image sizes and
particle seeding densities if required. In addition the test
blocks introduce imperfections such as background scatter
and image aberrations that are normally present on PIV
recordings but which are not normally present in computer
simulated data.
The PIV test blocks consist of hollow glass spheres
suspended in solid clear araldite blocks. The araldite blocks
were manufactured by permanently setting Ciba Geigy
Araldite resin M with hardener HY6150 into open-top glass
containers. The glass containers, with internal dimensions
20 mm 350 mm 250 mm, were constructed from
2 mm thick sheet glass pieces adhered together with silicone
sealant. This method gave acceptable viewing properties
with good non-scratch characteristics. Air bubbles which
formed in the resin during mixing with the hardener were
removed by placing the mould in a 100 kPa vacuum for
one hour before allowing the block to set.
The particles used to seed the test blocks were Potters
Ballotini SphericelTM hollow glass spheres. They were
chosen because of their excellent scattering properties in
the resin. The original 45100 m diameter range was
wire mesh sieved down to a 4560 m size range before
seeding the resin so as to improve the homogeneity of the
sample. Approximate seeding densities were determined
by the amount of seeding weight added to the resin before
moulding.
2.3. Calibration and data acquisition

Figure 3. Typical PIV image from PIV test block ( = 45 ).

these constants ensured that equal displacements of the test

block in the x and z axes would produce equal percentages
of movement in terms of a percentage of the interrogation
region length d, or a percentage of the light sheet width w
at the origin O.

A calibration stage is necessary for stereoscopic systems

since the imaging geometry generally causes significant
aberrations [12, 13, 16]. The calibration procedure will
output a space and calibration coefficient map across
the image where the calibration coefficients are related
to the spatial distortion of the image. This distortion
becomes significant when imaging inside a wide fluid
layer and complex stereoscopic equations can be developed
to output the calibration coefficients [12]. In this case,
however, a simpler direct in situ calibration method was
developed based on the PIV test blocks and which relied
on acceptable linearity between the object and image plane
by using an appropriate object distance. To this end
an object distance of 380 mm was selected and linearity
tested at the calibration stage for a range of PIV test
block translations. This approach resulted in a basic
image processing algorithm which used readily available
correlation and interpolation routines.
For calibration a grid was placed in the centre of the
light sheet or object plane. This grid allowed alignment of
1457

N J Lawson and J Wu

the camera positions in the x plane with a grid origin so

that both cameras were imaging a common object plane.
Figures 2(a) and 2(b) illustrate the left and right calibration
grid images which have 5 mm grid spacing in x and y
and where the grid origin can clearly be seen. The grid
images also allowed calibration of the x and y axis space
coordinates by using the data processing software described
in the next section. To obtain the calibration coefficients
two exposures were recorded from the PIV test block with
translations between exposures of 200800 m in the x or
z axes. These displacements corresponded to 7.530% of
the interrogation region length d at O or the light sheet
width w. The y axis was not required for this analysis.
Figure 3 illustrates a typical PIV image obtained from the
test blocks. These data were subsequently processed into x
and z calibration coefficients by cross correlation software
described in the next section.
For test data acquisition a similar procedure was used
as for the x and z calibration coefficients. In this case,
however, in order to obtain suitable data for the error
analysis, the block was translated by an equal amount in
both the x and z axes. As before the translations chosen
were between 200 and 800 m in z and x axes which
corresponded to a range of 7.530% of the interrogation
region length d at O and the light sheet width w.
2.4. Data processing
Figure 4 illustrates the data processing procedure used
to calculate the x, y and z components of particle
displacement 1x, 1y and 1z. The procedure can be split
into three major parts. Each part uses simple correlation and
interpolation algorithms due to the assumption of linearity
between the object and image planes and the decoupling of
the x and z characteristics. Three data processing routines
were developed to perform the calculations for each part.
The first two involved calibration of the system and the
third the calculation of the three-dimensional displacement
components. The y axis was not required for the error
analysis but the relevant equations will still be listed for
completeness.
In the first routine calibration grid image data were
initially processed into the X and Y pixel coordinate
functions X1 (x, y), X2 (x, y) and Y1 (x, y), Y2 (x, y). These
functions were produced by performing centroid analysis on
j
the image across an Mg Ng grid at points (X1i , Y1 ) and
j
i
(X2 , Y2 ) where i and j are the gridpoint numbers. These
centroid pixel positions were then fed the into a bicubic
spline interpolation routine ([24], pp 99122). The actual
positions of x and y in space were determined from the
gridpoint data and spacing. A threshold was also used to
eliminate background image noise with a value set from
a histogram analysis of the images. The resulting bicubic
spline interpolation functions were used in all subsequent
image processing to reference any image pixel position X1 ,
Y1 and X2 , Y2 into an actual space position x and y and
vice versa.
The second processing routine was used to produce the
x, y and z calibration coefficient functions from the PIV
calibration images. This was done by correlating the pairs
1458

of images on a regularly spaced i, j image grid using the

cross correlation functions
Z Z +
R1 (, )i,j =
G1 (X1cc , Y1cc )i,j

H1 (X1cc , Y1cc )i,j dX1cc dY1cc

and

ZZ

R2 (, )i,j =

G2 (X2cc , Y2cc )i,j

H2 (X2cc , Y2cc )i,j dX2cc dY2cc

where G1 (X1cc ,
Y1cc )i,j , H2 (X2cc ,

(1)

(2)

Y1cc )i,j , G2 (X2cc , Y2cc )i,j and H1 (X1cc ,

Y2cc )i,j are the first- and second-exposure

interrogation region images respectively and R1 (, )i,j ,

R2 (, )i,j are the correlation outputs at gridpoint (i, j )
[25]. The correlation routine was based on a fast Fourier
transform (FFT) algorithm ([24], pp 530602). A centroid
routine was performed on the three highest peaks in the
correlation plane and the peak positions stored for postprocessing. The interrogation region resolution was set
to 64 64 pixels with a 50% overlap and a 4 4 pixel
centroid sample to provide sufficient accuracy [26]. The
post-processing initially involved validating the vectors
using a range set by visual inspection of the output vector
maps. Finally the x and z calibration data were calculated
by taking the correlation vector component, R1 (, 0)i,j
or R2 (, 0)i,j , and interpolating and smoothing to reduce
correlation noise as recommended by previous workers
[27]. For the y calibration the Y component vector
was taken, R1 (0, )i,j or R2 (0, )i,j , and interpolated and
smoothed as before. To convert these data into the x, y
and z calibration coefficient functions the correlation data
were initially converted to a pixel mm1 quantity at each
j
j
grid point (X1i , Y1 ) and (X2i , Y2 ) such that:
 i=1,Mcal
R1 (, 0)i,j  j =1,Ncal
=
1xcal 

(3)

 i=1,Mcal
R2 (, 0)i,j  j =1,Ncal
=
1xcal 

(4)

 i=1,Mcal
R1 (, 0)i,j  j =1,Ncal
=
1zcal 

(5)

 i=1,Mcal
R2 (, 0)i,j  j =1,Ncal
1zcal i,j

(6)

 i=1,Mcal
R1 (0, )i,j  j =1,Ncal
1ycal 

(7)

 i=1,Mcal
R2 (0, )i,j  j =1,Ncal
1ycal i,j

(8)

j
a1 (X1i , Y1 )

j
a2 (X2i , Y2 )

j
b1 (X1i , Y1 )

b2 (X2i , Y2 ) =
c1 (X1i , Y1 ) =
c2 (X2i , Y2 ) =

where Mcal , Ncal is the grid size for the a, b and c

coefficients and 1xcal , 1ycal and 1zcal are the PIV
test block calibration displacements in the x, y and
z axes. These data were fed into a bicubic spline
interpolation routine to produce the calibration coefficient
functions a1 (X1 , Y1 ), a2 (X2 , Y2 ), b1 (X1 , Y1 ), b2 (X2 , Y2 )
and c1 (X1 , Y1 ), c2 (X2 , Y2 ) which allow estimates of the

3D PIV

Figure 4. Illustration of data processing procedure used to calculate 1x , 1y and 1z particle displacements.

x, y and z calibration coefficients at any pixel position on

the image. All the a1 , a2 , b1 , b2 , c1 , c2 coefficients and
X1 , X2 , Y1 , Y2 interpolation functions were subsequently
stored onto a batch file for use in the third and final data
processing routine.
The third processing routine was used to calculate threedimensional data from actual PIV flow data or in this case
test data recorded from the PIV test blocks. In the first stage
of the calculation, the two sets of left and right flow images
were cross correlated, validated and smoothed using a
regularly spaced i, j grid as described previously and stored
j
j
into the displacement arrays, 1X1 (X1i , Y1 ), 1Y1 (X1i , Y1 )
j
j
and 1X2 (X2i , Y2 ), 1Y2 (X2i , Y2 ) where
i=1,Mf
j =1,Nf

1X1 (X1i , Y1 ) = R1 (, 0)|i,j

(9)

i=1,Mf
j =1,N
R1 (0, )|i,j f

(10)

i=1,Mf
j =1,Nf
R2 (, 0)|i,j

(11)

1Y2 (X2i , Y2 ) =

i=1,Mf
j =1,Nf
R2 (0, )|i,j

(12)

j
1Y1 (X1i , Y1 )
j
1X2 (X2i , Y2 )
j

and Mf , Nf are the flow grid sizes. These data were

also fed into a bicubic spline interpolation to yield the
displacement functions 1X1 (X1 , Y1 ), 1Y1 (X1 , Y1 ) and
1X2 (X2 , Y2 ), 1Y2 (X2 , Y2 ). These functions allowed an
estimate of the particle image displacement in X or
Y at any pixel position on the left and right images.

From these displacement functions and the calibration

coefficient functions a1 (X1 , Y1 ), a2 (X2 , Y2 ), b1 (X1 , Y1 ),
b2 (X2 , Y2 ) and c1 (X1 , Y1 ), c2 (X2 , Y2 ) an estimate of the
three components of particle displacement, (1x, 1y, 1z),
can be made at any point (xi , yj ) inside the calibrated
flow area using previous simple gas interface stereoscopic
derivations as follows [12, 23, 26]:
1x(xi , yj )
j
j
j
j
b2 (X2i , Y2 ) 1X1 (X1i , Y1 ) b1 (X1i , Y1 ) 1X2 (X2i , Y2 )
=
j
j
j
j
a1 (X1i , Y1 )b2 (X2i , Y2 ) a2 (X2i , Y2 )b1 (X1i , Y1 )
(13)
1y(xi , yj )
j
j
j
j
c1 (X1i , Y1 ) 1Y1 (X1i , Y1 ) + c2 (X2i , Y2 ) 1Y2 (X2i , Y2 )
=
2
(14)
1z(xi , yj )
j
j
j
j
a2 (X2i , Y2 ) 1X1 (X1i , Y1 ) a1 (X1i , Y1 ) 1X2 (X2i , Y2 )
=
j
j
j
j
a2 (X2i , Y2 )b1 (X1i , Y1 ) a1 (X1i , Y1 )b2 (X2i , Y2 )
(15)
j

where the image coordinates (X1i , Y1 ) and (X2i , Y2 ) are

found from the X, Y coordinate functions:
X1i = X1 (xi , yi )
j

(16)

Y1 = Y1 (xi , yi )

(17)

= X2 (xi , yi )

(18)

Y2i = Y2 (xi , yi ).

(19)

X2i

1459

N J Lawson and J Wu

2.5. Error analysis

In order to compare the experimental data with the
theoretical error model [23], analysis was carried out on
the output displacement data 1x and 1z. This consisted of
examining the rms errors rms (1x) and rms (1z) in 1x and
1z across a regular i, j flow grid with M by N gridpoints
such that:
v
u M,N
u X
1
t
[1xcal 1x(xi , yj )]2
rms (1x) =
MN 1xcal i=1,j =1
rms (1z) =

1
MN 1zcal

(20)
v
u M,N
u X
t
[1zcal 1z(xi , yj )]2 .
i=1,j =1

(21)
Errors outside two standard deviations of the data [1xcal
1x(xi , yj )] and [1zcal 1z(xi , yj )] were rejected to filter
out spurious measurements and 4000 gridpoints were used
for the analysis. To compare the results to previous
theoretical analysis [23], the experimental error ratio, er ,
was found from:
er =

rms (1z)
.
rms (1x)

(22)

The next section will now show the results from the
error analysis for a range of camera angles and fluid
displacements.
3. Results
Considering the assumption of linearity, we must first
examine this characteristic for a range of camera angles
and test block translations. Figure 5 shows the a and b
calibration coefficients obtained at x/d0 = 0.1 or the edge
of the image where maximum distortion will occur. In each
case two sets of test block translations were made from 200
to 800 m in the x axes and z axes. These results show,
for camera angles of = 15 and = 30 , that the system
clearly behaves with acceptable linearity in both axes.
If we now examine the individual displacement errors
in x and z it can be seen from figure 6 that significant
errors in the z component begin to occur as the camera
angle is decreased from = 45 to = 10 . For example
the rms error in z at = 45 is 4% rising to a maximum
of 13% at = 10 . In contrast the x component error
slowly increases with camera angle from around 1% at
= 10 to 34% at = 45 . These counteracting effects
are caused by the geometric and imaging properties of
the system. In z the error will mainly be dependent on
camera geometry since an increase in camera angle will
result in greater imaging sensitivity to a z displacement and
hence a lower measurement error. In x, however, greater
camera angles will cause increased imaging problems such
as aberrations with a resultant increase in the x error.
Therefore a compromise must be sought between acceptable
imaging quality and the required z accuracy and these initial
results suggest the optimum range of camera angle to be
20 < < 30 .
1460

The results in figure 6 also show a dependence on

particle displacement with a greater displacement giving
improved accuracy. These effects can be attributed to the
noise characteristics of the correlation function which has
a random error primarily dependent on particle image size,
seeding density and velocity gradient [20, 22]. Therefore
because the noise is generally independent of particle image
displacement, a smaller particle displacement will result in
a greater rms % error as shown in the results.
If we now examine the error ratio characteristics for
the system, shown in figure 7, it can be seen that the
experimental trends match the theoretical predictions well.
The theoretical trend follows a 1/ tan function [23].
These theoretical errors are illustrated more clearly in
figure 8 which shows the percentage errors in the prediction
compared with the experimental results for a range of
camera angles and particle displacements. It can be seen
that the discrepancies in the predictions are worst at the
smallest particle displacement of 7.5% of d and w with a
overall error of 38%. This error, however, falls to around
8 and 18% at the larger particle displacements of 15 and
30% of d and w. This improvement is again related to
the noise characteristics of the correlation function and
the resultant fall in rms error with increasing particle
image displacement. The poorer performance with smaller
displacements is also apparent if the error characteristics in
x are examined. These are shown in figure 9 for a range
of 0.1<x/d0 <0.1 and the deviation of the experimental
results from the theoretical prediction clearly becomes
worst at the lower particle displacements.
These results have suggested an optimum range of
camera angle of 20 <<30 and also have shown the
theoretical predictions give an acceptable approximation to
the actual performance of the angular system. The next
section will now discuss these results and their application
in more detail.
4. Discussion
The angular stereo system tested has successfully shown
error ratios of er < 2 can be achieved by using a relatively
simple CCD camera set-up providing the camera angle is
greater than 30 and providing an f number of f 16 or
higher is used. This performance, however, does degrade as
the particle displacement being measured is reduced. This
limitation could be improved by using a higher-resolution
CCD or a 5 in 4 in photographic camera format as
demonstrated by previous workers [12, 13]. This would
allow the interrogation region resolution to be increased
which will produce in a fall in the x rms error. An
additional benefit in the increase in image resolution is
a corresponding increase in spatial resolution which must
be matched to the flow structures being measured [3].
The use of a high-resolution photographic system would
also allow the user to use a more sophisticated angular
configuration such as the Scheimpflug system proposed by
previous workers [16]. In this case the camera back was
tilted by an predetermined angle and a novel liquid prism
used to reduce image aberrations at a given f number. This
resulted in an equivalent image quality being achieved with

3D PIV

Figure 5. Graph to show linearity of system for a range of PIV test block translations and camera angles (x /d0 = 0.1).

Figure 6. Graph to show PIV system performance in terms of x and z rms errors.

a lower f number hence giving a light power advantage. If,

however, sufficient laser power is available and the spatial
resolution of the flow structures matches the resolution
provided by the CCD, these results have shown a relatively
cheap 3D PIV system with high measurement performance
is possible by using in situ calibration in conjunction with
simple processing algorithms.
Previous algorithms developed for particle tracking
velocity (PTV) have suggested an optimum camera angle
of = 45 [28] which matches the theoretical error
ratio analysis [23] but contradicts the optimum result here

of 20 <<30 . The previous analysis [28], however,

used computer generated images which did not include
the aberrations generated in this case by camera geometry
and by imaging through the PIV test blocks. These
optical effects would therefore account for the differences
in optimum camera angle of = 45 [28] and 20 <
< 30 . Hence the optimum set-up in this stereo
system should apply to similar CCD imaging systems.
It must be noted though that the test block width for
these experiments was limited to 20 mm. This was due
to the properties of the resin which contained impurities
1461

N J Lawson and J Wu

Figure 7. Graph to show experimental and theoretical error ratios for a range of camera angles and particle displacements.

Figure 8. Graph to show the deviations in the theoretical prediction of error ratio.

reducing its visibility and due to moulding imperfections

which formed during the gelling stage. Therefore it must
be assumed that a thicker block would degrade the image
quality further as found in previous work, when using a
thick liquid layer [12]. In this situation in order to retain
system performance a simple solution is to increase the
f number of the imaging system [12, 13]. Further tests
were carried to study these effects by placing a PIV test
block in the centre of a 600 mm wide water tank and
acquiring images using the same object distance. It was
1462

found an additional f number was necessary to achieve the

same image quality from simple visual inspection during
acquisition. Consequently, higher laser power would be
required when taking measurements in wider liquid flows.
These tests though did also demonstrate the flexibility of
the PIV test block design which can be customized to the
rig being measured and be used to calibrate in situ.
Previous work discussed in detail the problems of
image registration and calibration in stereo PIV systems
[23]. With photographic systems, moire techniques and

3D PIV

Figure 9. Graph to show the variation of error ratio across the image plane ( = 45 ).

double imaging onto a single transparency have been used

to overcome the registration uncertainty [12, 13]. In this
system, however, the use of bicubic spline interpolations
was relied upon on throughout the image processing
stage to ensure point to point correspondence of any
1X displacements and calibration coefficients. For the
simple uniform flow test images used in this study,
these techniques obviously provided acceptable accuracy.
If considerably more complex flow regimes are to be
measured, however, careful choice of the interpolation grid
resolution would be required to maintain the accuracy of
the interpolation functions. This issue was not investigated
in this study and further work is required in this area.
5. Conclusions
This paper has demonstrated and tested a simple design of
digital angular stereoscopic PIV system. Results from the
tests have shown optimum performance will be obtained
by using camera angles of between 20 and 30 and camera
f numbers of f 16 and higher. With this set-up and if
measuring a uniform flow the in plane and out of plane rms
errors will be 12 and 34% respectively. The results have
also shown a theoretical prediction of system performance
derived in previous work [23] which considers the ratio
of out of plane to in plane errors matches to within 8
and 18% of the experimental system performance providing
particle displacements are restricted between 15 and 30%
of the interrogation region length. Error ratios in this
case were around two. Below this displacement range,
the random errors in measurement caused unacceptable
deviation from the theoretical predictions and an increase in
imaging resolution was suggested to reduce this problem.
References
[1] Ower E and Pankhurst R C 1966 The Measurement of Air
Flow 4th edn (New York: Pergamon) pp 624

[2] Durst F, Melling A and Whitelaw J H 1981 Principles and

Practice of Laser Doppler Anemometry (New York:
Academic)
[3] Adrian R J 1991 Particle-imaging techniques for
experimental fluid mechanics Annu. Rev. Fluid Mech. 23
261304
[4] Pickering C J D and Halliwell N A 1985 Particle image
velocimetry: a new field measurement technique Optical
Measurement in Fluid Dynamics (Inst. Phys. Conf.
Ser. 77) session 4 (Bristol: Hilger)
[5] Mao Z Q, Halliwell N A and Coupland J M 1993 Particle
image velocimetry: high-speed transparency scanning
and correlation-peak location in optical processing
systems Appl. Opt. 32 508991
[6] Meinhart C D, Prasad A K and Adrian R J 1993 A parallel
digital processor system for particle image velocimetry
Meas. Sci. Technol. 4 61926
[7] Reeves M, Garner C P, Dent J C and Halliwell N A 1994
Particle image velocimetry measurements of barrel swirl
in a production geometry optical IC engine SAE Paper
940241
[8] Stolz W, Kohler J, Lawrenz W, Meier F, Bloss W H,
Maly R R, Herweg R and Zahn M 1992 Cycle resolved
flow field measurements using a PIV movie technique in
a SI engine SAE Paper 922354
[9] Hinsch K D 1995 Three-dimensional particle velocimetry
Meas. Sci. Technol. 6 74253
[10] Coupland J M and Halliwell N A 1992 Particle image
velocimetry: three-dimensional fluid velocity
measurements using holographic recording and optical
correlation Appl. Opt. 31 10057
[11] Barnhart D H, Adrian R J and Papen G C 1994
Phase-conjugate holographic system for high resolution
particle image velocimetry Appl. Opt. 33 715970
[12] Prasad A K and Adrian R J 1993 Stereoscopic particle
image velocimetry applied to liquid flows Exp. Fluids 15
4960
[13] Arroyo M P and Greated C A 1991 Stereoscopic particle
image velocimetry Meas. Sci. Technol. 2 11816
[14] Grant I, Zhao Y, Tan Y and Stewart J N 1991 Three
component flow mapping: experiences in stereoscopic
PIV and holographic velocimetry Laser Anemometry
Advances and Applications vol 1, ed A Dybbs and
B Ghorashi (New York: ASME) pp 36571
1463

N J Lawson and J Wu
[15] Westerweel J and Nieuwstadt F T M 1991 Performance
tests on 3D velocity measurements with a two-camera
digital particle image velocimeter Laser Anemometry
Advances and Applications vol 1, ed A Dybbs and
B Ghorashi (New York: ASME) pp 34955
[16] Prasad A K and Jensen K 1995 Scheimpflug stereocamera
for particle image velocimetry in liquid flows Appl. Opt.
34 70929
[17] Raffel M, Gharib M, Ronneberger O and Kompenhans J
1995 Feasibility study of three-dimensional PIV by
correlating images of particles within parallel light
sheets Exp. Fluids 19 6977
[18] Brucher C 1996 3D PIV via spatial correlation in a
color-coded light sheet Exp. Fluids 21 3124
[19] Grant I, Fu S, Pan X and Wang X 1995 The application of
an in-line, stereoscopic, PIV system to 3-component
velocity measurements Exp. Fluids 21 21421
[20] Keane R D and Adrian R J 1990 Optimisation of particle
image velocimeters. Part I: double pulsed systems Meas.
Sci. Technol. 1 120215
[21] Prasad A K, Adrian R J, Landreth C C and Offutt P W
1992 Effect of resolution on the speed and accuracy of
particle image velocimetry interrogation Exp. Fluids 13
10516

1464

[22] Keane R D and Adrian R J 1992 Theory of cross

correlation analysis of PIV images J. Appl. Sci. Res. 49
191215
[23] Lawson N J and Wu J 1997 Three-dimensional particle
image velocimetry: error analysis of stereoscopic
techniques Meas. Sci. Technol. 8 894900
[24] Press W H, Teukolsky S A, Vetterling W T and
Flannery B P 1992 Numerical Recipes in FortranThe
Art of Scientific Computing 2nd edn (New York:
Cambridge University Press)
[25] Goodman W G 1968 Introduction to Fourier Optics (New
York: McGraw-Hill) pp 27580
[26] Westerweel J 1993 Analysis of PIV interrogation with low
pixel resolution Optical Diagnosis in Fluid and Thermal
Flow (San Diego, 1993), SPIE vol 2005, ed S S Cha and
J D Trolinger (Bellingham, WA: SPIE) pp 62435
[27] Landreth C C and Adrian R J 1988 Measurement and
refinement of velocity data using high image density
analysis in particle image velocimetry Proc. 4th Int.
Symp. on the Applications of Laser Anemometry to Fluid
Mechanics (Lisbon, 1988) ed R J Adrian et al, pp 4917
[28] Guezennec Y G, Brodkey R S, Trigui N and Kent J C
1994 Algorithms for fully automated three-dimensional
particle tracking velocimetry Exp. Fluids 17 20919