6th International Symposium on Particle Image Velocimetry PIV05 Paper nnnn

Pasadena, California, USA, September 21-23, 2005

Tomographic particle image velocimetry

Abstract This paper describes the principles of a new 3D PIV system based on the illumination, recording and
reconstruction of tracer particles within a three-dimensional measurement volume. The technique makes use of
several simultaneous views of the illuminated particles and the three-dimensional light intensity reconstruction is
based on optical tomography. The technique is therefore referred to as Tomographic Particle Image Velocimetry
(Tomo-PIV). The analysis of the light intensity distribution in the 3D volume is performed with the MART
reconstruction algorithm, yielding a 3D array of light intensity discretised over voxels. The reconstructed tomogram
pair is then analyzed by means of FFT-based 3D cross-correlation returning the three-components velocity vector
distribution over the measurement volume. The principles and details of the tomographic algorithm are discussed
and a parametric study is carried out by means of a computer-simulated Tomo-PIV procedure. The study focuses on
the accuracy of the light intensity reconstruction process. The simulations also identify the most important
parameters governing the experimental setup and the tomographic algorithm parameters, showing their effect on the
reconstruction. A full-scale synthetic experiment of a 3D particle motion field finally demonstrates the capability
and potential of the proposed system with four 700x700 pixels cameras yielding 36x36x7 vectors over the
measurement volume and resolving the 3D flow structure.

1
Introduction
The instantaneous measurement of the three-dimensional velocity field is of great interest to fluid mechanic research
as it is able to reveal the complete topology of unsteady coherent flow structures. Moreover three-dimensional
measurements are relevant for those situations where the flow does not exhibit specific symmetry planes or axes and
several planar measurements have to be performed. Last but not least, flow turbulence is intrinsically three-
dimensional and its full description requires the application of measurements able to capture instantaneously its
three-dimensional structure, the complete stress tensor and the vorticity vector. The advent of PIV first and its
developments (stereo-PIV, Arroyo and Greated 1991, dual-plane stereo-PIV, Khler and Kompenhans 2000)
showed the capability of the PIV technique to quantitatively visualize complex flows. Several different methods
were also proposed to achieve a three-dimensional version of the technique (scanning light sheet, Brcker 1995,
holography, Hinsch 2002, 3D PTV, Maas et al 1993).
The present work proposes a novel system for 3D velocity measurements based on tomographic reconstruction
of the three-dimensional particle distribution within an illuminated volume from 2D particle image recordings taken
from various viewing directions simultaneously. The method is therefore referred to as Tomographic Particle Image
Velocimetry (Tomo-PIV). Provided that two subsequent exposures of the particle images are obtained, the
measurement technique returns the instantaneous velocity field within the measurement volume by means of three-
dimensional particle pattern correlation.
The motivation for such a 3D PIV system is presented in the next section together with an overview of current
techniques. Then the working principle and the tomographic reconstruction method are discussed in detail, followed
by a numerical simulation study showing the effect of the relevant experimental and reconstruction parameters and
finally the full-scale capabilities of Tomo-PIV.

2
Current 3D PIV techniques
Among the different 3D velocimetry techniques presently available certainly Holography-PIV received most
attention (Hinsch 2002, Chan et al 2004). It uses the interference pattern of a reference light beam with light

G.E. Elsinga, F. Scarano, B.W. van Oudheusden, Department of Aerospace Engineering, Delft University of Technology, Delft, The Netherlands
B. Wieneke, LaVision GmbH, Gttingen, Germany

Correspondence to:
G.E. Elsinga, Department of Aerospace Engineering, Delft University of Technology,
Kluyverweg 1, 2629 HS Delft, The Netherlands, E-mail: g.e.elsinga@lr.tudelft.nl

1
 PIV05

scattered by a particle, which is recorded on a hologram, to determine the particle location in depth. The in-plane
position in principle is given by the position of the diffraction pattern in the image. Illumination of the hologram
with the reference light beam reproduces the original light intensity field in the measurement volume at the time of
recording, the intensity being highest at the original particle location. The reconstructed intensity field is scanned by
a sensor, e.g. a CCD, to build a digital intensity map, which can be used for cross-correlation yielding the velocity
field. So far Holographic-PIV has shown a great potential in terms of a high data yield. However its drawbacks are
that the recording medium is a holographic film requiring wet processing, which makes the process time consuming
and somehow inaccurate due to misalignment and distortion when re-positioning the hologram for the object
reconstruction. The technique was successfully applied to measure a vortex ring in air, the wake of a mixing tab in
water (both in Pu and Meng 2000), a cylinder wake flow in air and a free air nozzle flow (both in Herrmann et al
2000) returning large numbers of vectors (up to 92,000 using particle pairing, Pu and Meng 2000).
Besides recoding the hologram on a photographic plate, it can also be captured directly by a CCD sensor (Digital-
Holographic-PIV, Cotmellec et al 2001). In that case the light intensity distribution in the measurement volume is
built numerically usually by solving the Fresnel diffraction formula on the hologram (near-field diffraction, Pan and
Meng 2002). CCD sensors, however, have a very limited resolution compared to the photographic plate returning
about 2 to 3 orders less particle images and velocity vectors. Moreover the large pixel pitch requires that the
recording is obtained at a relatively small angle (a few degrees between reference beam and scattered light) in order
to resolve the interference pattern, hence strongly limiting the numerical aperture and depth resolution (Hinsch
2002).
The scanning-PIV technique is directly derived from standard 2D or stereo PIV with the light sheet scanning
through the measurement volume (Brcker 1995). The volume is sliced by the laser sheet at sequential depth
positions where the particle image pattern is recorded. The second recording at that depth position can be taken
either directly after the first or after the complete scan of the volume. The procedure returns planar velocity fields
obtained slightly shifted in space and time, which can be combined to return a 3D velocity field. It requires high-
repetition systems (kHz) to ensure that the complete volume is recorded at almost the same time. The underlying
hypothesis of scanning-PIV is that the volume scanning time needs to be small if compared with the characteristic
time scale of the investigated flow structure. Cameras with high recording rates are thus required, which is not a
problem in low speed flow as shown by Hori and Sakakibara (2004) who measure a turbulent jet in water, but can be
limiting (reduced image resolution) or technically impossible at high speed. Moreover high repetition rate lasers are
expensive and provide relatively low pulse energy. It should also be remarked that the experimental setup is
significantly more complicated with a scanning laser.
As an alternative to scanning, dual plane stereo PIV (Khler and Kompenhans 2000) records the particle images
in different planes simultaneously using light polarization or different colors to distinguish the scattered light from
the two planes. In principle, measurements can be performed over more than two planes with each plane requiring a
double-pulse laser, however separation by polarization is the most commonly adopted solution and is possible only
over two planes. Furthermore using different colors complicates the optical arrangement.
Relatively recent options are 3D PTV and defocusing PIV (as presented by Maas et al 1993and Pereira et al
2000, respectively). They all rely on the identification of individual particles in the PIV recordings. The exact
position of the particle within the volume is given by the intersection of the lines of sight corresponding to a particle
image in the recordings from all viewing directions (typically three or four). The implementation of the particle
detection and location varies with the methods. In comparison with the previous two methods, the 3D-PTV approach
offers the advantage of being fully digital and fully three-dimensional without the requirement for moving parts. The
velocity distribution in the volume is obtained from either particle tracking or by 3D-cross-correlation with
artificially created particles (Schimpf et al 2003). However the identification of particles in an image can be
complicated and somewhat problem-dependent especially for a high seeding density with a high probability of
overlapping particle images. Moreover, the precision of the volume calibration or in the description of the imaging
optics is finite. This means the lines of sight for a particle almost never truly intersect and an intersection criterion is
needed. Consequently the maximum seeding density in 3D PTV is kept relatively low (typically 0.005 particles per
pixel for a 3 camera system, Maas et al 1993), and consequently also the resolution of the velocity spatial
distribution.
The development of Tomo-PIV as a new 3D PIV technique is motivated by combining the simplicity of the
optical arrangement of the photogrammetric approach with a robust particle volume reconstruction procedure not
relying on particle identification to allow a higher seeding (information) density compared to 3D PTV. Furthermore
the method is truly instantaneous, as opposed to scanning PIV, not needing high-speed cameras and allowing future
extension to high-speed flow application since it makes use of a fully digital recording and analysis procedure as
opposed to holographic PIV.

2
 PIV05

3
Working principle of Tomo-PIV
The working principle of Tomo-PIV is schematically represented in figure 1. Tracer particles within the
measurement volume are illuminated by a pulsed light source and the scattered light pattern is recorded
simultaneously from several viewing directions using CCD cameras similar to stereo-PIV. The entire volume is in
focus making use of a relatively large f/# and applying the Scheimpflug condition between the image plane, lens
plane and the mid-object-plane. The 3D particle distribution (the object) is reconstructed as a 3D light intensity
distribution, from its projections on the CCD arrays. The reconstruction problem (an inverse problem) is not
straightforward and it is in general underdetermined meaning that a single set of projections can result from many
different 3D objects. Determining the most likely 3D distribution is the topic of tomography (Herman and Lent
1976), which is discussed in the following section. The particle displacement (hence velocity) within a chosen
interrogation volume is then obtained by the 3D cross-correlation of the reconstructed particle distribution at the two
exposures. The cross-correlation algorithm is a direct extension of the commonly used 2D FFT cross correlation to
3D as proposed before in holographic PIV (Herrmann et al 2000 among others) and photogrammetric PIV (Schimpf
et al 2003).

Figure 1: Principal of Tomo-PIV

The relation between image (projection) coordinates and the physical space (the reconstruction volume) is
established by a calibration procedure common to stereo PIV. Each camera records images of a calibration target or
grid of known dimensions, which is either volumetric or planar and/or is moved in depth direction through the
volume. These images are then used to determine the viewing direction and field of view. For tomographic
reconstruction to be successful the calibration needs to be accurate within a fraction of a pixel or the particle image

3
 PIV05

size (see section 5). For mapping from space to camera coordinate system different functions have been used, e.g.
camera pinhole model (Tsai 1986) or third-order polynom in x and y (Soloff et al 1997).

4
Tomographic reconstruction algorithm
The novel aspect in Tomo-PIV is the reconstruction of the 3D particle distribution by optical tomography. Therefore
a separate section is devoted to the tomographic reconstruction problem and algorithms for solving it.
By considering the properties of the measurement system, one can already reject some reconstruction methods
beforehand. First the particle distribution is discretely sampled on pixels from a small number of viewing directions
(typically 3 to 6 CCD cameras) and involves high frequencies, which makes analytical reconstruction methods, such
as Fourier and back-projection methods, unsuitable (Timmerman 1997). Secondly statistical methods select out of
all possible solutions the most probable one based on prior knowledge of the object and of the noise (Timmerman
1997), e.g. the total number of particles, the particle intensity and size, which can be computationally demanding
given the large number of particles. Moreover these parameters are in fact not known and need to be estimated,
which makes the method not straightforward and somewhat arbitrary. Only algebraic reconstruction methods are
considered for further evaluation.
Algebraic methods (Herman and Lent 1976) iteratively solve a set of linear equations modeling the imaging
system, in which the object is written as a series of basis functions. The model is given by:

Wf=p (1)

where f is an array of unknown coefficients of the basis functions, p is an array containing the recorded pixel
intensities in the recorded images (projections) and W is a two-dimensional array of weighting coefficients
describing the contribution of each basis function to each element (pixel) in p. A cubic voxel element with non-zero
value inside and zero outside is chosen as a local basis function, because particles are very small/local features and
cross-correlation can be easily extended from a pixel to a voxel based object. A model of the physics of the
measurement establishes the coefficients in W. Diffraction limited imaging is, however, not taken into account, since
that would result in sharply reconstructed particles which is unfavorable in cross-correlation. Based on geometrical
optics, the recorded pixel intensity then is the object intensity integrated along the corresponding line of sight
(obtained from the calibration). In that case the reconstructed particle is represented by a 3D Gaussian-type blob,
which projection in all directions is the diffraction spot. The contribution of a voxel to a pixel (the weighting
coefficient) is given by the intersecting volume of a sphere representing the voxel (of equal volume) with a circular
cylinder representing the pixel along the line of sight (of equal cross sectional area) relative to the voxel volume.
The coefficients only depend on the relative size of a voxel to a pixel and the distance between the voxel center and
the cylinder axis (the line of sight) and not on viewing direction, which simplifies the calculation. Note that 0wi,j1
for all entries wi,j in W. The weighting coefficients can also be used to account for different camera sensitivities,
forward or backward scatter differences or other optical dissimilarities between the cameras. Alternatively one can
preprocess the recorded images in an appropriate way to compensate these effects.
A range of tomographic reconstruction algorithms exist to solve these equations. However due to the problem
being underdetermined (section 3), they may converge to different solutions, which means that they solve for a
different optimization/secondary criterion besides the set of equations of Eq. 1. The advantage for the present system
is that the basic shape of the object is known (sparse distribution of small particles in a confined volume), so that
with the appropriate optimization criterion the object can be accurately reconstructed. In this respect numerical
simulations of a Tomo-PIV experiment are performed to evaluate the performances of different tomographic
algorithms focusing the evaluation upon the reconstruction accuracy and convergence properties. Two different
algorithms are compared: additive and multiplicative. They are referred to as ART (algebraic reconstruction
technique) and MART (multiplicative algebraic reconstruction technique) respectively (Herman and Lent 1976).
Both algorithms start from an initial guess of f and consider only one equation (pixel) in each update. ART updates
the object f by adding a correction term as:

pi wi f k
f k +1 = f k + 2
w Ti (2)
wi

4
 PIV05

where k indicates the sequence in f, i indicates the equation or pixel considered, wi is the corresponding row in W
with norm w i and is a scalar relaxation parameter between 0 and 2. The magnitude of the correction depends on
the residual pi wi f multiplied by a scaling factor and the weighting coefficient, so that only the elements in f
k

th
affecting the i pixel are updated. A full iteration is completed after all equations i are used. Alternatively MART
updates each element j in the object f as:

p wi, j
k +1
=f j
ik
f j k
(3)
wi f

where 1 is a scalar relaxation parameter. The ratio of the measured pixel intensity pi and the projection of the
k
current object wi f determines the magnitude of the update. The exponent again ensures that only the elements in f
affecting the ith pixel are updated. The multiplicative MART scheme requires that p and f are positive.
The techniques are tested in a domain with reduced dimensionality: a 50x10 mm2 2D slice of the 3D volume in
order to save computational time and expand the number of parameters to be evaluated. The 2D particle field
contains 50 particles from three synthetic one-dimensional PIV signal recorded on a 1000 pixels array. Further
details of the simulation will be given in the next section. For the ART reconstruction the relaxation parameter is set
at 0.2 and the initial condition is a uniform zero intensity distribution, while for MART relaxation and initial
condition are both 1.

ART
10 70
60
z (mm)

5 50
40
30
0
-25 -20 -15 -10 -5 0 5 10 15 20 25 20
x (mm)

MART
10 70
60
z (mm)

5 50
40
30
0
-25 -20 -15 -10 -5 0 5 10 15 20 25 20
x (mm)

Figure 2: Particle field reconstructed using ART (upper) and the same field reconstructed using MART (lower). The
actual particle positions are indicated by circles. The color represents the intensity level.

The reconstructed particle field is converged after 20 iterations and the results are shown in figure 2. The
maximum intensity is 75 counts and values below 10 counts are blanked for readability. The ART algorithm leaves
traces of the particles along the lines of sight, while the MART reconstruction shows more distinct particles. The
additive ART scheme works similar to an OR-operator: in order to have a non-zero intensity in the object it is
sufficient to have a particle at the corresponding location in one of the PIV recordings. Still the intensity is highest at
the actual particle locations. The multiplicative MART scheme behaves as an AND-operator: non-zero intensity

5
 PIV05

only at locations where a particle appears in all recordings. The suitability of MART to reconstruct objects with
sharp gradients or spikes is confirmed by Verhoeven (1993). The artifacts in the ART reconstruction are undesirable
when it is to be used in cross-correlation later on. Therefore MART is chosen for Tomo-PIV. Moreover figure 3
shows that the individual particles reconstructed with MART are reconstructed at the correct position and have an
intensity distribution that is very suitable for cross-correlation.
Besides the sequential update algorithms of Eqs. 2 and 3 there exist simultaneous schemes that update the
reconstructed object using all equations in a single step, such as the conjugate gradient method (additive scheme)
and several simultaneous MART algorithms (Mishra et al 1999). Those methods return similar results and are
potentially computationally more efficient.

MART
4
70

60
3.5

50
z (mm)

3
40

30
2.5
20

2
-9 -8.5 -8 -7.5 -7
x (mm)

Figure 3: Detail of the MART reconstruction (figure 2) showing individual particles.

5
Parametric study
This section discusses the effect of experimental and algorithm parameters on the reconstruction based on numerical
simulations. The number of iterations, the number of cameras, the viewing directions, the particle density, the
calibration accuracy and the image noise are considered.
In the simulation the order of the problem is reduced from a 3D volume with 2D images to a 2D slice with 1D
images, which simplifies the computation and the interpretation of the results without changing the properties of the
tomographic reconstruction. Tracer particles are distributed in a 50x10 mm2 slice, which is imaged at 1000x1 pixel
resolution from different viewing directions by cameras placed at infinity (figure 4), so that magnification and
viewing direction are constant over the field of view. The magnification is identical for all views and is close to 1.
Furthermore the entire volume is assumed to be in focus. Given the optical arrangement the particle location in the
images is calculated and the application of diffraction (particle diameter is 3 pixels, which is justified by the large f/#
required) results in the synthetic recordings. The particle image intensity is constant. The 2D particle field is
reconstructed from these recordings at 1000x200 voxel resolution using the MART algorithm described in the
previous section. Unless stated otherwise the following (reference) experimental setup is used: 3 cameras at = -20,
0 and 20 degrees, 50 particles (0.05 particles per pixel), 5 iterations and no calibration errors or image noise.
To assess the reconstruction frec it is compared with fgaus in which the actual particles are represented by a
Gaussian intensity distribution of 3 voxel diameter (equal to the diffraction spot). The reconstruction quality Q,
which is a correlation coefficient, is defined as:

f rec f gaus
Q= (4)
f rec f gaus

The correlation coefficient is chosen, because the reconstructed object is to be used in cross-correlation and the
Gaussian particle fields are an appropriate reconstruction for that purpose.

6
 PIV05

Figure 4: Experimental setup used for the numerical simulation.

First the convergence of the residual and the reconstruction is considered. The residual (figure 5-left) always
decreases with the number of iteration in tomography, but that does not necessarily mean that the reconstructed
object converges to the desired solution (Watt and Conery 1993). However, for the present reconstruction algorithm
and noise free recordings, the solution also converges as shown in figure 5-right. A small divergence after 4 to 5
iterations was found for recordings with high noise levels (50% of the particle peak intensity). The divergence can
be countered by applying a smaller relaxation ( = 0.2) as suggested by Mishra et al (1999), although that requires
more iterations to reach a similar reconstruction quality Q.

0
10 1

0.9
residual [-]

Q [-]

-1
10 0.8

0.7

-2 0.6
10
0 5 10 15 20 0 5 10 15 20
iteration iteration

Figure 5: Convergence of the normalized residual (left) and the reconstruction quality (right).

7
 PIV05

1 1
0.9 0.9
0.8 0.8
0.7 0.7
Q [-]

Q [-]
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
2 3 4 5 0 10 20 30 40 50 60
number of cameras [deg]

1 1
0.9 0.9
0.8 0.8
0.7 0.7
Q [-]

Q [-]

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0 0.05 0.1 0.15 0.2 0 0.2 0.4 0.6 0.8 1
Np [ppp] calibration error [pixels]

Figure 6: Effect on the reconstruction quality of the number of cameras (upper left), viewing direction (upper right),
particle density (lower left) and calibration accuracy (lower right).

Figure 6 presents the reconstruction quality for the relevant experimental parameters, which are the number of
cameras, viewing direction, particle image density (in particles per pixel), and calibration error. The graphs show
trends and provide a first indication of the optimum experimental arrangement and limitations of the system. A
correlation coefficient of 75% with the desired solution (Q = 0.75) is used as a cut-off value, above which the
reconstruction is considered successful.
The effect of the number of cameras is clear: adding camera gives additional information on the object, which
increases reconstruction accuracy. A 2-camera system ( = -20 and 20) is largely insufficient returning a poor
result. The reconstruction quality rapidly improves when going to 3 and 4 cameras ( = -20, 0, 20 and 40) and is
nearly unity applying 5 cameras ( = -40, -20, 0, 20 and 40).
The viewing angles are changed maintaining the symmetric camera arrangement shown in figure 4, although
symmetry is not necessary. The angle indicated in figure 6 is the angle of the outer cameras with respect to the z-
axis. The graph shows an optimum near 30. For smaller angles the depth resolution decreases resulting in
elongation of the reconstructed-particle in depth. For larger angles the number of ghost particles increases.
Particles are reconstructed at locations where the lines of sight from all cameras correspond to a particle image
including locations where physically there is no particle, thereby forming a ghost particle. This is a problem of
ambiguity, which increases with the number of particles, the particle diameter and the length of the line of sight in
the volume (Maas et al 1993). The latter increases with the viewing angle in the present configuration, hence the
increase in ghost particles. The viewing angles returning good results range from 15 to 45 degrees.
As seen from the above discussion, increasing the particle density is expected to increase the number of ghost
particle and consequently decrease the reconstruction quality. On the other hand, increasing particle density will
return a better sampling of the flow through the tracers, hence more velocity information. Therefore a high particle

8
 PIV05

density is desirable. Based on the simulation results (figure 6) the maximum particle density yielding a good
reconstruction is 0.075 particles per pixel.
In real experiments the calibration accuracy is finite, which results in a dislocation of the lines of sight in the
reconstruction. As shown by Watt and Conery (1993) this reduces the accuracy of the reconstruction. To quantify
the error and to find the necessary calibration accuracy, a calibration error is introduced after recording the images
by shifting the center-camera (figure 4) to the right and the right-camera to the left by an equal amount (being the
calibration error). Lines of sight from the different cameras that originally intersected now form a triangle in the
reconstruction volume. As seen from figure 6 a calibration error of 0.4 pixel still returns a good reconstruction.
Stereo PIV calibration methods exist that can meet this requirement. For larger errors the reconstruction quality
rapidly decreases. Advanced self-calibration methods have been developed for planar Stereo-PIV (Scarano et al
2005 and Wieneke 2005), which can reduce calibration errors to less than 0.1 pixel. Along the same lines one can
correct the calibration function for the complete volume by optimizing the shape of reconstructed particles.

1 1
0.9 0.9
0.8 0.8
0.7 0.7
Q [-]

Q [-]
0.6 0.6
0.5 0.5
0.4 original images 0.4 outside volume
subtract sliding average inside volume
0.3 0.3
0 10 20 30 40 50 0 10 20 30 40 50
noise level [%] added particles [%]

Figure 7: Effect on the reconstruction of white image noise (left) and unmatched particles (right).

Finally two types of image noise are considered: random noise added to the recordings and background particles,
which are located outside the reconstruction volume. Figure 7-left shows the effect of random noise fluctuations,
which range is expressed as a percentage of the particle peak intensity. The image noise deteriorates particle shape
and increases the number of ghost particles, as noise is mistaken for particles. As a result the reconstruction quality
strongly depends on the random image noise. However image pre-processing can be applied to reduce number of
ghost particles. Subtracting a sliding average from the recoding using a window of 61 pixels significantly improves
the reconstruction results (figure 7-left). Random image noise up to 25% of the particle peak intensity has only a
small effect after pre-processing. For higher noise level (50%) the change of the particle image shape is important
and cannot be recovered by the pre-processing. The effect of added background particles (located outside the
volume) is compared with the same amount of particles added inside the volume in figure 7-right. The particle
located outside the volume have a stronger effect compared to the ones inside, therefore the reconstruction should
include the entire illuminated volume to yield the best result, or, the other way around, one should only illuminate
the volume which one wants to reconstruct

6
Synthetic 3D experiment
The potential of Tomo-PIV is demonstrated by a synthetic experiment of a 3D particle motion field of a ring vortex.
The vortex core is located in the center plane z = 0 mm and forms a circle of 10 mm diameter. The magnitude of the
displacement in voxels d is given by:

x R
4 R l
d = y = e (5)
l
z

9
 PIV05

where R is the distance to the voxel-center ring and l = 2 mm is a length scale that defines the width of the vortex.
The maximum displacement is 1.5 voxels. Four cameras are placed at infinity, all having a 30 viewing angle (to the
left, to the right, upward and downward) and recording at 700x700 pixel resolution. The measurement volume is
35x35x7 mm3, which is reconstructed at 700x700x140 voxel resolution using 10 iteration and = 1. The
magnification is 1 and the particle image diameter is 3 pixels. Inside the volume 24500 particles are distributed
corresponding to 0.050 particles per pixels. Standard single-pass cross-correlation with 40 3 voxels interrogation
volumes at 50% overlap is used to analyze the reconstructed volumes returning 36x36x7 vectors.

Y
Z X
15

X Z
15
10

10
5
Vorx [-]
0.50
5
0.43

y [mm]
0 0.36
y [mm]

0
0.29
0.21
-5 0.14
-5 0.07
0.00
-0.07
-10 -0.14
-10
-0.21
-0.29
-15 -15 -15 -0.36
-10 -0.43
-5
0 -0.50
5
x [mm 10
] 15 0 m] -2 0 2
z [m
z [mm]

Figure 8: Velocity field from a simulation of a vortex ring (left) and velocity vectors in the cross section at
x = 0.95 mm (right) the contours show the vorticity in x-direction (in voxels/voxel).

The measured vector field is shown in figure 8-left where the overall motion pattern is well captured. Figure 8-
right presents the vectors in the cross section at x = 0.95 mm with the corresponding vorticity in x-direction
(contours). The symmetry in the flow and the vortex core are clearly visible. The surface corresponding to a
vorticity magnitude of 0.30 voxels/voxel is shown in figure 9. A donut shape is returned as expected. It is concluded
that Tomo-PIV is capable of the instantaneous measurement for flow structures at a good resolution. Furthermore it
is expected that the implementation of a more sophisticated cross-correlation algorithm (i.e. multi-pass and volume
deformation) will improve the accuracy and robustness.

10
 PIV05

Z X

15

10

y [mm]
0

-5

-10

-15 -15
-10
-5
0
5
x [mm 10
] 15 0 m]
z [m

Figure 9: Iso-surface for the measured vorticity magnitude of 0.30 voxels/voxel.

7
Conclusions
Tomographic particle image velocimetry (Tomo-PIV) was presented as a novel technique for 3D velocity
measurements. The three-dimensional particle distribution within a volume is reconstructed by optical tomography
from 2D particle image recordings taken from various viewing directions simultaneously. Velocity information
results from three-dimensional particle pattern correlation of two reconstructions obtained from subsequent
exposures of the particle images. The technique is fully digital, truly instantaneous and allows relatively high
seeding (information) density compared to particle tracking methods.
A numerical simulation of a Tomo-PIV experiment showed that the multiplicative algebraic reconstruction
techniques (MART) are the best tomographic algorithms for particle reconstruction returning distinct particles
without many artifacts. Furthermore it was shown that the reconstruction algorithm converges to the desired solution
for noise free particle image recordings. Based on a parametric study the number of cameras, viewing direction,
particle density, calibration accuracy and image noise were identified as critical factors determining the quality of
the 3D particle reconstruction. The calibration must be accurate within 0.4 pixel. Image pre-processing (e.g. subtract
the sliding average) reduces the effect of random image noise. Adding a camera to the system provides extra
information, which can be used to increase measurement resolution (increasing the particle density) or accuracy.
Finally a full-scale simulation of a ring vortex flow showed the capability of Tomo-PIV to instantaneously
measure flow structures at a good resolution. A four-camera system, each with 700x700 pixel resolution, returned
36x36x7 vectors in a 35x35x7 mm3 volume, which was reconstructed at 700x700x140 voxel resolution. The first
results from wind-tunnel experiments are presented in Elsinga et al (2005).

Acknowledgements
The authors wish to thank Prof. Dave Watt of the University of New Hampshire and Dr. Dirk Michaelis of LaVision
GmbH for the introduction to tomography and the software implementation, respectively.

11
 PIV05

References
Arroyo MP; Greated CA (1991) Stereoscopic particle image velocimetry. Meas Sci Technol, Vol 2, 1181-1186
Brcker Ch (1995) Digital-Particle-Image-Velocimetry (DPIV) in a scanning light-sheet: 3D starting flow around a
short cylinder. Exp Fluids, Vol 19, 255-263
Chan VSS; Koek WD; Barnhart DH; Bhattacharya N; Braat JJM; Westerweel J (2004) Application of
holography to fluid flow measurements using bacteriorhodopsin (bR). Meas Sci Technol, Vol 15, 647-655
Cotmellec S; Buraga-Lefebvre C; Lebrun D; zkul C (2001) Application of in-line digital holography to
multiple plane velocimetry. Meas Sci Technol, Vol 12, 1392-1397
Elsinga GE; Wieneke B; Scarano F; Van Oudheusden BW (2005) Assessment of Tomo-PIV for three-
dimensional flows. 6th Int Symp on Particle Image Velocimetry, Pasadena, CA, USA
Herman GT; Lent A (1976) Iterative reconstruction algorithms. Compt Biol Med, Vol 6, 273-294
Herrmann S; Hinrichs H; Hinsch KD; Surmann C (2000) Coherence concepts in holographic particle image
velocimetry. Exp Fluids, Suppl, S108-S116
Hinsch KD (2002) Holographic particle image velocimetry. Meas Sci Technol, Vol 13, R61-R72
Hori T; Sakakibara J (2004) High-speed scanning stereoscopic PIV for 3D vorticity measurement in liquids. Meas
Sci Technol, Vol 15, 1067-1078
Khler CJ; Kompenhans J (2000) Fundamentals of multiple plane stereo particle image velocimetry. Exp Fluids,
Suppl, S70-S77
Maas HG; Gruen A; Papantoniou D (1993) Particle tracking velocimetry in three-dimensional flows. Exp Fluids,
Vol 15, 133-146
Mishra D; Muralidhar K; Munshi P (1999) A robust MART algorithm for tomographic applications. Num Heat
Transfer, Part B, Vol 35, 485-506
Pan G; Meng H (2002) Digital holography of particle fields: reconstruction by use of complex amplitude. Appl
Optics, Vol 42, 827-833
Pereira F; Gharib M; Dabiri D; Modarress D (2000) Defocusing digital particle image velocimetry: a 3-
component 3-dimensional DPIV measurement technique. Exp Fluids, Suppl, S78-S84
Pu Y; Meng H (2000) An advanced off-axis holographic particle image velocimetry (HPIV) system. Exp Fluids,
Vol 29, 184-197
Scarano F; David L; Bsibsi M; Calluaud D (2005) S-PIV comparative assessment: image dewarping +
misalignment correction and pinhole + geometric back projection. Exp Fluids, DOI: 10.1007/s00348-005-1000-x
Schimpf A; Kallweit S; Richon JB (2003) Photogrammetric particle image velocimetry. 5 th Int Symp on Particle
Image Velocimetry, Busan, Korea, paper 3115
Soloff SM; Adrian RJ; Liu Z-C (1997) Distortion compensation for generalized stereoscopic particle image
velocimetry. Meas Sci Technol, Vol 8, 1441-1454
Tsai RY (1986) An efficient and accurate camera calibration technique for 3D machine vision. Proc of IEEE Conf
on Computer Vision and Pattern Recognition, Miami Beach, FL, USA, 364-374
Timmerman BH (1997) Holographic interferometric tomography for unsteady compressible flow. PhD thesis, Delft
University of Technology, Delft, NL
Verhoeven D (1993) Limited-data computed tomography algorithms for the physical sciences. Appl Optics, Vol 32,
3736-3754
Watt DW; Conery DG (1993) Computational aspects of optical tomography for fluid mechanics and combustion.
3 rd ASME Symp on Exp and Num Flow Visualization, New Orleans, LA, USA, 241-251
Wieneke B (2005) Stereo-PIV using self-calibration on particle images. Exp Fluids, DOI: 10.1007/s00348-005-
0962-z

12