Measurement Science and Technology

Related content
- A simple single camera 3C3D velocity
On the calibration of astigmatism particle tracking measurement technique without errors
C Cierpka, R Segura, R Hain et al.
velocimetry for microflows - Hybrid micro-/nano-particle image
velocimetry for 3D3C multi-scale velocity
field measurement in microfluidics
To cite this article: C Cierpka et al 2011 Meas. Sci. Technol. 22 015401 Young Uk Min and Kyung Chun Kim

- In situ calibrated defocusing PTV for wall-

bounded measurement volumes
T Fuchs, R Hain and C J Kähler
View the article online for updates and enhancements.

Recent citations
- Volumetric velocimetry for fluid flows
Stefano Discetti and Filippo Coletti

- Development of multi-spectral three-

dimensional micro particle tracking
velocimetry
Wei-Hsin Tien

- Tomographic PIV: particles versus blobs

Frédéric Champagnat et al

This content was downloaded from IP address 193.175.232.78 on 02/05/2019 at 08:55

IOP PUBLISHING MEASUREMENT SCIENCE AND TECHNOLOGY
Meas. Sci. Technol. 22 (2011) 015401 (10pp) doi:10.1088/0957-0233/22/1/015401

On the calibration of astigmatism particle

tracking velocimetry for microflows
C Cierpka, M Rossi, R Segura and C J Kähler
Institute of Fluid Mechanics and Aerodynamics, Bundeswehr University Munich,
85577 Neubiberg, Germany
E-mail: christian.cierpka@unibw.de

Received 29 July 2010, in final form 2 September 2010

Published 22 November 2010
Online at stacks.iop.org/MST/22/015401

Abstract
Astigmatism particle tracking velocimetry (APTV) is a method to determine three components
(3C) of the velocity field in a volume (3D) using a single camera. The depth position of the
particles is coded by optical distortions caused by a cylindrical lens in the optical setup. This
technique is particularly suited for microfluidic applications as measurement errors due to
spatial averaging and depth of correlation, typically encountered with μPIV approaches, are
eliminated so that the measurement precision is enhanced. Unfortunately, the current state of
the technique is limited by the small measurement region achievable with the current
calibration procedures as well as by higher order image aberrations (Cierpka et al 2010 Meas.
Sci. Technol. 21 045401). In order to extend the size of the measurement volume and to
account for all image aberrations, a new intrinsic calibration procedure, based on the imaging
function of the particles, is proposed in the paper at hand. It provides an extended measurement
depth, taking into account all image aberrations. In this work, the calibration procedure was
applied to a μPIV arrangement but could also be implemented on macroscopic experimental
setups. The calibration procedure is qualified with synthetic data as well as Poiseuille flow in a
straight rectangular micro-channel with a cross-sectional area of 200 × 500 μm2 . The
three-dimensional velocity distribution of the whole channel was resolved via APTV with
uncertainties of 0.9% and 3.7% of the centerline velocity, uc , for the in-plane and out-of-plane
components, respectively. Further investigations using different cylindrical-lens focal lengths,
magnifications and particle sizes provide information about achievable measurement depths
and help to design and adapt the optimal system for the desired experiment.

Keywords: 3D3C PTV, microfluidics, astigmatism

(Some figures in this article are in colour only in the electronic version)

1. Introduction dimensions involved, the velocity distribution can be very

complex and fully three dimensional, e.g., in geometries
In 1998, Santiago et al [18] introduced μPIV for microflow where obstacles are present. This is why even though 2D2C
measurements. Since this time, the method has evolved to methods are currently on the edge of μPIV technology, there
be a major flow diagnostics tool for microfluidic devices is a growing need for reliable and easily applicable 3D
used in recent lab-on-a-chip, micro-chemical and micro- techniques. Moreover, 3D techniques have the capability of
biological applications, among others. The method provides
overcoming the limitation that arises from the relatively large
a 2D2C (two-dimensional, two component) measurement of
thickness of the measurement plane (depth of correlation [14]),
the velocity field. Recent reviews of the general state of
the art of μPIV and its most relevant applications were which may introduce bias errors in regions of strong velocity
published by Lindken et al [11], and Lee and Kim [9] in gradients.
2009, and by Wereley and Meinhart [21] in 2010. Although The adaptation of macroscopic, multi-camera 3D3C
micro devices often provide laminar flows due to the small techniques to microscopic applications is somewhat restricted

0957-0233/11/015401+10$33.00 1 © 2011 IOP Publishing Ltd Printed in the UK & the USA
Meas. Sci. Technol. 22 (2011) 015401 C Cierpka et al

Fxz
A
B

B A
x B
lens CCD
z
Fyz A
A ay
B
ax
z = f(ax,ay).
B A
y CCD
lens +
z
cylindrical lens

Figure 1. Principle setup of the optical components. In the xz-plane only the spherical lens acts, whereas in the yz-plane both lenses are
acting.

by the limited optical access. Therefore, techniques for In previous works, either the difference ax − ay [2] or
3D3C velocity measurements in microflows, based on the the axis ratio ax /ay [3] was used for the allocation of the
defocusing of the particle images to encode the depth position actual z-position. With either one of these approaches, the
of tracer particles, have been proposed and investigated in the detected particles must be situated between the two in-focus
past. Extended summaries of different defocusing techniques planes to preserve unambiguity. The depth of the measurement
successfully applied to microfluidics so far [1, 7, 15, 16, 19, 22] volume is therefore defined by the distance between the two
were reported by Chen et al [2] and Cierpka et al [3]. in-focus planes z = Fxz − Fyz . In this paper, however, a
One major issue for these methods is their lack of a new calibration approach is presented, based on the theory of
simple and reliable calibration procedure. In this paper, particle imaging by Olsen and Adrian [14], which increases
a new approach for the calibration of the astigmatism (or the measurement volume beyond the two in-focus planes. The
wavefront deformation) particle tracking velocimetry (APTV) calibration functions are derived directly from the measured
is discussed. In APTV, the astigmatic aberration is used data with no need for an external calibration procedure. The
to encode the out-of-plane position of a tracer particle in method was tested for robustness and accuracy with both
the measurement volume. This technique can be easily synthetic and experimental data.
implemented in a standard microscope system by properly
placing a cylindrical lens in front of the CCD chip. The optical 2. Material and methods
effect on the particle images is schematically shown in figure 1.
In the xz-plane, the microscope objective (symbolized by 2.1. Tracer particle characteristics
the lens) determines the in-focus plane (Fxz ), whereas the A defocused particle image has a larger diameter and a lower
cylindrical lens acts, in a first approximation, as a kind of peak intensity since the light emitted (or scattered) by the
flat window since its axis is aligned with the x-direction. In particle is distributed over a larger area. The performance
that plane, particle A is closer to the in-focus plane and will of measurement techniques based on defocusing, such as
be imaged with a smaller diameter ax , while particle B is the APTV, strongly depends on the quality and signal to noise
further away, therefore more defocused, and ax of B appears ratio (SNR) of the defocused particle images provided by the
to be larger. In the lower part of figure 1, the view of the imaging system and the tracer particles. The high-quality
system is rotated by 90◦ . In the yz-plane, the curvature of fluorescent particles, typically used in microfluidics, generally
the cylindrical lens causes a shortening of the distance from provide high-quality particle images, with acceptable SNR up
the lenses to the in-focus plane (Fyz ), compared to the xz-plane. to high levels of defocusing, and are particularly well suited
In this view, particle B is closer to Fyz , which will result in for this type of application.
a smaller image ay than that of particle A. The reconstructed In figure 2 the peak intensities and particle image
image of the particles shows the effect of the different positions diameters, obtained from different types of tracer particles,
according to the in-focus planes: particles close to Fxz will are shown as a function of their distance z from the
appear as prolate and particles close to Fyz as oblate in the in-focus plane. The tracer particles are polystyrene latex
same image. This information can be evaluated to estimate spheres with diameters of dp = 1, 2, 5 and 10 μm, fabricated
the particle position in the out-of-plane direction [3]. If the by Microparticles GmbH. Those particles with dp = 1 and
position (x, y, z) of a particle in the measurement volume is 10 μm were dyed with Rhodamine B, whereas those with
known in two (or more) successive frames, a PTV algorithm dp = 2 and 5 μm were dyed with a new proprietary fluorescent
is capable of estimating the three-dimensional displacement dye. The particles were left to dry off on a microscope slide and
vector with subpixel accuracy. were then moved by a precise stepper in the viewing direction.

2
Meas. Sci. Technol. 22 (2011) 015401 C Cierpka et al

4500 100
1 μm (RhB)
2 μm ax 2 μm
4000 ay 2 μm
5 μm
10 μm (RhB) 80 ax 5 μm
3500 ay 5 μm
ax 10 μm (RhB)
3000 ay 10 μm (RhB)

ax /px, ay /px
60
I/counts

2500

2000
40
1500

1000 20

500

0
-50 0 50 -50 0 50
z/μm z/μm
Figure 2. Peak intensity I for different dp and fluorescent dyes (left), the width ax and height ay of the particle images (right) versus z for
different particle sizes. Rhodamine B labeled particles are indicated by RhB.

The same data processing used for the measurements, reported

2000
in section 3, was used to study the defocusing effects on the ax /px
140 ay /px
particle images. I /counts
In the left part of figure 2, the peak intensity I of the 120
particle images is shown against their depth position z. The
difference in the maximum intensities observed in the graphs 100
ax /px, ay /px

is due to the different dyes used.

The maximum value in the excitation spectrum for 80
1000
Rhodamine B is observed at a wavelength of λ = 510 nm,
60
whereas the new dye presents maximum absorption at
λ = 530 nm. Since the Nd:YAG laser’s emission wavelength,

I /counts
40
λ = 532 nm, is very close to the maximum absorption of the
new dye, the images of particles dyed with it (dp = 2 μm 20
and dp = 5 μm) are much brighter than the ones labeled with
Rhodamine B. This results in an significant increase of the 0 0
-50 0 50
SNR, which allows for reliable width and height detection of z/μm
highly defocused particles. This can be observed in the right Figure 3. Peak intensity I, width ax and height ay as a function of
plot of figure 2, where the width ax and height ay of the particle the z-position for the case of a 5 μm diameter particle viewed with a
images are shown. Since no cylindrical element is introduced magnification of M = 20 and f cyl = 150 mm. Highlighted with gray
in the light path, the in-focus planes for the x- and y-directions is the distance between both in-focus planes.
fall on top of each other.
Now, including a cylindrical element between the imaging
lens and the CCD, as outlined in section 1, results in different that the intensity of the particles beyond the two in-focus
in-focus planes for the x- and y-directions. This effect can planes remains high enough for the reliable determination of
be clearly seen in figure 3 by the change of ax and ay with the particle image shape, which allows for the use of the new
z for dp = 5 μm and M = 20. For ay , the in-focus plane is calibration approach discussed in section 2.4.
located at z = 0 μm and for ax at z = 32 μm. In contrast to
the intensity distributions measured without a cylindrical lens
(figure 2), two local maxima in the intensity distribution appear 2.2. Influence of different focal lengths and magnifications
in the vicinity of the respective in-focus planes. Between
these maxima, both respective intensities keep a fairly high The distance between the two in-focus planes z = Fxz −
level. When the particle images get out of focus for both, i.e. Fyz is the quantity that primarily defines the depth of the
the x- and y-directions, the intensity drops immediately. Due measurement volume and can be adjusted by the focal length
to the larger particle images, the peak intensity at the center of the cylindrical lens or the distance between the imaging lens
also decreases and is, with I ≈ 2000 counts, two times lower of the CCD and the cylindrical lens. In figure 4, z is plotted
than without the cylindrical lens (compare figure 2, left). The for f cyl = 100–700 mm at the magnifications of M = 10, 20
portion of the plot highlighted in gray indicates the region and 40. The analysis was performed using different particles
between the two in-focus planes. It is clear from the plot (filled symbols) and a calibration grid (open symbols).

3
Meas. Sci. Technol. 22 (2011) 015401 C Cierpka et al

channel or the detailed analysis, with maximum resolution in

120
Δ z = C *fcyl
-1.0 the z-direction, of large out-of-plane gradients present in shear
flows or near-wall regions.
100

d P = 10 μm
2.3. Mathematical model for the particle image diameter
80 d P = 5 μm
d P = 2 μm A model to predict the particle image diameter as a function of
Δ z/μm

the distance z from the in-focus plane was proposed by Olsen

60 and Adrian [14]. This model is built under the assumption
M = 10, extension that (i) the size of the tracer particles, (ii) diffraction and
-0.9
40 Δz = C *fcyl (iii) defocusing contribute to the particle image and that all
three terms can be approximated by a Gaussian function.
20
Now, according to Olsen and Adrian’s model, with the added
Δz = C *fcyl-0.7 assumption that the working distance of the lens is significantly
larger than z, the particle image diameter can be estimated by
0 the following equation [13, 17]:
0 200 400 600

fcyl /mm  2   2 −1
n0 n0
Figure 4. Measurement depth in air for different focal lengths of the a(z) = M dp2 + 1.49λ2 − 1 + 4z 2 − 1 ,
N A2 N A2
cylindrical lens for a magnification of 10 (right triangles), 20
(circles) and 40 (diamonds). Color coded is the particle diameter; no (1)
color indicates the use of a calibration grid; extension means an
increased distance between the cylindrical and spherical imaging where dp denotes the particle diameter, λ is the wavelength
lenses. of emitted light, n0 is the refractive index of the immersion
medium of the lens, and M and NA are the magnification
Using the thin lens formula, Chen et al [2] came up with and numerical aperture of the lens, respectively. Equation (1)
the conclusion that z ≈ u2m /f cyl , where um is the distance represents the arc of a hyperbola described by the general
from the center of the measurement range to the first principal formula

plane. The lines in figure 4 correspond to the best fit with f (z) = a 2 (z − F )2 + b2 , (2)
−1
n
Cfcyl , where C = const. The behavior of z ∼ Cfcyl was
found for the use of the infinity corrected microscope at where F is the position of the in-focus plane. As equation (1)
M = 10. However, using higher magnifications, n increases also remains valid, for the principal axis, when the cylindrical
to −0.9 at M = 20 and −0.7 at M = 40. For measurements in lens is included in the system, the following functions can
water at room temperature, zH2O = nH2O /n0 z ≈ 1.33z. be introduced to describe the particle image diameters in the
In order to cover channel widths in the xy-plane of up to several x- and y-directions:

hundreds of μm, the magnifications of M = 10 or 20 are used
ax (z) = c12 (z − Fxz )2 + c22 + c3
in most cases. This results in a measurement depth range of  (3)
z ≈ 30 μm for f cyl = 100, M = 20 to z ≈ 120 μm for ay (z) = c42 (z − Fyz )2 + c52 + c6 ,
f cyl = 150, M = 10, which covers the dimensions of a myriad
of microfluidic devices. Moreover, the depth can be increased where Fxz and Fyz are the positions of the respective in-
further by moving the imaging lens away from the cylindrical focus planes for the x- and y-directions, and c3 and c6 are
lens. For example, an extension of approximately 20 mm additional terms to account for an offset along the ordinate
between both lenses increases z to 38 μm for f cyl = 700 axis. To validate the accuracy of the model, the agreement
and M = 10. However, strong unwanted aberrations by the between the defocusing functions (3) and real data was tested
spherical imaging lens were observed for this configuration. for different sets of tracer particles and optical arrangements.
The resolution in the z-direction is determined by the limit of As an example, results obtained for the case presented in the
the CCD’s capability to detect changes in the image diameter previous section (dp = 5 μm, M = 20 and f cyl = 150 mm) are
as well as by the data processing algorithm. Nevertheless, it shown in figure 5.
can be adjusted by changing the magnification and numerical A confidence region, highlighted in gray in figure 5,
aperture. Increasing the magnification usually increases the where the fit model shows the best agreement can be clearly
numerical aperture as well and results in higher resolution identified. Outside this region the fit is still acceptable but
in the z-position determination for a given limit by the data a deviation from the model is observed due to higher order
processing. Nevertheless, the light intensity decreases rapidly image aberrations and low SNR, which influence the particle
with increasing image diameter and larger particle images image diameter determination. Nevertheless, it is evident from
on the CCD might restrict the measurement depth in this the figure that the confidence region extends far beyond the
case. Changing either the magnification or the focal length volume between Fxz and Fyz . Similar results were consistently
of the cylindrical lens makes the adaptation of the technique obtained for the other configurations as well, which provided
to different experiments quite flexible. This flexibility allows fitting correlation coefficients of R2 ≈ 1 in the confidence
for the instant characterization of the flow field across a whole region for all cases.

4
Meas. Sci. Technol. 22 (2011) 015401 C Cierpka et al

ay

ax, ay
ax
ay
fit data
fit

ay(Fxz)

ay(Fyz)

Fyz Fxz ax(Fxz) ax(Fyz)

z ax

Figure 5. Fit of ax and ay using equations (3) for the case of a 5 μm diameter particle viewed with a magnification of M = 20 and
f cyl = 150 mm. Data are presented for ax and ay as a function of z (left), and for ay as a function of ax (right).

2.4. Intrinsic calibration procedure The approach was initially tested on synthetic data. A
generic calibration curve was chosen, with ax, min = ay, min =
Equations (3) can be seen as parametric equations for a 1, and used to generate a set of 10 000 points, randomly
calibration curve in the (ax , ay ) space, as shown in the graph distributed between z = −0.5 and z = 1.5 (with Fxz = 1
on the right of figure 5. Given the position of the in-focus and Fyz = 0). An error for each ax and ay was randomly
planes Fxz and Fyz and the six coefficients ck , each position generated using a normal distribution with a standard deviation
z always identifies a unique point [ax , ay ] in the calibration of 20% of the minimum particle image diameter ax, min =
curve. Thus, theoretically, it is always possible to reverse ay, min . This error is quite large compared to what is obtained
the problem and go back to the z-position from a given couple in real measurements (usually an order of magnitude lower)
[ax , ay ]. However, when dealing with real data, a random error
but was chosen to test the procedure under extreme conditions.
[εax , εay ] is always associated with the measurement, and the
The calibration curve (in red) and the synthetic data
measured values will then be given by
generated are plotted in the graph on the left of figure 6. The
axi = ax (zi ) + εax i calibration procedure was then applied to obtain the estimated
(4) calibration curve (in black) and the estimated z-position of
ayi = ay (zi ) + εay i ,
the particles. The right plot in figure 6 shows the estimated
where i = 1, 2, . . . , N and N is the number of particles z versus the actual z-position. It can be seen from the graphs
in the recording. Equations (4) show that a measured that the procedure is able, first of all, to correctly estimate
couple [axi , ayi ] will fall outside the calibration curve thus the calibration curve from the data, and secondly, to correctly
excluding the possibility of directly deriving the z value. To determine the z-position of the particles without bias error.
overcome this problem, the measured couple [axi , ayi ] has The random error, which was quite large in this synthetic test
to be associated with a point on the calibration curve, i.e. case (σ = 0.065 for normalized z units), reduces as the random
[ax (zi ), ay (zi )] where zi will be the measured z-position for error in ax and ay determination is reduced.
the ith particle. Assuming the random error as normally and The intrinsic calibration procedure, presented so far, is
equally distributed in the x- and y-directions, it is proposed able to provide the z-position of particles from a data set of
here that a determination of zi be made by finding the value elliptical particle image recordings, obtaining the calibration
that minimizes the Euclidean distance between the two points curve directly from the same data set. Nevertheless, the
[axi , ayi ] and [ax (zi ), ay (zi )]. z-position is not provided in real units yet, but it is normalized
On the other hand, the same approach can be used to obtain based on the values assigned to Fxz and Fyz . To finally
the calibration curve from a given set of data. Assuming two convert z to real, physical space units, it is necessary to know
arbitrary values for the position of the in-focus planes, for the distance between the two in-focus planes z or the real
instance Fxz = 1 and Fyz = 0, the calibration curve can be z-position of at least two particles in the volume (such
determined by finding the six parameters ck in equation (3) as particles adhering to the top and bottom walls of a
that minimize the function microchannel).

N The intrinsic calibration was also tested on real data.
f (ck ) = (ax,i − ax (zi ))2 + (ay,i − ay (zi ))2 . (5) Results for the case of M = 20 and f cyl = 150 mm, with the
i particle diameters of dp = 2 μm and dp = 5 μm, are shown
Thus, the calibration curve, obtained from this procedure, can in figure 7. ax and ay are plotted versus the corresponding
be applied to the data once again to obtain the z-position of original (gray) and estimated (color) z-positions along with the
the particles. corresponding error. The experiments confirm the reliability

5
Meas. Sci. Technol. 22 (2011) 015401 C Cierpka et al

12 2

10 1.5

8
1

z measured
6
ay

0.5

4
0

2
-0.5

0
-1
0 5 10 -0.5 0 0.5 1 1.5
ax z

Figure 6. Application of the intrinsic calibration to synthetic data. Left: distribution of ax and ay with actual calibration curve (in red) and
estimated calibration curve (in black). Right: estimated z versus real z.

dp= 2 μm dp= 5 μm
50 5 50 5
ax /Px ax /Px
45 ay /Px 45 ay /Px
ax/Px (est) 4 ax/Px (est) 4
40 ay /Px (est) 40 ay /Px (est)
εcalib /μm εcalib /μm
35 3 35 3
ax /Px, ay /Px

ax /Px, ay /Px

30 30
2 2
εcalib /μm

εcalib /μm
25 25
1 1
20 20

15 0 15 0

10 10
-1 -1
5 5

0 -2 0 -2
0 20 40 0 20 40
z/μm z/μm

Figure 7. Application of the intrinsic calibration to real data. ax and ay versus the corresponding original (gray) and estimated (color)
z-position and the corresponding error are shown. The astigmatic particle images were obtained using M = 20 and f cyl = 150 mm and
particle with dp = 2 μm (left) and dp = 5 μm (right).

of this calibration approach, with an error in the determination information needed to correctly estimate the z-position in each
of z always under 1 μm. region of the CCD. An example of the application of this
An additional advantage of this calibration approach is procedure on real data is given in section 3.
that it can account for distortions in the xy-plane. Even
though objective lenses installed in high-quality microscopes
3. Experimental results
are corrected for image aberrations, the lenses that focus the
infinity corrected image onto the CCD of the camera are 3.1. Setup
often of standard quality and cause image aberrations in the
xy-plane, especially when coupled with standard cylindrical To illustrate the performance of the technique and to qualify the
lenses. In fact, the last study performed with astigmatism new calibration approach, the flow through a straight channel
flow diagnostics techniques [3] only used a small portion in with a rectangular cross section was investigated. This flow
the center of the sensor due to the strong image aberrations was chosen because it has a streamwise component, in the
outside this region. On the other hand, applying the intrinsic x-direction, that varies over the cross section but does not
calibration locally allows us to map the variations of the provide flow in the y- or z-directions. Thus, the measured
calibration coefficients through the whole area of the CCD. fluctuations of v and w could be used to estimate the
z can also be locally determined, thus providing all the uncertainty of the measurement technique. Furthermore, an

6
Meas. Sci. Technol. 22 (2011) 015401 C Cierpka et al

20

10

z/μm
0 200

Cylindrical -10
400 Z

lens 600
px
-20 800 x/
200
400 1000 Y
Figure 8. Photograph of the μPIV system. The extension to a fully 600
X
3D system is done by the implementation of a cylindrical lens in y /px 800
front of the CCD chip on the inverted microscope.
Figure 9. The two in-focus planes z0 x and z0 y with image
aberrations. The distance between two lines correspond to 0.5 μm.
analytical solution for the channel flow is provided by the
Hagen–Poiseuille law. The application of APTV to fully three- a calibration plate with a grid reticule. The objective was
dimensional flows is reported in several publications and the displaced along the z-direction, with steps of 1 μm, using the
reader is referred to [4, 5, 8, 12] for further details. optical stage of the Zeiss microscope, which has a spatial
The micro-channel was fabricated out of elastomeric translation resolution of 10 nm. The position of the in-focus
polydimethylsiloxane (PDMS) on a 0.6 mm thick glass plate by planes z0 x and z0 y was reconstructed using a local focus
the Institute for Microtechnology of the Technical University function based on the variance of the image intensity [20].
Braunschweig. It had a rectangular cross section of h × w = The two in-focus planes z0 x and z0 y are shown in figure 9,
200 × 500 μm2 and a length of 30 mm. In order to have a fully where z is already corrected for water. The mean distance
developed flow, the measurements were performed 20 mm is z = 35.1 μm. The deformation from a plane can clearly
from the inlet. For details on the procedure and fabrication, be seen. The standard deviations of z for both planes are
the interested reader is referred to Lesche et al [10]. std(z0 x ) = 1.5 μm and std(z0 x ) = 0.6 μm. As can be expected,
The fluid was deionized water, seeded with polystyrene the distortions become larger further away from the center.
latex particles (Microparticles GmbH) with a diameter of Since it is known that the distortions are due to the image
dP = 2 μm (standard deviation = 0.04 μm). The particle aberrations caused by the cylindrical and spherical lenses, an
diameter is 1% of the channel height, which is at the two ellipsoid surface fit was used to build a correction function for
orders of magnitude limit and was found to be small enough z0 x , z0 y = f (x, y).
to follow the flow precisely [11]. To calibrate the images in the xy-plane, the position of
The particles originally come dissolved in water at a the line intersections on the grid was determined with sub-
concentration of 25 mg/ml. All experiments were made with pixel accuracy via cross correlation, with a template of the
a concentration of 0.01 mg of particle solution, dissolved in intersection of two lines with Gaussian intensity distribution.
60 mg of water. The solution was stirred with an ultrasonic If the grid was properly aligned horizontally, no change for
homogenizer (model 300 by Biologics Inc.). Dissolution the intersection’s position was observed across the whole
during the experiment was not observed. scanning range z = 135 μm. This is probably due to
The flow was generated by pushing the homogeneous the careful precautions taken for the leveling of the channel
seeded solution through the channel using a high precision before performing the measurements and renders out-of-plane
neMESYS syringe pump (by Cetoni). The uniformity of aberrations for the in-plane coordinates negligible.
the provided fluid velocity versus time was checked in a To account for the in-plane image aberrations, the mean
preliminary study using conventional μPIV. distance between each intersection point and its closest
The measurement system is installed on an optical table neighbors was determined in the x- and y-directions. The
to damp vibrations and consists of an Axio Observer Z.1 mean conversion from real space into image space was
inverted microscope manufactured by Carl Zeiss AG, a two
cavity frequency-doubled Litron Nano S Nd:YAG laser system kx = 0.620 +0.005
−0.003 and ky = 0.751 −0.008 μm/pixel.
+0.010
(6)
and a 12-bit, 1376 × 1040 pixel2 , interline transfer CCD The absolute difference is 1.3% for kx and 2.4% for ky across
camera (Sensicam QE, PCO GmbH). A photograph of the the span of the entire sensor. Once again, an ellipsoid surface
full experimental setup is shown in figure 8. fit was used to provide a function for kx , ky = f (x, y). This fit
The imaging optics arrangement in the microscope function was later used to correct the in-plane positions of the
consists of an EC Plan-Neofluar objective lens from Carl Zeiss particle images.
with a magnification of M = 20, an imaging lens (Zeiss part The recordings were made using the DaVis 7.4 software
number 426113) with a magnification of M = 0.63 and a package from LaVision. To resolve the whole channel, six
cylindrical lens with focal length f cyl = 150 mm. The distance successive positions, each 25 μm above the next, were selected
between the two in-focus planes z was determined using to scan across the full cross section with overlapping volumes

7
Meas. Sci. Technol. 22 (2011) 015401 C Cierpka et al

Figure 10. Scatter plot of ax and ay for valid (blue ≈73%) and rejected (gray ≈27%) particles. ax and ay of particles in the neighborhood of
two different [x, y] coordinates are also shown in two different colors (red and green) together with the corresponding calibration curves,
coded with the respective colors.

and no error due to the depth of correlation. For each volume, and a first outlier detection was performed by rejecting
2000 double frames with approximately 55–45 particle images particles with [ax , ay ] values which strongly deviated from
per double frame were taken and processed. This results the estimated calibration curve, which resulted in the removal
in approximately 0.000 06 particle images per pixel for the of approximately 6000 outliers. The calibration was then
illuminated area of the sensor. The low particle image refined by locally calibrating data from particles in different
density was chosen to avoid overlapping particle images which [x, y] neighborhoods. Second-order polynomial surfaces
require special treatment in the data processing. Furthermore were used to fit the coefficient parameters ck and to get
a nearest-neighbor algorithm is used for the later particle an analytical expression for the calibration functions in the
tracking. For reliable detection of valid particle links the x- and y-directions. Next, another, more restrictive, outlier
seeding density has to be low. Since half of the measurement detection filter was then performed rejecting particles with
volume lies outside the channel in the top and bottom scanning [ax , ay ] at a distance >1 px from the corresponding calibration
position, particle density is lower in the near-wall regions. curve, resulting in approximately 85 000 valid particles. The
The time delay between the two successive frames was set to scatter plot of the [ax , ay ] for valid (blue) and rejected (gray)
t = 200 μs. particles is shown in figure 10. In the figure, one can also
note how the [ax , ay ], corresponding to particles detected in
different xy-regions, follow different calibration curves and
3.2. Results
how the presented calibration approach is able to follow these
Image preprocessing was applied for background suppression changes. Altogether, the [ax , ay ] of the valid particles present a
by subtracting a sliding minimum over time. A threshold filter standard deviation from the respective calibration curve of σ <
was used to identify possible particle image regions to be used 0.25 pixel.
by the following processing steps on the original images with For the velocity determination, a simple nearest-neighbor
suppressed backgrounds. A wavelet algorithm was applied for approach was used. The data of all individual particles were
the xy-position determination of the particles and the particle filtered using a local universal PTV outlier detection algorithm
image width and height were determined by an auto correlation [6]. The authors proposed a weighting of the neighboring
method. The interested reader is referred to [3] for further values by their distances. The normalized residuum or
details. fluctuation at the position r0∗ was set to be lower than 2
The algorithm identified a total number of approximately for valid data, taking the ten closest neighboring points into
116 000 particles, each associated with an [ax , ay ] value. consideration. Rejecting data with a residuum higher than 2
The intrinsic calibration was run on the entire set of data for one of the three velocity components results in a outlier

8
Meas. Sci. Technol. 22 (2011) 015401 C Cierpka et al

0.8

u /mms-1: -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 0.6

u* vs. z/h

*
u
0.4
150
z /μm

u* vs. y/w u* vs. z/h

100 *
u vs. y/w
0.2 theoretical profile
50

100 200 300 400 0

0 0.2 0.4 0.6 0.8 1
y /μm z/h, y/w

Figure 11. Streamwise velocity in the cross section of the channel (left) and profiles of the normalized streamwise velocity u∗ versus z/h at
y/w = 0.5 and u∗ versus y/w at z/h = 0.5 (right).

removal of 5%. In total, 81 000 vectors were considered to be motion. The mean standard deviation of the measured in-plane
valid and taken for the following analysis. The data were then component v was σ v = 0.18 mm s−1 , which is comparable to
averaged in the streamwise direction and interpolated onto a μPIV measurements [4]. On the other hand, for the out-
Cartesian grid with grid size y = z = 7.5 μm and 25% of-plane component w, it was σ w = 0.73 mm s−1 . The
overlap. Using this interpolation, approximately 110 single uncertainty of the measured w, equal to 3.72% of uc , is
PTV measurements points contribute to the mean at each point therefore four times higher than for the in-plane component
on the Cartesian grid. The streamwise velocity over the cross with 0.9% of uc . This value is still lower than the out-of-plane
section of the channel is shown on the left side of figure 11. error for stereo μPIV, which is supposed to be ten times the
The expected parabolic Poiseuille flow profile can clearly be in-plane error due to the small viewing angle and the slight
seen. On the right side of the same figure, profiles of the misalignment of the two in-focus planes for the respective
normalized streamwise velocity u∗ = u/uc can be seen. The cameras [4].
agreement is quite good in both directions. The profiles clearly
show that the quality of the data is comparable for both the 4. Conclusion and outlook
in-plane and out-of-plane measurements. One big advantage
of the method is that no bias due to depth of correlation The applicability of the APTV technique to measure the whole
is present and the gradients close to the wall are not three-dimensional velocity field in a volume was shown in
underestimated due to spatial averaging like in μPIV. several studies [2–5, 8, 12]. For the calibration of most of these
approaches, only the linear part in between the two in-focus
planes could be used and the measurement volume depth was
3.3. Error analysis
therefore limited. In the paper at hand, an intrinsic calibration
The error for the in-plane position was estimated to be below procedure based on a local fit of the defocused particle images
0.1 pixel for an SNR of ≈ 6 for the measurements presented was proposed. The main advantages of this new calibration
here [3]. This results in an in-plane error of 0.075 μm. For the procedure are as follows.
out-of-plane component, the error was calculated by adding the • The measurement volume depth is not limited by the in-
measurement uncertainty to that of the measurements where focus planes of the setup and can be substantially extended
the z-position was known. First, the uncertainty was estimated by the use of high-quality fluorescent particles.
using an allowed residual of 0.5 pixel for the calibration • The intrinsic calibration is based on the measurement
function. The corresponding difference between the measured itself, and the optical path through the different media
z-position and the z-position for the maximum residual was does not change between calibration and experiment.
then determined and the standard deviation of the difference No complicated scanning procedure is needed. The
was σ z = 0.28 μm. Now, the estimated error in the z-position image preprocessing is the same for calibration and
for measurements, when the exact position of the particles was measurements.
known, was 1 μm. • All image aberrations are taken into account. The
As a measure for the uncertainty of the APTV technique, difference in the defocus function for different
the standard deviation of the single measurements, σ (vi − xy-positions, as well as the curvature of the in-focus
v mean ), that contribute to the mean of a certain grid point can planes, contributes to the calibration. The whole CCD
be taken. In the laminar channel flow, v and w should have sensor can therefore be used for reliable measurements.
a zero mean and the scatter of the single PTV data points • All measured data points are taken into account; therefore,
is caused purely by the measurement technique or Brownian the highest statistical relevance is achieved.

9
Meas. Sci. Technol. 22 (2011) 015401 C Cierpka et al

Furthermore, the APTV technique does not suffer from [5] Cierpka C, Tschulik K, Segura R, Uhlemann M and Kähler C J
bias error due to depth of correlation and spatial averaging [3] 2010 Zeitaufgelöste 3D3C Geschwindigkeitsfeldmessungen
and it only requires one observation direction. Compared mit der fernmikroskopischen Astigmatismus PTV zur
Analyse der elektrochemischen Kupferabscheidung 18.
to other defocusing techniques, the particle shape can be Fachtagung ‘Lasermethoden in der Strömungsmesstechnik’
adjusted to show a larger size in the flow direction close to the (Cottbus, Germany, 7–9 September)
wall, which allows for very precise measurements at channel [6] Duncan J, Dabiri D, Hove J and Gharib M 2010 Universal
walls. The method shows further potential in the analysis outlier detection for particle image velocimetry (PIV) and
of mixing devices. Since the particle positions are known particle tracking velocimetry (PTV) data Meas. Sci.
Technol. 21 057002
in the whole volume, the particle distribution can be used to [7] van Hinsberg N P, Roisman I V and Tropea C 2008
reconstruct interfaces between fluids and to characterize the Three-dimensional, three-component particle imaging using
mixing process at the micro scale [12]. Moreover, it can also two optical aberrations and a single camera 14th Int. Symp.
be applied for time-resolved measurements [8]. on Applications of Laser Techniques to Fluid Mechanics
With the new calibration procedure, the method can (Lisbon, Portugal, 7–10 July)
[8] Kumar A, Cierpka C, Williams S J, Kähler C J and
be easily applied to existing μPIV setups without expert Wereley S T 2010 3D3C velocimetry measurements of an
knowledge in optics and data processing. electrothermal microvortex using wavefront deformation
It should also be highlighted that even though the PTV and a single camera Microfluid. Nanofluid. at press
presented calibration approach was qualified for an optical doi:10.1007/s10404-010-0674-4
system including a microscope, the concept behind the [9] Lee S J and Kim S 2009 Advanced particle-based velocimetry
techniques for microscale flows Microfluid. Nanofluid.
procedure remains valid for all measurement scales (micro and 6 577–88
macro) as long as the quality of the ellipsoidal particle images [10] Lesche C, Demming S, Kampen I, Kwade A and
is high enough. Taking this calibration approach as a starting Büttgenbach S 2008 Dispersion of nanoparticles in
point, there is high potential for the significant improvement micro-channel geometries for pharmaceutical screening
and extension of the method. For instance, one could use applications Proc. 1st European Conf. on Microfluidics
(Bologna, Italy, 10–12 December)
the particle image intensity as a weighting aid to improve [11] Lindken R, Rossi M, Grosse S and Westerweel J 2009
the accuracy of the particle image diameter determination, Micro-particle image velocimetry (μPIV): recent
or use high-speed multi-frame approaches to reconstruct the developments, applications, and guidelines Lab Chip
particle’s trajectories. 9 2551–67
In general, the simplicity, robustness, flexibility and [12] Mastrangelo F, Rossi M, Cierpka C, Segura R and Kähler C J
2010 Reconstruction of the interface between two fluids in
potential for further developments make this method a good microfluidic devices using astigmatic particle imaging 2nd
candidate to be used as a tool for 3D3C velocity measurements European Conf. on Microfluidics (Toulouse, France, 8–10
in microfluidics, and eventually in a broad range of other December)
applications. [13] Meinhart C D and Wereley S T 2003 The theory of
diffraction-limited resolution in micro particle image
velocimetry Meas. Sci. Technol. 14 1047–53
Acknowledgments [14] Olsen M G and Adrian R J 2000 Out-of-focus effects on
particle image visibility and correlation in microscopic
Financial support from Deutsche Forschungsgemeinschaft particle image velocimetry Exp. Fluids 29 S166–74
(DFG) in the frame of the priority program SPP 1147 and [15] Pereira F and Gharib M 2002 Defocussing digital particle
the Mikropart research group is gratefully acknowledged. image velocimetry and three-dimensional characterization
of two phase flows Meas. Sci. Technol. 13 683–94
The authors would also like to thank the Institute for
[16] Ragan T, Huang H, So P and Gratton E 2006 3D particle
Microtechnology at the Technical University of Braunschweig tracking on a two-photon microscope J. Fluorescence
for their kind support. 16 325–36
[17] Rossi M, Cierpka C, Segura R and Kähler C J 2010 On the
References effect of particle image intensity and image preprocessing
on depth of correlation in micro-PIV measurements 15th
[1] Angarita-Jaimes N, McGhee E, Chennaoui M, Campbell H I, Int. Symp. on Applications of Laser Techniques to Fluid
Zhang S, Towers C E, Greenaway A H and Towers D P Mechanics (Lisbon, Portugal, 5–8 July)
2006 Wavefront sensing for single view three-component [18] Santiago J G, Wereley S T, Meinhart C D, Beebe D J
three-dimensional flow velocimetry Exp. Fluids 41 881–91 and Adrian R J 1998 A particle image velocimetry system
[2] Chen S, Angarita-Jaimes N, Angarita-Jaimes D, Pelc B, for microfluidics Exp. Fluids 25 316–9
Greenaway A H, Towers C E, Lin D and Towers P D 2009 [19] Shinohara K, Sugii Y, Jeong J H and Okamoto K 2005
Wavefront sensing for three-component three-dimensional Development of 3D scanning micro-PIV system 6th Int.
flow velocimetry in microfluidics Exp. Fluids 47 849–63 Symp. on PIV (Pasadena, USA, 21–23 September)
[3] Cierpka C, Hain R and Kähler C J 2010 A simple single [20] Sun Y, Duthaler S and Nelson B J 2004 Autofocusing in
camera 3C3D velocity measurement technique without computer microscopy: selecting the optimal focus
errors due to depth of correlation and spatial averaging for algorithm Microsc. Res. Tech. 65 139–49
micro fluidics Meas. Sci. Technol. 21 045401 [21] Wereley S T and Meinhart C D 2010 Recent advances in micro
[4] Cierpka C, Rossi M, Segura R, Mastrangelo F and Kähler C J particle image velocimetry Annu. Rev. Fluid Mech.
2010 Comparative study of the uncertainty of stereoscopic 42 557–76
micro-PIV, wavefront-deformation micro-PTV, and standard [22] Yoon S Y and Kim K C 2006 3D particle and 3D velocity field
micro-PIV 15th Int. Symp. on Applications of Laser measurement in a microvolume via the defocussing concept
Techniques to Fluid Mechanics (Lisbon, Portugal, 5–8 July) Meas. Sci. Technol. 17 2897–905

10