FlowMaster D80

LaVision
We count on Photons
Product-Manual
FlowMaster
Item-Number(s): 1105011-4
Product-Manual for DaVis 8.0
LaVision GmbH, Anna-Vandenhoeck-Ring 19, D-37081 Gttingen
Produced by LaVision GmbH, Gttingen
Printed in Germany
Gttingen, February 14, 2011
Document name: 1003005_FlowMaster_D80.pdf
Contents
Contents
1 Safety Precautions 7
1.1 Laser Safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Seizures Warning . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Camera / Image Intensifier Safety . . . . . . . . . . . . . . . . . 9
2 Introduction 11
3 How PIV Works 13
3.1 Auto correlation (single frame / double exposure) . . . . . . . . 15
3.2 Cross correlation (double frame / double exposure) . . . . . . . 16
4 Recording 17
4.1 Live mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.1.1 Camera-AOI selection . . . . . . . . . . . . . . . . . . . . 18
4.1.2 Show information . . . . . . . . . . . . . . . . . . . . . . 18
4.2 dt optimizer wizard . . . . . . . . . . . . . . . . . . . . . . . . . 21
5 Vector calculation 23
5.1 Vector calculation - double frames . . . . . . . . . . . . . . . . . 24
5.2 Vector calculation - time series of single frames . . . . . . . . . 26
6 Image Preprocessing 29
6.1 Inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6.2 Subtract a sliding background . . . . . . . . . . . . . . . . . . . 30
6.3 Subtract an offset . . . . . . . . . . . . . . . . . . . . . . . . . . 30
6.4 Particle intensity normalization (min/max-filter) . . . . . . . . . 31
6.5 Test current settings . . . . . . . . . . . . . . . . . . . . . . . . 32
7 Define Mask 33
7.1 Operations to mask out an image . . . . . . . . . . . . . . . . . 33
7.1.1 Define mask . . . . . . . . . . . . . . . . . . . . . . . . . 35
7.1.2 Geometric mask . . . . . . . . . . . . . . . . . . . . . . . 35
7.1.3 Algorithmic mask . . . . . . . . . . . . . . . . . . . . . . 39
7.1.4 Load mask from file . . . . . . . . . . . . . . . . . . . . . 42
7.1.5 Advanced mask settings . . . . . . . . . . . . . . . . . . 43
3
Contents
8 Vector Calculation Parameter: PIV 47
8.1 Use Mask . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
8.2 Correlation Mode . . . . . . . . . . . . . . . . . . . . . . . . . . 49
8.3 Data source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
8.4 Iterations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
8.5 Window Size and Weight . . . . . . . . . . . . . . . . . . . . . . 53
8.5.1 Adaptive PIV . . . . . . . . . . . . . . . . . . . . . . . . . 56
8.6 Overlap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
8.7 Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
8.7.1 Use image correction . . . . . . . . . . . . . . . . . . . . 59
8.7.2 Pixelwise window deformation . . . . . . . . . . . . . . . 60
8.7.3 Display intermediate results . . . . . . . . . . . . . . . . 61
8.7.4 Vector scale . . . . . . . . . . . . . . . . . . . . . . . . . . 61
8.7.5 High accuracy mode for final passes . . . . . . . . . . . . 62
8.8 Multi-pass options . . . . . . . . . . . . . . . . . . . . . . . . . . 64
8.8.1 Initial window shift for image reconstruction . . . . . . . 65
8.8.2 Correlation function . . . . . . . . . . . . . . . . . . . . . 66
8.8.3 Deformed interrogation windows . . . . . . . . . . . . . . 71
8.8.4 3D vector validation . . . . . . . . . . . . . . . . . . . . . 71
8.9 Multi-pass postprocessing . . . . . . . . . . . . . . . . . . . . . 72
9 Vector Calculation Parameter: PIV Sum Of Correlation 73
9.1 Description of method . . . . . . . . . . . . . . . . . . . . . . . 74
9.2 When to use this method? . . . . . . . . . . . . . . . . . . . . . 74
9.3 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
10Vector Calculation Parameter: PIV + PTV 79
10.1Description of the method . . . . . . . . . . . . . . . . . . . . . 79
10.2PTV Calculation Parameter . . . . . . . . . . . . . . . . . . . . . 80
10.3PTV Postprocessing . . . . . . . . . . . . . . . . . . . . . . . . . 81
10.4When to use this method? . . . . . . . . . . . . . . . . . . . . . 83
11Vector Calculation Parameter: PIV on GPU 85
12Vector Postprocessing 89
12.1Allowable vector range . . . . . . . . . . . . . . . . . . . . . . . 89
12.2Peak Ratio or Correlation Value . . . . . . . . . . . . . . . . . . 90
12.3Median Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
12.4Remove groups . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
12.5Fill-up empty spaces (interpolation) . . . . . . . . . . . . . . . . 98
12.6Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4
Contents
12.7Make mask permanent . . . . . . . . . . . . . . . . . . . . . . . 100
12.8Apply allowable vector range again . . . . . . . . . . . . . . . . 100
12.9Troubleshooting . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
13Stereoscopic PIV 103
13.1Image correction for planar PIV . . . . . . . . . . . . . . . . . . 103
13.2Stereoscopic viewing . . . . . . . . . . . . . . . . . . . . . . . . 105
13.3Camera Arrangement . . . . . . . . . . . . . . . . . . . . . . . . 105
13.4Scheimpflug criterion . . . . . . . . . . . . . . . . . . . . . . . . 107
13.5Calibration for stereoscopic PIV . . . . . . . . . . . . . . . . . . 108
13.6Vector Field Computation . . . . . . . . . . . . . . . . . . . . . . 110
13.6.1Camera Mode . . . . . . . . . . . . . . . . . . . . . . . . . 111
13.6.2Vector Calculation Parameter . . . . . . . . . . . . . . . . 111
13.6.3Procedure for Stereo Vector Calculation . . . . . . . . . . 112
14Stereoscopic PIV Self-Calibration 117
14.1Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
14.1.1Disparity Vector . . . . . . . . . . . . . . . . . . . . . . . 118
14.1.2Disparity Map . . . . . . . . . . . . . . . . . . . . . . . . . 118
14.1.3Correlation Map . . . . . . . . . . . . . . . . . . . . . . . 120
14.2Requirements for the self-calibration procedure . . . . . . . . . 120
14.3Step 1: setup the initial disparity map . . . . . . . . . . . . . . 121
14.4Step 2: calculate disparity map / update calibration . . . . . . . 124
14.4.1Parameter Setup . . . . . . . . . . . . . . . . . . . . . . . 125
14.4.2Sum of correlation / correlation of sum and passes . . . . 125
14.4.3Advanced settings . . . . . . . . . . . . . . . . . . . . . . 126
14.4.4Info text output . . . . . . . . . . . . . . . . . . . . . . . . 126
14.4.5Accept / discard new calibration . . . . . . . . . . . . . . 127
14.5Step 3: refine new calibration . . . . . . . . . . . . . . . . . . . 128
14.5.1Advanced settings: Refit camera parameter . . . . . . . 129
14.5.2When should internal camera parameters be refitted? . . 129
14.6Warning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
14.6.1Old results become invalid . . . . . . . . . . . . . . . . . 130
14.6.2Original scaling pixel/millimeter is changed . . . . . . . . 130
14.7Calibration History . . . . . . . . . . . . . . . . . . . . . . . . . 131
15Proper Orthogonal Decomposition (POD) 133
15.1Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
15.2POD Wizard step 1: Decomposition . . . . . . . . . . . . . . . . 134
15.2.1The data source . . . . . . . . . . . . . . . . . . . . . . . 135
5
Contents
15.2.2The window manager . . . . . . . . . . . . . . . . . . . . 135
15.2.3The decomposition subdialog . . . . . . . . . . . . . . . . 135
15.2.4Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
15.2.5Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
15.3POD Wizard step 2: Approximation . . . . . . . . . . . . . . . . 138
15.3.1The data source . . . . . . . . . . . . . . . . . . . . . . . 139
15.3.2The window manager . . . . . . . . . . . . . . . . . . . . 139
15.3.3The Approximation subdialog . . . . . . . . . . . . . . . . 139
15.3.4Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
15.3.5Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
16Display Options 141
16.1Correlation Map & Vector Edit . . . . . . . . . . . . . . . . . . . 141
16.2Scatter Plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
16.3Probability Density Function . . . . . . . . . . . . . . . . . . . . 145
17Vector related Expert Settings 149
17.1How to open the dialog . . . . . . . . . . . . . . . . . . . . . . . 149
17.2What can be set in the dialog . . . . . . . . . . . . . . . . . . . 150
18Customer Service 153
18.1Order and Dongle Number . . . . . . . . . . . . . . . . . . . . . 153
18.2Customer Settings . . . . . . . . . . . . . . . . . . . . . . . . . . 154
18.3Log.txt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
18.4Export data set for support . . . . . . . . . . . . . . . . . . . . . 156
18.5Camera Activity Logging . . . . . . . . . . . . . . . . . . . . . . 158
18.6Shipment of defective items . . . . . . . . . . . . . . . . . . . . 160
6
1 Safety Precautions
Before working with your LaVision system we recommend to read the
following safety precautions. Observing these instructions helps to avoid
danger, to reduce repair costs and downtimes and to increase the reliability
and life of your LaVision system.
1.1 Laser Safety
If a laser
1
is integrated in your system it is important that every person
working with it has fully read and understood these safety precautions and
the laser manual of the specific laser.
Lasers included in LaVision systems may belong to CLASS 4 laser de-
vices, which are capable of emitting levels of both visible and invisible ra-
diation that can cause damage to the eyes and skin. It is absolutely neces-
sary that protective eyewear with a sufficiently high optical density is worn
at any time when operating the laser. The goggles must protect against all
wavelengths that can be emitted, including harmonics. See your Lasers
manual for further details.
Class 4 laser beams are by definition a safety and fire hazard. The use of
controls, adjustments or performance of procedures other than those spec-
ified in the LaVision manual and laser manual may result in hazardous
radiation exposure.
AVOID EYE AND SKIN EXPOSURE TO DIRECT OR SCATTERED RADIATION.
FOLLOWTHE INSTRUCTIONS YOU CAN FIND IN THE CORRESPONDING LASER
MANUAL FOR PROPER INSTALLATION AND SAFE OPERATION. USE PROTEC-
TIVE EYE WEAR ALL THE TIME WHEN OPERATING THE LASER.
Important instructions for safe laser handling:
Before operating the laser contact your laser safety officer.
1
In the following laser means any kind of laser, in particular Nd:YAG- and dye laser as
well as Optical Parametric Oscillators at any wave-length and output-energy.
7
Read and understand the instruction manual of the particular type of
laser. Take special care with respect to laser emission, high voltage
and hazardous gases if in use.
Declare a controlled access area for laser operation. Limit access to
trained people. Never operate the laser in a room where laser light
can escape through windows or doors. If possible, cover beam paths
to avoid obstacles getting into the beam.
Provide adequate and proper laser safety-goggles to all persons
present who may be exposed to laser light. The selection of the gog-
gles depends on the energy and the wavelength of the laser beam
as well as the operation conditions. Check the Lasers manual for a
detailed description.
While working with lasers do not wear reflective jewelry like watches
and rings, as these might cause accidental hazardous reflections.
Avoid looking at the output beam, even diffuse reflections can be dan-
gerous.
Operate the laser at the lowest beam intensity possible.
Avoid blocking the output beam or reflections with any part of the
body. Use beam dumps to avoid reflections from the target.
Wear clothes and gloves which cover arms and hands to avoid skin
damage when handling in the optical path. Especially UV-radiation
can cause skin cancer.
1.2 Seizures Warning
WARNING: HEALTH HAZARD! STROBE LIGHTING COULD TRIGGER SEIZURES
Some people (about 1 in 4000) may have seizures or blackouts triggered
by flashing lights or patterns. This may occur when viewing stroboscopic
lights or objects illuminated by such devices, even if a seizure has never
been previously experienced. Anyone who has had a seizure, loss of aware-
ness, or other symptoms linked to an epileptic condition should consult a
doctor before operating systems which include flashing lights, strobe lights,
or a pulsed or modulated laser.
Stop operating the system immediately and consult a doctor if you have
one of the following symptoms:
8
1.3 Camera / Image Intensifier Safety
convulsions, eye or muscle twitching, loss of awareness, altered vi-
sion, involuntary movements, disorientation
To reduce the likelyhood of a seizure when operating a system:
Do not look directly at flashing light sources or on illuminated objects,
e.g. into a strobe light or a flashing LED panel.
Operate the system in a well-lit room.
Take frequent breaks in normally illuminated areas.
1.3 Camera / Image Intensifier Safety
The camera integrated in your system is based on a CCD (Charge Coupled
Device) or CMOS (Complementary Metal-Oxide Semiconductor) sensor with
high resolution and high sensitivity. Optionally your system is equipped
with a built-in or external image intensifier.
A LASER BEAM FOCUSED ON THE CHIP OR INTENSIFIER, EITHER DIRECTLY
OR BY REFLECTION, CAN CAUSE PERMANENT DAMAGE TO THE CHIP OR IN-
TENSIFIER. ANY LASER POWERFUL ENOUGH TO PRODUCE LOCALIZED HEAT-
ING AT THE SURFACE OF THE CHIP OR INTENSIFIER WILL CAUSE DAMAGE
EVEN WHEN THE CAMERA OR INTENSIFIER POWER IS OFF. A CHIP OR IN-
TENSIFIER DAMAGED BY LASER LIGHT IS NOT COVERED BY ITS WARRANTY.
Important instructions for safe camera handling:
Fully read and understand the instruction manual of the specific type
of camera.
Put the protection cap on the camera lens whenever you do not take
images, especially when the laser beam is adjusted. Switching off the
camera / image intensifier does not protect the chip from damage by
laser light.
Use full resolution of the sensor and always read out the complete
chip to have control of the intensity on all areas of the sensor.
Make sure that no parts of the image are saturated, i.e. the intensity
is below maximum gray level (< 4095 counts for a 12 bit camera,
< 65535 counts for a 16 bit camera, ...).
9
Start measurements with the lowest laser power and a small aperture
of the camera lens.
Increase laser power step by step and check the intensity on the cor-
responding image. Make sure that the sensor does not run into satu-
ration.
Bright parts in the experiment, like reflections on walls or big parti-
cles, will limit the maximum laser power. Modify the optical arrange-
ment of your setup in order to remove bright reflections from the
camera image.
10
2 Introduction
Welcome to the LaVision FlowMaster Software manual. This manual
explains the additional software functions that are available in the DaVis
FlowMaster module. For details on universal DaVis functionalities please
consult the DaVis manual.
In the How PIV works chapter you find a short description of the working
principle of Particle Image Velocimetry. You will find the basic experimental
setup and the fundamental steps for the vector calculation.
The Recording dialog is the central dialog for the optimization of recording
parameter. From here you have access to hardware components like laser
and camera.
The PIV Vector Calculation chapter describes the options and parameter
for the vector calculation. DaVis offers access to these parameter in the
Processing dialog. These parameter need to be optimized for the specific
particle image quality. The Stereoscopic PIV chapter describes the partic-
ularities of stereoscopic PIV measurements. You find different experimental
setups and additional information regarding image acquisition and vector
calculation with stereo FlowMaster systems.
11
2 Introduction
12
3 How PIV Works
Many technical and scientific developments require a measuring technique
that can measure the velocity distribution across an extended area of a
flow field. This can be achieved by scanning a point velocity probe across
the flow but then the instantaneous structure is lost and only the average
flow field is obtained. Flow visualization techniques can often reveal in-
stantaneous flow structure but are only qualitative or semi-quantitative at
best. An optical, non-intrusive method that is related to both flow visual-
ization and optical point techniques has developed over the last 20 years
called Particle Image Velocimetry (PIV). This technique can provide an ac-
curate quantitative measure of the instantaneous flow velocity field across
a planar area of a flow field.
t
PIV laser
sheet
optics
seeded
flow
field of
view
lens
interrogation
window
image
plane
y
x
1
t
2
light sheet
Figure 3.1: General setup of a 2D PIV System.
PIV has become an important tool for quantitative and instantaneous mea-
surement of laboratory flows. The flow is seeded with tiny, neutrally buoy-
ant particles - so called tracers - e.g. oil or water aerosols in air and solid
particles in fluids or flames. Using a light sheet, formed by passing a dou-
ble pulsed laser beam through an optical arrangement including cylindrical
13
3 How PIV Works
lenses, the particles in the flow are illuminated twice with a small time
separation between.
The displacement of particles in the time between the laser pulses is
recorded as either a single image exposed twice or as a pair of two sin-
gle exposure images. The recorded particle displacement field is measured
locally across the whole field of view of the images, scaled by the image
magnification and then divided by the known pulse separation to obtain
flow velocity at each point. A camera positioned typically perpendicular
to the plane of the light sheet is shuttered to capture the light scattered
from the particles. Depending on the flow velocity and the factor of magni-
fication of the camera lens the delay of the two pulses have to be chosen
such that adequate displacements of the particle images on the CCD are
obtained. From the time delay between the two illuminations and the dis-
placement of the tracers velocity vectors can be calculated.
For the evaluation of the particle images it is assumed that the tracers fol-
low the flow into the local flow velocity between the two illuminations. The
(digital) PIV recording is divided in small subareas - so called interrogation
windows. Using statistical correlation techniques one local displacement
vector is determined for each interrogation window. For this reason the size
of this interrogation cell is selected such that all particles within this area
have moved homogeneously in the same direction and the same distance.
For good results the number of particles within one interrogation cell should
be at least ten.
Figure 3.2: Seeded vortex and corresponding flow field.
The evaluation of the particle images depends on the way these images
have been recorded by the used camera. One possibility is to record
the scattered light of both illuminations in one frame what is called sin-
gle frame / double exposure. These pictures can be evaluated by auto-
14
3.1 Auto correlation (single frame / double exposure)
correlation. The other possibility is to record the scattered light from the
first illumination in one frame and the scattered light from the second illu-
mination in another frame. This is called double frame / double exposure.
These double frame images can be evaluated by cross-correlation.
3.1 Auto correlation (single frame / double expo-
sure)
The scattered light from first and second exposure of the particles is
recorded in one image. The complete image is subdivided in so called
interrogation windows and each window is evaluated by auto correlation.
auto
correlation
peak search
local velocity
vector v(t)
dx
dy
Figure 3.3: Evaluation of PIV recordings using Auto-Correlation.
The auto correlation function is characterized by two identical correlation
peaks rotationally symmetrical about the highest central peak indicating
zero displacement. This is a consequence as you can not detect the sign
of the displacement because you do not know which particle is illuminated
by the first and the second laser pulse. So the information from the auto-
correlation is ambiguous and not conclusive if you can not use some a priori
information about the observed flow. Also the detection of very small dis-
placements is a problem as in this case the correlation peaks are very close
to the central peak. Compared to cross-correlation the auto-correlation
peaks are significantly smaller. High noise increases the possibility that the
displacement correlation peak disappears in the background.
15
3 How PIV Works
3.2 Cross correlation (double frame / double ex-
posure)
The scattered light from first and second exposure of the particles is
recorded in two different images. The complete image is subdivided in
interrogation windows and each window is evaluated by cross correlation.
t
t+dt
cross
correlation
peak search
local velocity
vector v(t)
dx dy
Figure 3.4: Evaluation of PIV recrodings using Cross-Correlation
For cross-correlation a fast double shuttered CCD camera is typically used
to record images with double frame / double exposure. The minimum time
delay between the two laser pulses is limited by the time necessary for the
frame transfer of the camera. As the second frame can not be shuttered
the two images have different background intensity and you have to use
a suitable filter in front of the camera lens. But compared to the auto-
correlation you get a higher and unambiguous correlation peak.
References
Kompenhans J, Raffel M, Willert C (1998), Particle Image Velocimetry A
practical guide, Springer, Berlin (this book is also available in Japanese
translation).
Hinsch K (1993), Particle Image Velocimetry in Speckle Metrology (ed.
Sirohi R.S.), Marcel Dekker Inc., N.Y., 235-324.
Westerweel J (1993), Digital particle image velocimetry: theory and appli-
cation, Delft University Press.
16
4 Recording
In addition to the standard option of the Recording dialog there are spe-
cific options available that go alongside with a PIV project. This chapter
explains the additional options, for details on the standard features please
refer to the DaVis manual (#1003001).
4.1 Live mode
In Live mode no data will be stored, it is just an online display of live
camera images.
17
4 Recording
4.1.1 Camera-AOI selection
Here up to three rectangular area of interests (AOI 1 to AOI 3) are selected
for which various quantities are calculated and shown as overlay graphics
in the image. Switch on and off the AOIs of each camera by checking the
boxes. Define the position of each AOI by clicking on the button to the right
of the check box. A short notify message will appear to select opposite
corners of the rectangular AOI with the mouse in the image : first click one
corner, then move the mouse to the opposie corner and click again. Repeat
until satisfied.
4.1.2 Show information
Here different quantities can be selected that are calculated and displayed
as overlay on the image:
AOI rectangles shows the selected AOI as rectangular overlay in the
color blue (AOI 1), red (AOI 2) or green (AOI 3) with a text e.g. Cam 1
AOI 2 in the upper left corner.
18
4.1 Live mode
dynamic range shows the maximum count number inside the AOI and as
text next to it the corresponding number of bits for easy comparison
to the capability of your camera. This is useful to avoid saturation of
the image and to make use of the full sampling range of your chip.
The open rectangle shows the highest value found in the last 10 im-
ages for comparison. The filled rectangle the current value. The max-
imum is the (known) maximum count number of the camera.
focus quality shows the focus quality is in focus. Blury images will get
lower numbers. Higher numbers correspond to sharper edges and
higher dark-bright gradients in the image. The open rectangle shows
the highest value found in the last 10 images for comparison. The
filled rectangle the current value.
Whenever the value becomes greater than the formerly found max-
imum value the range readjusts so that the maximum value found
so far fills 95% of the maximum area. You can readjust the range
manually by pressing the Reset button.
19
4 Recording
particle size shows the mean diameters of the particle images where the
intensity drops down to half its value. This quantity is calculated by
the width of the auto correlation peak of the pixels inside the AOI.
Ideally for PIV this value lies in the range of 2 4 pixels.
particle density shows the average number of particles per pixel (ppp)
which ideally for PIV is a value close to 0.05 ppp. This value is de-
termined by a simple segmentation process with threshholds three
times higher than the background noise to distinguish particles from
background. Please note that in very densly seeded regions or with
regions with no particles this quantity might get erroneous results.
beam overlap shows the global normalized correlation value of all pixel
intensities from first and second frame of the camera
correlation value =
i
I
1
I
2
I
2
1

I
2
2
for very small dt the two frames should show identical particle pat-
terns for a good beam overlap whcih results in a correlation value
close to 1. If the beams are misaligned the correlation value drops
20
4.2 dt optimizer wizard
down and becomes close to 0 if there is no beam overlap and the
particle patterns are uncorrelated.
Text size Here the text size may be adjusted to be either more readable
or to make more space for further entries.
Display now Pressing this button redraws all overlay graphics which is
useful if the image in the window has been zoomed up or down so
that not all items are visible.
Scheimpflug adapter
Using a Scheimpflug adapter, it is a good idea to put one AOI in the center
of the image to focus the distance and two AOIs to either side to adjust
the Scheimpflug adapter until all three AOIs have a high focus quality. For
details regarding the use of Scheimpflug adapters ( mounting, adjustment,
etc. ), refer to product manual (1005619, 1004090).
4.2 dt optimizer wizard
In this wizard dialog the optimal pulse separation dt of the two laser beams
can be found. There are three modes: manually, interactively and auto-
matically.
This feature will be available in Davis 8.1
21
4 Recording
22
5 Vector calculation
The essential principle of PIV is to illuminate a seeded flow-field with two
short laser light pulses and record the particle images with a camera. To-
day modern CCD cameras use progressive-scan, interline transfer technol-
ogy that allows to record each of the two exposures on separate frames
followed by analysis based on cross-correlation of the two frames. We call
the technique of positioning the two laser pulses on sequential CCD frames
double-frame mode. DaVis will store the two frames of one double-
frame recording into one file.
If the flow velocity is very small and the resulting particle image shift on the
image sensor very small or the repetition rate of camera and light source is
very fast (e.g. using a modern CMOS camera and a diode-pumped Nd:YLF
laser) it is also an option to use the camera in single frame mode followed
by analysis based on cross-correlation of two consecutive images. DaVis
will store the single frames recorded in a time-series in separate files.
The PIV project offers two Groups for Vector calculation in the DaVis
Processing dialog. These are Vector calculation - double frames and
Vector calculation - time series of single frames. These two groups
reflect the two different recording options for the PIV data.
23
5.1 Vector calculation - double frames
This group needs to be used if the PIV data has been acquired using the
camera(s) in double-frame mode. In this case the cross-correlation will
work on the two frames of one recording.
There are several operations available in this group:
Operation: PIV (particle image velocimetry):
This operation will work on the frames recorded with each camera
and will use the PIV (particle image velocimetry) function for vector
calculation. For details on the parameter cards refer to chapter 8.
Operation: PIV sum of correlation:
This operation will work on the frames recorded with each camera and
will use the PIV sum of correlation function for vector calculation. For
details on the parameter cards refer to chapter 9.
Operation: PIV+PTV (particle tracking velocimetry):
24
5.1 Vector calculation - double frames
This operation will work on the frames recorded with each camera
and will use the PIV + PTV (particle tracking velocimetry) function for
vector calculation. For details on the parameter cards refer to chapter
10.
25
5.2 Vector calculation - time series of single frames
This group needs to be used if the PIV data has been acquired using the
camera(s) in single-frame mode. In this case the cross-correlation will work
on the two consecutive frames of one recorded time-series.
There are several operations available in this group:
Operation: PIV time series:
This operation will work on consecutive frames of the time-series
recorded with each camera and will use the PIV (particle image ve-
locimetry) function for vector calculation. For details on the parame-
ter cards refer to chapter 8.
Operation: PIV time series sum of correlation:
recorded with each camera and will use the PIV sum of correlation
function for vector calculation. For details on the parameter cards
refer to chapter 9.
26
5.2 Vector calculation - time series of single frames
Operation: PIV+PTV time series:
recorded with each camera and will use the PIV + PTV (particle track-
ing velocimetry) function for vector calculation. For details on the
parameter cards refer to chapter 10.
27
28
6 Image Preprocessing
Using the image preprocessing you have the possibility to manipulate the
particle image before the vector calculation is performed. Often this helps
to improve the quality of the results, especially if you have to work with
high or locally changing background intensities.
6.1 Inversion
You can do an intensity Inversion. This can be necessary when the particle
image intensity is less than the background. For the intensity inversion you
have the following options:
image=COUNT-image: the image intensity is subtracted from a con-
stant value COUNT that can be specified.
image=MAX-image: the image intensity is subtracted from the
maximum intensity MAX in the image.
image=LocalMax-image: the image intensity is subtracted from a
local maximum LocalMax that is calculated in an area of a specified
ScaleLength in pixel dimensions.
image=LMin + Factor*(LMax-LMin) - image: the image intensity
is determined according to the specified equation while the local in-
tensity minimum LMin and the local intensity maximum LMax are
determined on an area of a specified ScaleLength in pixel dimen-
sions and a Factor that may be selected arbitrarily. For a factor=1
the result will have a dynamic range of 0, ... , (LMax LMin).
29
6.2 Subtract a sliding background
You can Subtract a sliding background. This can be useful when you can
not prevent intensity fluctuation in the background due to reflections etc.
Like a high pass filter the large intensity fluctuations in the background may
be filtered out while the small intensity fluctuations of the particle signal
will pass through. So you receive an image with constant background level
without affecting the particle signal you correlate on. The scale length
in pixel dimensions should be at least double the size of the mean particle
diameter.
original particle image after sliding background filter
Figure 6.1: Example for sliding background filter.
6.3 Subtract an offset
30
6.4 Particle intensity normalization (min/max-filter)
You can Subtract an offset. The offset may be specified in counts (inten-
sity) in the corresponding textbox. This feature is useful when the normal-
ized correlation function is used for vector calculation.
6.4 Particle intensity normalization (min/max-filter)
This option applies a local particle intensity correction. This is especially
useful if work on particle images with a high intensity fluctuations, e.g. due
to inhomogeneous particle diameter.
This filter will determine the maximum and minimum intensity on a local
window defined by the scale length [pixel]. The MINMAX-filter will work in
the following way:
1. Take the image and extract sliding minimum over a window of <scale-
length> pixel as specified in dialog.
2. Subtract this sliding minimum from the original image. This procedure
will remove the local background intensity level. The result result0 is
used later in step (5).
3. Compute a strict sliding maximum over a window of <scale-length>
pixel as specified in dialog = result1.
4. Compute sliding maximumover a windowof 10<scale-length>pixel
as specified in dialog to get a global maximum = result2.
5. Divide result2/result1 to get result3. This is the ratio of local to global
maximum.
6. Multiply the intermediate result0 from step (2) by result3.
As a result you obtain homogeneous particle intensities so that small par-
ticles will also contribute to the correlation. Note that this filter will only
manipulate the particle intensity and not the particle position.
31
original particle image after normalization
Figure 6.2: Example for particle intensity normalization filter.
6.5 Test current settings
If you press the Test current settings button the image preprocessing will
be applied to the currently active source image (left window) and the result
will be shown in the result window (right window). This allows to verify if
the current parameter setup is suitable to improve the image quality for
the vector calculation.
32
7 Define Mask
Using the Define Mask card allows to add masking information to every
specific frame that is marking a pixel valid or masked out. This informa-
tion is used in the display and by many algorithms so that only valid pixels
are used for further calculations. The Statistics card in the data prop-
erties dialog card for example calculates the average/rms/min/max values
only from the visible pixels/vectors (i.e. of those that are not masked out).
Please note that the information (pixel count values in images or shift val-
ues in vector fields) of masked out positions are not deleted, but they are
simply not used. Deleting the mask can bring them back. Some batch op-
erations let you choose if you want to use all available information or just
the part that is marked as valid information (and not masked out).
7.1 Operations to mask out an image
There are several ways to do create a mask for an image:
By defining areas with a composition of geometrical forms like filled
rectangles, circles and polygons.
By using the image intensity information itself to get a local criteria
like intensity thresholds or local intensity variations (RMS).
By applying an existing mask stored at a hard disk created otherwise.
These operations exist as individual functions and are named correspond-
ingly:
Define geometric mask
Define algorithmic mask
Load mask from disk
It is also possible to combine operations under mask out image that cal-
culates the intersection of several methods to define a mask.
33
Define algorithmic
mask
Load mask from disk
Define geometric
mask
7 Define Mask
Press the Test processing button in the batch processing dialog to use the
image displayed in the Source window on the left side as source for the
calculation and display the resulting mask for the currently set parameters
on the right side in the Result window.
The following figure shows an example of the masking procedure for a
densely seeded particle flow. You see the original image the result of each
individual card and the resulting intersection of the individual masks.
In the following the individual parameter cards are explained.
34
7.1.1 Define mask
The define mask major card allows to select which method you want to
use to define the mask.
Use existing mask allows optionally to use the existing mask (se-
lect) or to start the mask definition from scratch which removes any
existing mask information from the images if any is present before
continuing (deselected).
Define geometric mask allows to select from a number of user de-
fined geometric shapes as filled rectangles, circles, polygons for a
specific frame, for all frames or for all frames of an individual cam-
era or for all camera frames in the same coordinate system (mapped
calibration).
Define algorithmic mask allows to use the local intensity informa-
tion inside the image to decide if a pixel will be masked out or not
Load mask from disk loads a single mask from disk and uses it as
look up table to add the mask information to all files or load a set
of mask files (one per image) and attachs it as mask information to
the corresponding image
If several selections are made all individual masks will be combined with
AND to calculate the intersection of all masks.
In the following sections the parameter dialogs for the individual modes will
be explained in more detail. Please note that if you do not enable a mask
definition mode in this dialog the corresponding parameter dialog for that
mode is disabled and no changes can be made.
7.1.2 Geometric mask
This is the standard mask definition mode suitable for fixed geometrical
shapes. Typically this is required in the case of limited optical access and
35
7 Define Mask
as a result the image has certain fixed areas with no useful information. An-
other typical reason is that you want to take out areas with reflections and
other artifacts that stay at a fixed position and would negatively influence
in the derived result of an algorithm (e.g. would appear as ghost particles).
You can define masked regions for a certain frame or for a certain camera
(please see the following screen shots). In the latter case a masked area is
automatically attached to all frames of the specified camera (e.g. double
frame cameras).
You define a region by clicking on one of the following 8 buttons and then
select the positions on the source image in the batch processing with the
mouse. The blue buttons enable all pixels inside the defined element and
masking out all pixels at the outside. The red buttons enable all pixels
outside the defined element and masking out all pixels inside.
The first mouse click defines a corner of the rectangle, the second
mouse click defines the opposite corner. Here it is not important which
corner you click first. You might start with the upper left, followed by
the lower right, or vice versa.
For a polygon any number of points can be added. Pressing "c"
closes the polygon and stops adding further points. "d" deletes the
last entry and "q" quits the entry and discards the polygon.
The first mouse click defines the center position of the ellipse.
Then move the mouse to a new position and then click again a second
time to define a corner of a bounding box that totally encloses the
ellipse (see figure below).
36
The first mouse click defines the center position of the circle.
Then move the mouse to a new position and then click again a second
time to define a point on the circle perimeter.
If you define more than one element then the intersection of all elements is
calculated the following way: first the single blue enabled areas are added
up (if a pixel is inside of any one of the elements then it is enabled). Inde-
pendently the single red enabled areas are added up (if a pixel is inside of
any one of elements then it is disabled). Then the intersection of those two
areas is calculated (if a pixel is enabled in the added red AND enabled in
the added blue areas then it is enabled in the final result). Please refer to
the following figure for an example:
Additionally you can select in the box to the right side of the blue / red
button a the frame dependent mask card if the element is used to mask
out one frame (that is the currently visible frame) or all frames of the
image. In the camera dependent mask card you can make a similar
selection to apply the elements of the mask definition only to one camera
that is to the frames of the currently visible camera or to all cameras that
is to all frames of cameras in the same coordinate system. In order to
use this functions a proper calibration for mapped cameras is required (e.g.
as necessary in a stereo system) and a z-position has to be given that is
used to map the masked camera raw image are to world coordinates first
and then determine which pixels need to be disabled in the other camera
raw images. This is also the reason why a rectangle / circle in the original
camera most likely looks deformed in related camera images. Usually you
would specify z = 0 mm which is at the position of the calibration plate in
the calibration process.
37
7 Define Mask
All defined elements appear in the Existing items: list of elements and
can be deleted individually if needed. You do so by selecting an item in
the list (active items are drawn in green color in the buffer to identify the
selection) and pressing Delete. You can clear the whole list by pressing the
Delete all button. Also you can make copies of single elements pressing
Duplicate and edit them manually pressing Edit. If you do the latter a
dialog appears that let you adjust the element to a certain degree (see
following screen shot):
You can change the following Mask properties: switching from enable
inside to disable inside and vice versa, specify if the mask should be used
for all cameras / all frames or specify one camera/one frame. In case
of a camera dependent mask you can also change the z-position for
which the mask mapping is done.
Furthermore you may change e.g. the Point= pixel coordinates of a rect-
angle or the Center=, Radius= pixel coordinates of a circle in the field Ed-
itable information. They correspond to the positions where you clicked
on the image with your mouse. This is useful e.g. when you missed the
desired coordinate while clicking (e.g. (0,0) for the very upper left corner of
the image) or if you duplicated an item and want to adjust the coordinates
of this second item manually to a know position (e.g. shift a circle center
by 200 pixels to the right).
38
Some things cannot be changed in the dialog - those are the displayed in
the Fixed information group.
The Create manually button is not yet available.
In the bottom row a number of options are available to help you in the mask
generation process. First you can save the current settings and restore old
selections by pressing the and buttons. A dialog opens to define a
set that should be stored to or loaded from disk. Pressing Show overlays
will draw the defined elements to the currently visible image in the Source
window of the batch processing. Selecting the include overlays in result
checkbox will copy the graphical overlay items to the result image.
7.1.3 Algorithmic mask
Sometimes you cannot define a fixed mask for a series of images because
the sample under study is moving or deformation or because the seeding of
a flow is changing density and distribution. Rather than defining a different
mask for many images the solution might be to define some criteria to
distinguish areas of interest and areas that should not be considered in
further calculations. One criteria is for example an intensity threshold.
So if you need to take into account a moving area of interest you should
adjust your experimental setup before the recording to facilitate this pro-
cess. E.g. use a wavelength filter corresponding to your laser illumination
on your camera lens or just hang a black curtain in the background when
using normal lighting to see the reflected illumination but suppress back-
ground light and hence make it easy to see the contour of your sample or
distinguish highly and poorly seeded areas.
In the algorithmic mask mode a copy of the image is taken to run some
algorithms using the local pixel intensity information to decide if the mask
for that pixel is enabled or disabled. The resulting mask is applied to the
original image without changing the intensity values in the result.
To create a mask in algorithmic mode a number of different operations
can be selected in the list on the right.
39
7 Define Mask
Eliminate all 0 pixels: If desired all zero count pixels in the original
image are replaced and will hold the given count value. This may be
required if a previously masked image contain many zero count pixels
and disturb subsequent algorithms. Here they can be given a count
value close to the average of the rest of the image.
Invert image and add an intensity offset N using the formula
I (x, y) = I
max
I (x, y) + N.
Smoothing over a neighborhood of N pixels that is over an area of
2N + 1 2N + 1 pixel centered at the current pixel position.
Sliding average/maximum/minimum filter over N N pixel.
Median filter over N N pixel. The median filter is a non-linear fil-
ter not working with the simple matrix equation above. The median
filter takes a N N area and sorts all elements according to their in-
tensity. Then the middle element is stored back as the new center
pixel. Just like the other smoothing filter, it will eliminate high fre-
quency noise. As an advantage it does not eliminate sharp edges as
much as the other filters. On the other hand this filter is non-linear,
i.e. the average intensity of the image may change. The median filter
works mathematically very similar to the erosion and dilatation filter
discussed below, which have a very different behavior.
Above/below threshold sets the intensity above or below a given
threshold to 0 counts.
Erosion: Just like the median filter this filter will sort the pixels of
a 3 3-area according to their intensity. But instead of taking the
middle element, the erosion will take the element before. The effect
is a kind of erosion, valley become deeper and finally the image will
be eroded away toward the background level.
Dilate: The dilatation filter does exactly the opposite of the erosion.
It will take the next element after the middle one in the sorting list.
The effect is that edges become sharper while plateaus flatten and
become more even.
Binarize image to 0, 10000 counts: will set all pixels to either 0 if
they were 0 before or to 10000 otherwise. This options is useful to
distinguish low-count pixels from 0. The image may also be inverted
afterwards. Note that sometimes it is easier to define the areas which
must not be used e.g. for vector calculation.
40
Local RMS: Any pixel value is replaced with the local RMS deviation
to the average of its neighbors. This is a useful filter operation to
mask out areas with high particle seeding for example. Image ar-
eas without or with poor particle seeding show a homogeneous (low)
count distribution while areas with many particles have high local
contrast (count value variations). The same applies to probes with
speckle pattern.
Replace value: to 0, 10000 counts: will set all pixels of the specified
value to 0. This options is useful to use masks with a mask out value
different from 0. Note that sometimes it is easier to define the areas
which must not be used e.g. for vector calculation.
User function: A CL macro of the form macro(int inputoutputbuffer,
int parameter) can be specified that is called with the image in the
currently active state of the operation pipeline and the given param-
eter value. The macro should return the changed buffer. This way it
is possible to write your own filters.
A typical scenario is using a intensity threshold to divide the the image in
areas with seeding (high pixel count values) and without seeding (low pixel
count values). As between single particles there might be low count pix-
els as well it is recommended first to apply a sliding maximum filter over
a certain distance N (typically the mean distance between seeding parti-
cles). This will put the values to the highest value in a N neighborhood
in x and y direction. This will effectively fill in all small low count areas
with high values if high pixel count seeding particles are in the neighbor-
hood! Areas where no high pixel count seeding is nearby maintain their
small intensity values. See the following screen shot for an example with
the operation below threshold of 400 counts alone and a combination of
Sliding maximum over 10 pixels and below threshold of 400 count.
41
7 Define Mask
If the average intensity of the seeded area differs very little from the back-
ground intensity with no seeding it is recommended to calculate the local
RMS value first which is the intensity variation which gives high values
for seeded areas and low values for non seeded areas. Typical values are
RMS=10 for highly seeded regions and RMS=0 for poorly seeded regions.
Starting from this you can apply the above Sliding maximum and below
threshold functions.
As the Sliding maximum filter tends to enhance the region of interest to
non seeded areas you can erode the region using the erosion filter as last
function. This effectively eats the outer N pixels of the mask away like
peeling several layers from an onion.
7.1.4 Load mask from file
You may have a different source to calculate the masked areas. Using
this mode you can load an image (or a set of images) holding intensity
values equal zero (disabled pixels) and pixel count values greater than zero
(enables pixels). This intensity information is used to define a mask for the
source camera images. There are four different modes available how this
can be done (see following screen shot)
Same image mask for all files: Here you specify a single image in
any supported format (IM7, IMX, BMP, TIF,...) that will be applied as
mask to all images of the source set.
42
Same image mask for all files from locally stored MASK.SET
(1st file). Look for the first file found in the locally stored set
MASK.SET below the current set and applied it as mask to all images
of the source set. This is useful when different masks for different
recordings should be applied and Hyperloop is used to do the batch
processing overnight on all the recordings.
Separate image mask for each file from specified SET Here
you specify a set of images that will be applied as mask to each cor-
responding images of the source set (first mask file on first image,
second mask file on second image...)
Separate image mask for each file fromlocally stored MASK.SET
Look for the locally stored set MASK.SET below the current set and
apply the set files as mask to each corresponding images of the
source set (first mask file on first image, second mask file on second
image...). Again this is useful if different masks for different record-
ings should be applied and Hyperloop is used to do the batch process-
ing overnight on all the recordings.
Below the selection radio group you see the currently selected file or set
name and the button to the right that opens a file selection box to change
the currently active selection.
7.1.5 Advanced mask settings
On the Advanced mask settings card you may select how to take into
account the masked out pixels for the vector field computation. These
parameter concern only those interrogation windows are partially masked
out. There are two different concepts.
No vector if vector position is outside mask: Allows to disable
a vector if the center pixel of the corresponding interrogation window
is unselected by the mask. As this is often an insufficient criterion it
43
7 Define Mask
is unselected by default. The recommended criterion is based on the
residual percentage of the corresponding interrogation window after
removing the information from the masked out pixels (see below).
Only use pixel inside mask: The remaining pixel masking will de-
termine the percentage of the residual information after the mask has
been applied. Depending on this value the software will decide if the
vector will be displayed or if it will determine the vector using direct
correlation.
Disable vector if ratio of masked out / all pixels: Allows to de-
termine the minimum ratio of masked out pixels and all pixels in the
corresponding interrogation window before disabling the vector. This
ratio may be selected separately for the first pass (default=0.3)
and for all later (multi) passes (default=0.6).
Example:
0.0 = vector is disabled if >0% of the interrogation win-
dow is masked out (even if this is only one pixel)
0.5 = vector is disabled if > 30% of the interrogation win-
dow is masked out
1.0 = no vector is enabled by masked out pixels in the
interrogation window
Do direct correlation (slow, better) if ratio of masked out / all
pixels: Allows to determine the maximum allowed ratio of masked
out pixels and all pixels in the corresponding interrogation window
before switching to direct correlation. This ratio may be selected
separately for the first pass (default=1) and for all later (multi)
passes (default=0.1).
The direct correlation will only use the residual information of the in-
terrogation window and not the masked out area. This avoids that
the border of the mask will give a significant contribution to the cor-
relation and will determine the the mathematical exact correlation for
each separation of both interrogation windows. For the same reason
this procedure is much more time consuming than the standard FFT.
Therefore this option is disabled by default for the first pass. This may
be modified depending on the image quality.
44
Example:
0.0 = always do direct correlation
0.1 = do direct correlation if > 10% of interrogation win-
dow is masked out
1.0 = never do direct correlation
45
7 Define Mask
46
8 Vector Calculation
Parameter: PIV
Using the vector calculation parameter you can select the data source and
the mode for the correlation. The correlation can be done in a single step
or iteratively via multi pass. The interrogation window size and overlap can
be determined. Furthermore you can enable the image correction and the
display of intermediate results and enable the high accuracy mode for the
final pass(es).
The standard cross-correlation PIV algorithm computes first the 2D-cross-
correlation plane from the correlation of the two input image interrogation
windows I
1
and I
2
(e.g. 32 32 pixel) from the light exposure at two dif-
ferent times shortly after another (typically with double-pulse laser). The
correlation computation is usually done using FFT which is much faster than
directly summing up the correlation terms.
Correlation plane of size n n pixel at image position x
0
, y
0
:
C(x
0
, y
0
) = I
1
(t
1
, x
0
, y
0
) I
2
(t
2
, x
0
, y
0
).
Then in the correlation plane the highest correlation peak is detected,
whose position corresponds to the most likely displacement of the particle
ensemble in the interrogation window. Vector V=(Vx,Vy) at image position
x
0
, y
0
:
V(x
0
, y
0
) = Position of highest peak in C(x
0
, y
0
).
Each cross-correlation of a single pair of interrogation windows leads to a
single displacement vector. All vectors together are then combined to yield
a complete 2D-vector field of the whole image.
Usually a large number of images are acquired and the corresponding vec-
tor fields are computed and the vector fields are locally averaged to com-
pute an average vector field.
V
avg
(x
0
, y
0
) =
V(x
0
, y
0
)/n, n = number of images / vector fields.
47
8 Vector Calculation Parameter: PIV
Note that it is also possible to compute a standard deviation from the n
vectors and other qualities derived from there (e.g. turbulent energy).
8.1 Use Mask
The Use Mask option determines if the mask that has been defined on the
Define Mask (see chapter 7) is taken into account for the vector calcula-
tion. If the Use Mask option unselected the complete image is used for
correlation, if it is selected the predefined binary mask is used. In order
to obtain the best spatial resolution you should try to arrange the experi-
mental setup in a manner that you can use the complete image for vector
calculation. Sometimes this is not possible due to experimental boundary
conditions. The mask consists information which parts of an image or a
frame must not be used for evaluation.
48
8.2 Correlation Mode
Figure 8.1: Example of a mask generated from a particle image.
8.2 Correlation Mode
For the evaluation of images via correlation the complete image is divided
in so-called interrogation windows. The correlation function operates on
the intensities inside each interrogation window and passes through the
complete PIV recording with a specified window shift. The evaluation will
yield one velocity vector for each interrogation window.
Figure 8.2: Dividing the image in in-
terrogation windows.
image
interrogation
window
Using the radio buttons the corresponding evaluation mode may be se-
lected. Note that the choice of the correlation mode depends on the cam-
era mode that has been used for the data acquisition.
cross-correlation: calculates a vector field on two single-exposed
images.
49
sequential cross-correlation: calculates a multi frame vector field
applying cross-correlating to frame pairs 0+1, 2+3, 3+4, ...
stereo cross-correlation: calculates a 2D-3C vector field with three
velocity components V
x
, V
y
, V
z
applying cross-correlation to images
of two cameras.
image (double frame)
cross
correlation
ds
(0,0)
1st frame
2nd frame
Figure 8.3: Cross-correlation.
In the cross-correlation mode the original image consists of two frames
(1
st
and 2
nd
frame). The first frame (= frame 0) contains the 1
st
exposure
while the second frame (= frame 1) contains the 2
nd
exposure. The algo-
rithm computes the cross-correlation of all interrogation windows between
frame 0 and frame 1. The highest peak in the cross-correlation image is
the most likely displacement vector ds.
8.3 Data source
If the Advanced settings flag is off the correlation is done on frame 0
and frame 1 of camera 1 for a planar PIV system and on camera 1+2 for
a stereo PIV system by default. If the flag is set you may select the data
50
8.4 Iterations
source for the correlation arbitrarily from frame 0, ... , frame 15 or camera
1, ... , camera 8.
8.4 Iterations
Using the Multi-pass radio buttons you can determine whether the PIV
evaluation is done in a Single pass with constant window size or in a
Multi pass with constant or decreasingly smaller window sizes. For each
option the number of iterations Nx may be selected.
Single pass: The vector calculation is done in one pass. The interro-
gation window size and the overlap can be selected in active selection
boxes in the second row and is constant for the evaluation. There is
no postprocessing applied during vector calculation.
Multi pass (constant size): The vector field is calculated by an ar-
bitrary number of iterations Nx on the same image with a constant
interrogation window size. In each pass a reference vector for each
interrogation window is processed. The window size is determined by
the active selection boxes in the second row. The computed vector
field information is used as reference vector field for the next pass.
Using the information of the previous pass the position of the interro-
gation windows in the new pass is shifted according the determined
particle image shift. This helps to correlate the right particles and
improves the signal-to-noise.
Multi pass (decreasing size): The vector field is calculated by an
arbitrary number of iterations Nx with a decreasing interrogation win-
dow size. The initial interrogation window size can be determined in
the selection boxes in the first row and a final interrogation window
size in the selection boxes in the second row. The evaluation starts in
the first pass with the initial interrogation window size and calculates
51
interrogation
window
(1st frame)
interrogation
window
(2nd frame)
1st pass 2nd pass
3 matching particles,
poor correlation
5 matching particles,
good correlation
= +
result of 1st pass
window
shift
cross-
correlation
cross-
correlation
estimator
high precision
result
Figure 8.4: Adaptive multi-pass with constant interrogation window size.
a reference vector field. In the next pass the window size is half the
size of the previous pass and the vector calculated in the first pass
is used as a best-choice window shift. In this manner the window
shift is adaptively improved to compute the vectors in the following
steps more accurately and more reliably. This ensures the same parti-
cles are correlated with each other even if you use small interrogation
windows where less particles enter into or disappear from the interro-
gation window. The options for postprocessing during multi-pass may
be selected using the Multi-pass Postprocessing file card.
The multi-pass with decreasing window size allows to use a much
smaller final interrogation window size than it would be possible with-
out adaptive window shifting. This improves the spatial resolution of
the vector field and produces less erroneous vectors.
For example if a final interrogation window size of 32 32 pixels is
used then with a fixed window shift of 0 pixel it is only possible to
calculate vectors with 16 pixels particle displacement maximum (in
practice only 8 pixel). In a vector field with large fluctuations vectors
can still be computed in areas with big particle separation using the
multi-pass algorithm since the window shift is locally adapted to the
mean particle displacement. Using an initial 64 64 pixel window
size even large vectors are calculated with high confidence since the
gradient ds is still small (< d
P
) for the corresponding window size.
52
8.5 Window Size and Weight
Those vectors are then used as a window shift for the final 32 32
calculation.
The Window size defines the size of the interrogation window in width and
height. In case of a decreasing interrogation window size the first row spec-
ifies the initial and the second row specifies the final interrogation window
size.
For the vector calculation the corresponding image is divided into so called
interrogation windows. The rectangular size can be selected from 4, 6, 8,
12, 16, 32, 64, 128, 256, 512 or 1024 pixels (the default is a square). Each
correlation operates only on the intensities inside the corresponding inter-
rogation window.
During a multi-pass the initial interrogation window size is divided by two
in each step until the final interrogation window size is reached.
Example: With the interrogation window size set to 64 64 pixels, an
image of e.g. 1280 1024 pixels is divided into 20 16 interrogation
windows (assuming an overlap of 0%). With the interrogation window size
set to 32 32 pixels, a data range inside the image of e.g. 1280 1024
pixels would be divided into 40 32 windows (again assuming an overlay
of 0%).
53
An image evaluated
with a fixed interrogation
window size of 16x16pixel,
0% overlap and initial shift
of Vy=-5
(no Postprocessing).
The same image
evaluated with a fixed
window size of 8x8pixel
and an initial shiftof Vy=-5
(no Postprocessing).
Result of an evaluation using
multi-pass from 32x32pixel
down to final window size
of 8x8pixel, 0% overlap
(no initial shift and
no Postprocessing).
Figure 8.5: Example for an evaluation using single or multi-pass. The sepa-
ration of the signal in pixel is in the range of 4 < Vx < 4 and 3 < Vy < 12.
54
Additionally you have the possibility to select a weighting for the interro-
gation window in the Select interrogation window weighting function
dialog. This dialog can be opened by a click on the weight button in the
Vector calculation parameter dialog. The weighting is disabled by de-
fault (= square) as this is the fastest mode. Depending on the quality of
the raw data and the dominant flow direction it might be useful to use a
Gaussian weighting function.
Internally an interrogation window of twice the width and height is used for
the calculation and the weights are adjusted in such a way that the integral
of all weighted pixels is equivalent to the interrogation window specified.
E.g. if 64 64 is selected, then internally the calculation is done with
128 128 pixels centered on the current position and the Gaussian bell
adjusted in such a way that the many additional outer pixel add with very
small weights so that the sum is comparable to 64 64 whole pixels.
In general a round weighting factor (1:1) will improve the accuracy in the
vector field computation as the resulting vector weights the local informa-
tion higher than the pixel information further away and such represents the
local shift better. A drawback is the higher calculation time.
The elliptical weighting factors (1:2) and (1:4) is a good choice if high gra-
dients exist in a given direction depending on a dominant flow. E.g. in a
laminar flow through a pipe there is a gradient perpendicular to the pipe
wall. So it is a good idea to use more pixels in the flow direction and use
less pixels in the direction perpendicular to the wall that would smear the
correlation peak due to contributions with different velocities.
55
Note: If the image size is not a multiple of the interrogation window size
the remaining parts on the right and bottom of the image may not be taken
into account for the evaluation or the missing space to fill the specified in-
terrogation window size is padded up with zero intensity. The criteria for
the both options is whether the center of the corresponding interrogation
window is still inside or outside the image size. If the center of the inter-
rogation window is outside the image size the area is ignored, when the
center is inside the image the interrogation window is padded with zero
intensity.
8.5.1 Adaptive PIV
Choosing this option the interrogation window size and shape will be
changed automatically and optimized locally according to the local seed-
ing density and flow gradients. Smaller interrogation windows are desired
since a smaller window leads to a higher spatial resolution of the resulting
vector field. On the other hand a certain size is required to guarantee a
minimum number of particles inside the window for the robustness of the
correlation.
Choosing this option is recommended to achieve the highest possible accu-
racy, robustness and spatial resolution. The only reason to deactivate this
option is a lower computation time. The adaptive PIV calculation method is
applied during the last multi-pass iteration only. Previous passes are calcu-
lated with round Gaussian weighted interrogation windows. It requires at
least two passes and an overlap of 50% or larger in the last pass. Although
the window sizes will be adaptively changed the grid spacing of the vectors
will be kept by adapting the local overlay to equalize the varying window
sizes.
Working principle of the shape adaption
The method is increasing the size of the window in the direction of larger
flow structures indicated by a smaller velocity gradient, keeping the same
overall size of the window by enlarging the window in the perpendicular
direction. A variable shape is implemented by weighting all pixels of both
interrogation windows with a round or elongated Gaussian elliptical func-
tion. The orientation of the ellipse is such that the major axis of the ellipse
is minimizing the variation of the displacements dV taken from the previous
iteration over all angles from 0
to 180
:
56
dV
=
180
=0
|v
0
v
|,
where v
0
is the displacement at the center of the interrogation window,
summing over all displacements v
along the line through the center of the

interrogation window at a certain angle . The length of this line is two
times the windows size. The velocity values along this line are determined
by a bilinear interpolation of the 33 surrounding velocity vectors.
The eccentricity of the ellipse, i.e. the aspect ratio of major to minor axis,
is computed from dV
and the value perpendicular to this direction dV

+90
:
= 4/(1 + 3(dV
/dV
+90
)).
The eccentricity is thereby limited to values between 1 and 4 and is always
applied symmetrically, i.e. the major axis is maximum two times the origi-
nal interrogation length and the minor axis half of it, respectively.
Working principle of the size adaption
The size of the interrogation window should be large enough to ensure
enough matching particles for robustness and accuracy. The criteria used
in the adaptive PIV method is the correlation coefficient C (range -1 to +1)
calculated from the previous iteration pass. This normalized correlation co-
efficient provides a measure how well the two interrogation windows have
been matched by the previous vector calculation pass and incorporates dif-
ferent degrading effects like low seeding density, high image noise, out-of-
plan motion leading to (partly) non-matching particles, and high gradients
inside the interrogation window. A low correlation coefficient C should be
compensated by larger interrogation windows to ensure stability and accu-
racy. On the other hand high local gradients should rather be tackled by
smaller windows - if image quality i.e. robustness permits - and this is given
by the second term of the local weighting factor W based on the rms of the
displacement field inside the interrogation window:
W = 1/(C ).
The median of W over the entire vector field is calculated. Windows with
minimum W will be enlarged by a factor of 0.5, the windows with W value
close to the median value with a factor of 0.75 and windows with maximum
W will be enlarged by a factor of 1.
57
References
Wieneke B, Pfeiffer K. (2010) Adaptive PIV with variable interrogation win-
dow size and shape, 15th Int Symp on Applications of Laser Techniques to
Fluid Mechanics, Lisbon, Portugal.
8.6 Overlap
The Overlap defines the overlap among neighboring interrogation win-
dows. The bigger the specified overlap, the closer is the grid of computed
velocity vectors (the number of pixels for each interrogation window is not
affected).
interrogation
window
neighbors
(50% overlap)
Figure 8.6: Example for 50% interroga-
tion window overlap.
Example: Using a interrogation window size of 64 64 pixels and an over-
lap of 0% for an image of 12801024 pixels for FlowMaster3 it gives 2016
interrogation windows and 320 vectors. The same sizes with an overlap of
50% lead to 39 31 interrogation windows and 1209 vectors.
The interrogation window size and the window overlap determine the grid
size of a vector field. This is the spacing between two neighboring vectors
in the vector field. E.g. a window size of 32 32 pixel and an overlap of
0% will give a grid size of 32 pixel, a window size of 32 32 pixel and an
overlap of 50% will give a grid size of 16 pixel. The position of the first
vector at the top left corner of a vector field is determined by the grid size
only. The convention is that the pixel position (x/y) is determined by half
58
8.7 Options
the grid size, i.e. if the grid size is 32 pixel the top left vector is located at
the position (16/16) and if the grid size is 16 pixel the vector is located at
the position (8/8). This convention allows vector arithmetic on vector fields
that have been calculated with different interrogation window size, but with
the same grid size.
(48/16)
(16/16)
32
(48/16) (16/16)
64
32
32
Interrogation window: 32x32 pixel
Overlap: 0%
Grid: 32 pixel
16
Overlap: 50%
Grid: 16 pixel
(24/8) (8/8)
32
Overlap: 50%
Grid: 32 pixel
Figure 8.7: Vector position depending on interrogation window size and
overlap.
8.7 Options
8.7.1 Use image correction
The Use image correction flag specifies whether the image correction is
used during vector field calculation or not. The image correction can be
defined using the Calibration dialog. Find details on the Calibration in the
Imaging Tools manual (#1003012).
For planar PIV you may activate or deactivate this option depending if a
calibration and image mapping is required (e.g. if the camera viewing di-
rection is not perpendicular to the light sheet) or not. For stereoscopic PIV
measurements the image mapping is required in any case and the Use
image correction option is activated by default.
59
8.7.2 Pixelwise window deformation
If a vector field is calculated based on a given reference vector field or the
multi-pass is used for calculation the raw image is mapped according to
the velocity gradient of the signal. The raw image is mapped pixel-wise
according to the given flow field and the pixel intensity is recalculated via
bi-linear interpolation. The effect of the pixel mapping is a better signal-to-
noise ratio of the correlation function, an increased accuracy of the velocity
measurement and a suppression of the peak locking effect.
without
window deformation
with
window deformation
flow field with high shear
- peak is smeared
- low Q-factor
- peak is circular
- high Q-factor
Figure 8.8: Correlation quality depending on the use of interrogation win-
dow deformation. On flow fields with high shear the benefit is a higher and
more symmetric correlation peak.
Similar to the adaptive multi-pass technique the information on the parti-
cle image shift from the first pass is used for the image correction. With an
existing velocity field the software can calculate a imaginary deformation
vector for each pixel and determines the intensity at the new position by
bilinear interpolation. To avoid redundant calculations e.g. for a big interro-
gation window overlap the image reconstruction is executed only once for
all pixel before the correlation itself is executed. The image reconstruction
takes into account the image correction and the vector pre-shift in one step
and leads to a interpolation. The customer may decide if the interpolation
is bilinear (default) or using the Whittaker reconstruction (see 8.7.5).
60
8.7 Options
References
Fincham A, Delerce G (2000) Advanced optimization of correlation imaging
velocimetry algorithms, Experiments in Fluids [Suppl.], Vol.29, p.13-22.
Scarano F, Riethmuller ML (2000) Advances in iterative multigrid PIV image
processing, Experiments in Fluids [Suppl.], Vol.29, p. 51-60.
8.7.3 Display intermediate results
The Display intermediate results flag may be used to enable the display
of intermediate results when a vector field is calculated via multi pass. Here
it is essential to check whether the initial correlation returns a reasonable
result. This is necessary to make sure that the position of the interrogation
windows are shifted properly in the next pass to match the right particle
images.
8.7.4 Vector scale
The Vector scale flag defines the units for the vector scale. There are fol-
lowing options possible that are selected by default by the project type:
(default): = m/s (default for PIV projects systems)
displacement: [pixel]
displacement: [mm]
velocity: [m/s]
61
8.7.5 High accuracy mode for final passes
Using the High accuracy mode for final passes check box you can use
a more sophisticated reconstruction algorithm for the image correction and
reconstruction. You have following options:
Lanczos reconstruction: This reconstruction type is often used for im-
age scaling and resizing. The Lanczos function indicates which pixels
in the original image, and in what proportion, make up each pixel of
the final image. It is a windowed form of the sinc filter, a theoretically
optimal brick-wall low-pass filter. The sinc function is infinite in ex-
tent, and thus not directly usable in practice. The Lanczos filter is a
windowed form of the filter an approximation of the sinc. Using larger
windows allows one to improve the accuracy in exchange for more
computation.
The Lanczos window is the central lobe of a horizontally-stretched
sinc, sinc(x/a) for -axa. Due to its form, the Lanczos window is
also called the sinc window. Thus the Lanczos filter, on its interval, is
a product of two sinc functions. The resulting function is then used as
a convolution kernel to resample the input field. In one dimension its
formula is:
L(x) =
sinc(x) sinc(x/a) a < x < a, x = 0

1 x = 0
0 otherwise
with a a positive integer, typically 2 or 3, controlling the size of the
kernel. a corresponds to the number of lobes of sinc, as the nor-
malized sinc has zeros at integers - thus a = 1 corresponds to just
the (positive) central lobe, while a = 2 has the central lobe and the
62
8.7 Options
second lobe on each side, which are negative. Integer values of a
are used as having the end of the window fall at a zero of the win-
dowed function means that the product (the Lanczos kernel) vanishes
to second order there, yielding smoother vanishing as the kernel is
differentiable at the end of the window: it does not have a corner. In
the above, we can expand the sinc terms:
sinc(x) sinc(x/a) =
a sin(x) sin(x/a)
2
x
2
.
References
Duchon CE (1979) Lanczos Filtering in One and Two Dimensions,
Journal of Applied Meteorology 18 (8), pp. 1016-1022.
Whittaker reconstruction: The Whittaker-Shannon interpolation formula
is a method to reconstruct a continuous-time bandlimited signal from
a set of equally spaced samples. The sampling theorem states that,
under certain limiting conditions, a function x(t) can be recovered
exactly from its samples, x[n] = x(nT), by the Whittaker-Shannon
interpolation formula:
x(t) =
n=
x[n] sinc
t nT
T
where T = 1/f
s
is the sampling interval, f
s
is the sampling rate, and
sinc(x) is the normalized sinc function.
This reconstruction type avoids any smoothing of the image (if the
particle diameter are above 1 pixel) so that it should be able to restore
the original image without loss of information if you apply the inverse
transformation.
References
Whittaker ET (1915) Expansions of the Interpolation-Theory,
Proc. Roy. Soc., Edinburgh, Vol. 35, 1915, p. 181.
Bicubic spline interpolation: Bicubic interpolation is an extension of
cubic interpolation for interpolating data points on a two dimensional
regular grid. The interpolated surface is smoother than corresponding
surfaces obtained by bilinear interpolation or nearest-neighbor inter-
polation. In this case bicubic interpolation is accomplished using cu-
bic splines. In image processing, bicubic interpolation is often chosen
63
over bilinear interpolation in image resampling, when speed is not an
issue. Images resampled with bicubic interpolation are smoother and
have fewer interpolation artifacts.
Figure 8.9: Nearest-neighbor (left), bilinear (middle) and bicubic (right)
interpolation on a the same dataset. The black dots are the locations of the
prescribed data being interpolated. The color indicates the function value.
For the bilinear interpolation the derivatives of the surface are not continuous
over the square boundaries. For the bicubic interpolation the color samples
are not radially symmetric, they are more square-based.
Note: If the High accuracy mode for final passes is unselected the
software will use bilinear interpolation.
By default, Lanczos reconstruction is set as algorithm in High accuracy
mode. You can activate Whittaker reconstruction and bicubic spline inter-
polation in the Vector related expert settings menu (see 17).
8.8 Multi-pass options
On the Multi-pass options card you may activate the Initial window
shift for image reconstruction, select the used Correlation function
for first and final pass(es), select the option for the Deformed interroga-
tion windows and specify the maximum reconstruction error for the 3D
vector validation.
64
8.8.1 Initial window shift for image reconstruction
Entering a suitable initial shift can enhance the signal-to-noise ratio of the
correct correlation peak. It should be used whenever a good estimation of
the displacement vector is possible (known from experimental conditions
or from a first correlation without window shift). Using this shift the particle
image is reconstructed in a first step and the corresponding interrogation
windows are positioned symmetrically around the position of the final vec-
tor. This procedure is applied to the first pass only.
Constant: You may use a constant initial shift. If the mean veloc-
ity in x- or y-direction is known the corresponding separation can be
entered in the dx and dy text box (values can be entered in pixel or
scaled). The default value for the shift is dx = 0, dy = 0.
Use reference vector field: This mode uses a reference vector
buffer, where each vector in the reference buffer specifies the inter-
rogation window shift at that position. In this mode the interrogation
window shift is varied across the size of the imaged area but is fixed
temporally. This is a useful feature for analyzing a series of records
where each individual vector field does not deviate much from the
mean flow, e.g. in stationary flow phenomena. This mode is used
repetitively by the adaptive multi-pass processing mode.
cross
correlation
ds
(0,0)
1st frame
2nd frame
image (double frame)
Figure 8.10: Window shift for cross-correlation.
In the cross-correlation mode the interrogation windows are located in
two discrete frames. Each interrogation window is shifted by (
dx
2
/
dy
2
) so
that the location of the interrogation window in the second frame is shifted
by (dx/dy) with respect to the interrogation window in the first frame.
65
In example (1) of Fig. 8.11 a 32 32 interrogation window size with 0%
overlap is specified, consequently the grid spacing between the computed
vectors is also 32 pixels. The origin of first vector at the top left is displayed
at pixel (16/16) in the middle of its interrogation window. Since the initial
shift is (0/0) the interrogation windows can be found at the same position
in both frames. They span from (0/0) to (31/31).
Example (2) of Fig. 8.11 shows an interrogation window shift = 0: Assuming
a 3232 interrogation window with 0% overlap and a initial shift of V
x
= 10,
V
y
= 0 the position of the top left vector is at (16/16) again. The window
shift is symmetrically divided between the first and the second interroga-
tion window. So the first window spans from (-5/0) to (26/31), while the
second window spans from (5/0) to (36/31). The region outside the image
is set to zero intensity, so it does not give a contribution to the correlation.
Placing the window shift symmetrically gives the best performance in re-
lation to the error introduced to the fact of non-zero dt between the laser
pulses. The vector computed this way represents an average vector at a
time something like dt/2 in the middle between the laser pulses.
In example (3) of Fig. 8.11 we find case (2) with an overlap of 50%. Here
the final vector grid spacing is 16 pixels, so the first vector at the top left
is displayed at (8/8). With an interrogation window size of 32 pixels, the
interrogation window spans from (-8/-8) to (23/23), centered around (8/8) if
the initial shift is still (0/0). The region outside the image is set to zero in-
tensity. When the middle of the interrogation window is outside the image,
no vector is computed. As long as there is enough data at the rim of the
image, the vector is computed, and the missing data is filled with zeros.
8.8.2 Correlation function
Using the Correlation function file card you may select the Correlation
function itself, you can enable the Second-order correlation, the High accu-
racy mode and the image reconstruction for the final result. Furthermore
you have the possibility to save the correlation map.
The correlation function selection box offers two algorithms, a standard and
a normalized function.
Standard I1*I2 (via FFT, no zero-padding): This correlation uses
following equation, which is the standard cyclic FFT-based algorithm
66
(16/16) (48/16) (16/16) (48/16)
(16/16) (48/16)
1st frame 2nd frame displayed vector field
(48/16) (16/16) (21/16) (11/16)
(5/0) (-5/0)
1st frame 2nd frame
displayed vector field
1st frame 2nd frame
displayed vector field
(8/8)
(-8/8)
(8/8)
(-8/8)
(24/8) (8/8)
32
16
32
(0/0)
0
0
0
(1)
(2)
(3)
0
0
Figure 8.11: Three examples (1)-(3) for interrogation window shift and over-
lap. The position (x/y) of the center and edges of the interrogation window is
given in pixel.
that calculates a cyclic correlation of the interrogation window and is
similar but not exactly the same as the mathematical true correlation:
C(dx, dy) =
x<n,y<n
x=0,y=0
I
1
(x, y)I
2
(x + dx, y + dy),
n
2
< dx, dy <
n
2
I
1
and I
2
are the image intensities of the 1
st
and 2
nd
interrogation win-
dow and the 2D-array C gives the correlation strength for all integer
displacements (dx, dy) between the two interrogation windows. n is
the size of the interrogation window and usually also the size of the
correlation plane, i.e. n/2 is the maximum displacement computed.
Cyclic means that the correlation is computed as if the two inter-
rogation windows of size n n are repeated again in 2D-space and
on this pattern the true correlation is computed. Mathematically from
the two nn interrogation windows the complex 2D-FFT is calculated,
they are multiplied complex conjugated, and the inverse FFT is com-
puted, which yields the cyclic correlation function. This is much faster
(e.g. 50) than directly computing the correlation according to the
above equation except for very small interrogation windows.
A disadvantage is that this introduces a weighting of the correlation
coefficients with an emphasis on the displacement (0,0) (assuming
67
no additional window shift right now for simplicity) with decreasing
weighting toward larger displacement. For example, for an interroga-
tion window of size 3232 pixel the weighting is 1 for a displacement
of (0,0) and
1
2
for a displacement of (16,16) pixel and 0 for a displace-
ment of 32 pixel. This is because e.g. for a displacement of 16 pixel,
only signal in left half of the first interrogation window really add to
the correlation strength. The other right half of the first interrogation
window is actually correlated with the wrong left half of the other in-
terrogation window, only adding unwanted false random correlation
noise. Therefore this correlation function should only be used, when
the displacement is less than about 1/3 of the interrogation window
size. Or the other way around, the interrogation window size should
be big enough to accommodate possible large vectors.
This weighting is not only a disadvantage. If one knows the approxi-
mate vector length, e.g. 20 pixel 3 pixel, one can set the window
shift to (20,0), in which case vectors around this window shift are pre-
ferred in comparison to other false random correlation peaks. Using
the multi-pass approach with iteratively smaller interrogation window
sizes and adapted window shifts, this cyclic FFT-correlation performs
very well, because the adapted window shift places the correlation
peak of the true displacement always in the middle of the correlation
plane and therefore gives the peak a large weighting factor.
One more disadvantage of the cyclic FFT-based correlation is that it
introduces a bias toward smaller displacements. E.g. if the displace-
ment is 10 pixel, calculated is actually 9.8 pixel. Again this is due to
the above weighting feature which changes a perfect Gaussian cor-
relation peak to one which is slightly shifted toward higher weighting
functions, that is toward lower displacements. Using adaptive multi-
pass this bias can be eliminated almost totally, because here the rel-
evant displacement is reduced to a range of less than 1 pixel.
What is left is still a so called peak locking bias effect (see references
for details), which shifts the measured displacement slightly toward
the next integer. The peak locking effect depends on the size of the
signal relative to the pixel dimensions on the CCD, the size of the
interrogation window and the density of the signal. The effect is of
68
the order of 0.03 to 0.1 pixel. The peak locking bias together with the
always present r.m.s. noise limits the possible maximum accuracy of
the vector computation.
69
Normalized (I1-I1avg)*(I2-I2avg)/rms: This correlation function
uses the following equation:
C(dx, dy) =
x<n,y<n
x=0,y=0
I
1
(x, y) I
1avg
(x, y))(I
2
(x + dx, y + dy) I
2avg
(x + dx, y + dy)
RMS(I
1
(x, y) I
1avg
(x, y))RMS(I
2
(x + dx, y + dy) I
2avg
(x + dx, y + dy))
For the normalization the average of the individual interrogation win-
dows is used as reference. This means that even matching dark areas
contribute to the correlation. Note that the standard deviation of all
pixels in the first interrogation window RMS(I
1
I
1avg
) is constant (it
is actually not computed), while RMS(I
2
I
2avg
) is a function of dx
and dy and normalizes the correlation relative to the standard devia-
tion of the second interrogation window. This effectively compensates
varying background intensities and signal densities, which otherwise
would bias the correlation function toward higher intensities.
The equation above is calculated again via FFT, but this time using
a twice as large FFT-window and filling up appropriately the first in-
terrogation window with zeros (zero-padding, see papers below for
details). The computation time is more than doubled due to double
the FFT-size.
This correlation function does not have a weighting or a bias, but still
shows a small peak locking effect.
In practice one would test different correlation functions to see which one
supplies the best results, i.e. as few false vectors as possible. The stan-
dard FFT-correlation is about factor 5 faster, while the normalized correla-
tion function works better with a varying high background intensity. Very
good results have been achieved with the adaptive multi-pass approach
and the fast standard FFT-correlation with three iterations together with
an image preprocessing consisting of a high-pass filter to remove varying
backgrounds. One may also try the second-order correlation additionally,
which often is more reliable, but on the other hand introduces a slightly
larger error.
References
Fincham AM, Spedding GR (1997), Low Cost, High Resolution DPIV for Mea-
surements of Turbulent Fluid Flow, Experiments in Fluids, Vol. 23, pp. 449-
462.
Ronneberger O, Raffel M, Kompenhans J (1998), Advanced Evaluation Algo-
rithms for Standard and Dual Plane Particle Image Velocimetry, Proc. 9
th
International Symposium on Applications of Laser Techniques to Fluid Me-
chanics, Lisbon, July, p. 10.1.
70
8.8.3 Deformed interrogation windows
Allows to specify if the vector is positioned symmetrically in the middle
between interrogation windows / image (usually the case in PIV) or at the
position of the first interrogation window / image (usually in time series
experiments like deformation processes that are calculated relative to the
first image).
8.8.4 3D vector validation
The 3D Vector Validation option can be used as additional validation
criterion for Stereo systems. This is the triangulation error when trying to
calculate a 3-component vector from two 2-component vectors plus the
image calibration.
In Stereo PIV system a second camera is used. Each camera returns a vec-
tor field and on each recording a correlation is executed. So you obtain two
correlation peaks V1=(Vx1,Vy1) and V2=(Vx2,Vy2). Taking into considera-
tion the stereo calibration the four velocity components (V
x1
, V
y1
, V
x2
, V
y2
)
are used to calculate the stereoscopic (2D-3C) vector field with three ve-
locity components (V
x
, V
y
, V
z
). The linear equation system to compute the
z-shift is over-defined:
Vx1 = Vxdxx1 + Vydxy1 + Vzdxz1,
Vy1 = Vxdyx1 + Vydyy1 + Vzdyz1,
Vx2 = Vxdxx2 + Vydxy2 + Vzdxz2,
Vy2 = Vxdyx2 + Vydyy2 + Vzdyz2.
dxx1, dxy1, ... are the gradients of the image calibration function in the
given camera at the position of the 2-component vector. The solution is
determined using the Normal equation. The error is the error of the so-
lution. In the triangulation process the two lines of sight do not intersect
but pass one next to the other, i.e. the error is the distance between lines
of sight. This error may be used as an indicator for the quality of the 2C
vector fields and their conformity. The error textbox can be used to enter
a maximum deviation between the two vector fields from both cameras.
Since the vector field of each camera image contains the 4 most probable
solutions (positions of the 4 highest correlation peaks) for the 2C-vectors
this 3D vector-field validation is not simply rejecting 3C vectors. This vali-
dation selects the 4 most probable 3C vectors out of the 16 combinations
of the four 2C-vectors obtained from each camera image.
71
8.9 Multi-pass postprocessing
Using the Multi-pass Postprocessing file card you may select the param-
eters for the postprocessing during the multi-pass.
Using the multi-pass it is advisable to run a validation on each calculated
vector field before using the corresponding vector field as a reference in the
next pass. When spurious vectors are not removed it is possible that the
correlation with a reduced window size returns worse results as the right
particles are not matched any more.
The Median Filter and the Fill-up empty spaces (interpolation), Fill-
up all and Smoothing options are used by default. By entering the cor-
responding values you may select how restrict the filter is. For details see
Vector Postprocessing chapter 12.
Additionally the peak ratio Q and Remove groups can be applied for
filtering the intermediate calculated vector fields.
72
Parameter: PIV Sum Of
Correlation
The PIV Sum of correlation or average correlation method is extremely
useful if one is only interested in the average vector field and the standard
correlation method does not work good enough due to poor seeding or high
noise. For these cases the average correlation method can deliver average
vector fields with high stability and accuracy. A priori the average correla-
tion method is only useful for datasets with stationary flow conditions such
as phase-locked recordings in periodic flows.
All settings in the vector calculation parameter dialogs may be used iden-
tical to the standard PIV vector calculation.
Please note that for a multi pass approach the batch processing works re-
peatedly through the whole set. So due to the higher data transfer with
73
9 Vector Calculation Parameter: PIV Sum Of Correlation
the hard disk the sum of correlation method might take longer than the
standard multi pass approach plus averaging of the vector fields.
9.1 Description of method
Usually for a steady flow you would take N images to calculate N vector
fields and average them to suppress spurious vectors and time dependent
structures. This way you get the steady component of your flow.
The disadvantage of that procedure is that if a single vector field is too un-
reliable and noisy the normal averaging does not work. Because then too
many false vectors lead to a bad average result and only statistical averag-
ing over spurious vectors will not lead to the desired result or the desired
precision for the result. These problems usually arise with poor seeding,
low light levels or a high degree of turbulent (random) components in your
flow overlaying steady flow component.
The sum-of-correlation algorithm computes from each acquired image only
the correlation planes using the cross-correlation of the individual interro-
gation windows.
Correlation plane of size n n pixel of image i at position x
0
, y
0
:
C
i
(x
0
, y
0
) = I
1
(t
1
, x
0
, y
0
) I
2
(t
2
, x
0
, y
0
).
Then all correlation planes are summed up to lead to a single ensemble of
correlation planes.
C
avg
(x
0
, y
0
) =
C
i
(x
0
, y
0
) summing over all images.
From here a single vector field is computed from all the summed-up corre-
lation planes. Averaged vector V
avg
= (Vx,Vy) at image position x
0
, y
0
:
V
avg
(x
0
, y
0
) = Position of highest peak in C
avg
(x
0
, y
0
).
References
This algorithm is described in more detail in the following paper:
Meinhart CD, Wereley ST, Santiago JG (1999), A PIV Algorithm For Estimat-
ing Time-Averaged Velocity Fields, Proceedings of Optical Methods and Im-
age Processing in Fluid Flow, 3rd ASME/JSME Fluids Engineering Conference,
San Francisco.
9.2 When to use this method?
The average correlation method should be used if:
74
Standard PIV algorithm does not work good enough due to following
reasons:
not enough seeding particles
high image noise
particles are barely visible and/or very large (high particle inten-
sity noise)
high out-of-plane vector component (unmatched particles)
The result is that the vector field computed from a single image shows
false vectors even after a good vector post processing scheme, which
will lead to a large error, when such vector fields are averaged in
order to compute an average vector field. A single false vector can
change a computed average vector by a large amount making the
results very questionable. For large number of images it is also not
possible to manually go through all vector fields to throw out false
vectors.
Standard Particle Tracking schemes (PTV) also do not work. For a low
seeding density the PTV algorithm it is often more advantageous to
use compared to the correlation based PIV algorithm. But PTV might
also not work with enough stability maybe due to the same reasons
as given above. Then the average correlation method should be used.
One is only interested in the average vector. When there is no need
for a calculation of standard deviation of vectors (or derived quanti-
ties), the average correlation method directly gives the average vec-
tor field with a high stability and accuracy.
The average correlation method should not be used
If the flow field is highly unstable. For a very turbulent or unstable
vector field, the summation of the correlation planes might not lead
to a single correlation peak. For example if the flow changes between
two preferred states (bi-modal), then two correlation peaks might ap-
pear.
If your flow is so steady and well seeded that calculating a single
vector field is no problem. Then averaging over N vector fields will
lower the statistical error of the averaged vector field by N compared
to the noise of a single vector field. In the sum of correlation method
the averaged vector field only has the error of a single vector field
75
(statistical error from correlation peak fit).
Example: for 100 images the average vector field of a series of vector
field may achieve a precision of 0.01/sqrt(100) = 0.001 pixel while
the sum of correlation method is limited to a precision of 0.01 pixel
accuracy of each vector component.
Note that the average correlation method can still be used for moderately
turbulent flows as long as enough images are summed up to yield one
correlation peak.
9.3 Example
The acquired images in this example are taken on an injection nozzle for an
IC engine. The diameter of the nozzle is approx. 1 mm. As the velocities are
quite high the pulse separation for the recording has been set to dt=0.45s.
The raw images are of very poor quality. Seeding is very low, particles have
a size of up to 20 pixels in diameter and the particle images are quite noisy,
which means that particle tracking would have a lot of problems to deal
with such data. The image itself is also quite noisy and the illumination is
uneven, which becomes visible when averaging e.g. 100 images. So this is
a classical case, when standard PIV has a lot of difficulty.
Also the flow in the bowl in the left image region is quite slow ( 5 pixel
displacement, 10 m/s) while in the small channel to the right (< 200 m)
the flow becomes very fast (200 m/s) with particle displacements above
60 pixels. This alone poses a quite severe problem to any evaluation of
the images. One could reduce the time between the two exposures, but
this reduces the accuracy in the left part of the image considerably. Also
the time spacing used was 450 ns, which was at the limit of the camera in
conjunction with an image intensifier, because the phosphor of the image
intensifier needs this decay time. The other option to use is the adaptive
multi-pass scheme to start with e.g. 128 128 pixel interrogation window
size to adaptively compute a suitable interrogation window shift for the
76
9.3 Example
later passes with e.g. 3232 pixel. This often works in such cases, but not
here, because the seeding density is too low and the vector gradients are
very high especially when the flow enters the channel. The intermediate
vector fields are simply not stable enough.
When evaluating the image using a fixed interrogation window size of
32 32 pixels and no interrogation window shift, the result is very poor.
The vectors in the channel are of course totally random since the displace-
ment there is 60 pixels. In the bowl there are a lot of false vectors but also
some good ones.
The average correlation method gives a significant improvement. The dif-
ference of the correlation map for the correlation of a single image and the
sum of correlation is obvious. The correlation map of a single correlation on
the left shows correlation peaks only for a few interrogation windows. The
correlation map of the average correlation method shows significant corre-
lation peaks in most of the interrogation windows. Note that often there is
only a single broad correlation peak. It is so wide because of the big size of
the particles and - maybe - due to the variations of the velocity field itself.
But this can not be checked, because only the average vector field is com-
puted. No further information about the standard deviation of the vectors
is available.
Find below the corresponding vector field in the channel calculated by the
average correlation method. Note how smooth the vector field is. No post-
processing was done, i.e. no false vectors.
77
Two regions of the flow are shown with 20 m/s = 5 pixel shift and 200 m/s
= 50 pixel shift respectively.
78
Parameter: PIV + PTV
This option will use the PTV (Particle Tracking Velocimetry) algorithm for the
vector calculation in a final step in order to get maximum resolution for the
velocity field. In addition this procedure will make sure that the position
of the vectors is exactly at the position of the detected particles. But the
PIV+PTV option is doing the PIV vector calculation in a first step and the
resulting vector field is used as an estimator for the PTV algorithm. All
parameter dialogs concerning the PIV functionality may be used identical
to the standard PIV vector calculation.
10.1 Description of the method
The PIV + PTV method implemented in DaVis works as a sort of in-
tegrated PIV+PTV method. The evaluation starts with the standard PIV
evaluation to determine a velocity field. This can be used to get a robust
estimator of the local velocity field. Then the resolution of the vector field
if further refined by the detection and tracking of individual particles (PTV).
In order to detect the particles you have to specify a particle image diam-
eter range and an intensity threshold. This will allow to separate single
particles from the background noise. The next step is the pairing of par-
ticles between the frames. The obtained displacement field from the PIV
evaluation will allow to interpolate a vector centered on the predicted par-
ticle position. This is used to narrow down the number of possible partners
and reduces the number of false measurements. On the basis of the deter-
mination of successive particle positions the displacement and velocity is
calculated. The particle detection method adapted in DaVis will simply
look for the position of the highest intensity and will arrange an interro-
gation window of the size of the particle diameter around the position of
the maximum intensity. The interrogation window in the second frame is
shifted by the interpolated vector from the PIV result. Finally an individual
particle correlation will correlate the intensity pattern of both windows to
79
10 Vector Calculation Parameter: PIV + PTV
determine the exact shift.
Particle image (1st frame) PTV result PIV result
Figure 10.1: A PIV and PTV result on the same particle image.
The obtained vector field may be further processed. For example you can
convert the vector field to a fixed grid by interpolation in order to calculate
other derivatives like rotation or divergence.
References
This algorithm is described in more detail in the following paper:
Scarano F, Riethmuller M (2000), Advances in iterative multigrid PIV image
processing, Exp. in Fluids, 29-S1, 51-60.
Theunissen R, Stitou A, Riethmuller M (2004), A novel approach to improve
the accuracy of PTV methods, 12th International Symposium, Lisbon.
Stanislas M , Okamoto K, Kahler C and Westerweel J (2005), Main results of
the Second International PIV Challenge, Exp. in Fluids, 39, 2, 170-191.
10.2 PTV Calculation Parameter
On the PTV Calculation Parameter card you can activate the Do PTV as
final step option if you want to continue the PIV vector calculation by a fi-
nal pass doing a PTV vector calculation. In this case the software will do an
80
10.3 PTV Postprocessing
image reconstruction based on the resulting vector field of the PIV calcula-
tion, i.e. the original particle image will be deformed pixelwise so that each
particle in both frames should overlap more or less perfectly. Then another
correlation is executed using the grid=1. You need to specify the corre-
sponding Correlation window size that should be around 88 pixel and
an Intensity threshold. For all detected particles above the threshold the
software will do a correlation based on the interrogation window arranged
around the pixel with the maximum intensity. You may specify the Allowed
vector range relative to reference as a limit for the allowable deviation
from the shift that has been predicted by the PIV evaluation. The result is
a sparse vector field with grid of 1 pixel.
10.3 PTV Postprocessing
The result of the PTV vector calculation can be postprocessed with following
filter on the PTV Postprocessing card.
Do PTV postprocessing: This option allows to switch on and off the
complete PTV postprocessing.
Median filter: The median filter is working similar to the median
filter in the vector postprocessing (see paragraph 12.3). In contrast
to the PIV vector postprocessing you can not expect the vectors on a
fixed grid >1 pixel. In order to find the minimum of eight neighbors
the filter will enlarge the neighborhood area of the concerned vector
in steps of 33, 55, 77, ... until it detected at least the required
eight vectors. The limit of the neighborhood extension is 20 pixel. If
this limit is reached the filter will work on the N < 8 vectors that have
been detected in this area.
Denoising filter: This filter will correct the velocity values for the
vectors obtained by the PTV vector calculation using the velocity in-
81
formation in the neighborhood. It will search all valid vector posi-
tions and check for vectors in the neighborhood up to a separation
of 80 pixel depending on the specified Number of vectors used
for smoothing that needs to be 3. Based on the velocity values
of these vectors the filter will interpolate the velocity field. You can
select one of the following fitting functions:
simple averaging / strong filter, fast
polynomial fit (1st order)
polynomial fit (2nd) / preserve gradients, slow
The velocity information of the original vector will be replaced by the
interpolated velocity value of the neighboring vectors. Using the Ig-
nore up to N outliers box you can specify a number of neighbors
that will not be taken into account for the average calculation. The fil-
ter will determine all fitting functions based on the difference between
detected vectors and outliers. Finally the function with the least error
is selected. This procedure allows to identify spurious vectors and to
ignore them for the interpolation.
The result of this filter is a smoothed velocity field where spurious
vectors have been replaced by the interpolation of the neighboring
vectors while these spurious vectors do not contribute to the interpo-
lation themselves.
The drawback of this filter is that it is very time consuming as it needs
to calculate N fitting functions for N detected neighbors and for two
outliers there are N(N+1) combinations for ignored vectors that need
to be investigated.
Convert to grid: This filter is working similar to the denoising filter
but in addition to the interpolation it will fill up the vector field based
on a given regular grid. For the search area extension you need to
specify the Number of neighbor vector used (3) for the interpo-
lation and the New grid size. We recommend to use a final grid of
8 pixel, for a finer grid the calculation is to time-consuming. You can
select the option Disable output, if no vector within a distance
of +-N pixel to make sure that all interpolated vectors will be dis-
abled, if the distance between new grid position and nearest original
vector (i.e. particle) is above a certain limit.
82
If the final vector field is based on a regular grid you may apply any func-
tion in the BatchProcessing that works on vector fields to calculate further
derivatives (basic vector arithmetic, vector mapping, extract scalar field).
without postprocessing after denoising filter after convert to grid
Figure 10.2: A PTV vector result after applying different postprocessing fil-
ter. Generally the denoising and convert-to-grid operations will modify every
vector. The changes introduced by the denoising filter are most clearly visible
on questionable vectors e.g. at the gray mark in the vector field.
The PTV method should be used if the seeding density is quite low or if
a higher spatial resolution of the vector field is required. The PTV vector
calculation will allow to display the vector at the position where the particle
has been detected in the recorded image. In particle images with high
seeding density this method may suffer from the statistical ambiguity in
the particle image.
83
84
Parameter: PIV on GPU
PIV on GPU (CUDA) is an optional operation in the vector calculation
- double frame function group. The operation is only accessible with an
additional software license. It calculates the PIV cross-correlation on the
Nvidia graphics processing unit (GPU) instead of the computers central
processing unit (CPU). The data processing using this operation can be up
to 15 times faster compared to the standard PIV cross-correlation. Nev-
ertheless, both methods give the same vector field results with the same
precision and robustness in almost all cases.
This chapter will cover the differences to the standard PIV calculations. For
further information on the installation of the Nvidia cards please consult the
manual "Graphics Processing Unit" (manual no. 1006396).
Most parameters and dialogs are identical to the standard PIV. In the vector
calculation parameter card shown in the dialog below, two additional
parameters need to be explained here. Those parameters define the size
in pixels of the resulting correlation function in which the correlation peak
will be searched.
85
11 Vector Calculation Parameter: PIV on GPU
1. max. shift must be set at least 1 pixel larger than the maximum
occurring particle image separation.
2. The shift for all other passes must be larger than the particle image
separation difference across the initial interrogation window size if
you select the multi-pass decreasing size option. The maximum
accepted value for both parameters is 30 pixels!
If the particle image separation is larger than these shift-settings, wrong ve-
locity vectors will be obtained without any warning! On the other hand the
calculation speed is strongly influenced by these parameters. The larger
those are the slower the calculation will be.
LaVisions Cuda-PIV code is determining only the highest correlation peak.
The standard CPU-PIV code used in DaVis determines the 4 highest
peaks. As a consequence three options are not available:
1. The 3C-vector-validation in the multi-pass option card for Stereo PIV
calculation. With only one highest peak per interrogation window in-
stead of 4, no choice can be made to determine the best combination
of the 4 x 4 positions. Accepting or removing the 3C vector is possi-
ble in the post processing functions using the same criteria. See also
chapter 8.8.4 and 12.2.
2. The median filters used in the post processing and multi-pass valida-
tion have less options: only removing is possible not replacing. See
also chapter 12.3.
3. Data validation according the peak ratio is not available.
As a consequence of determining only the highest peak the behavior of the
3C-vector validation and the median filters are different to the standard PIV.
86
This may result in differences to the standard PIV especially in multi-pass
and Stereo-PIV calculations of noisy images.
Cuda-PIV is does not work using a Windows remote session! If you need to
control the computer remotely and while using the CUDA-PIV feature you
have to use a third party remote control program which transfers the dis-
play without using the local graphics board as the Windows remote desktop
does.
87
11 Vector Calculation Parameter: PIV on GPU
88
12 Vector Postprocessing
Once a vector field has been calculated, vector validation algorithms can be
applied to eliminate spurious or false vectors. The parameter for this vali-
dation can be specified in the Vector Postprocessing file card. The oper-
ations are processed from top to bottom when the corresponding checkbox
is active.
For noise reduction and validation in vector fields there are various possible
methods.
12.1 Allowable vector range
The Allowed vector range will restrict the filtered vectors to a user spec-
ified range in units of particle separation [pixel] or velocity [m/s]. A range
may be specified for each velocity component V
x
, V
y
and V
z
(value
range). Any vectors outside this range will be removed.
Note: The information obtained from the Scatter plot function may be
useful to select the range for each velocity component (see corresponding
paragraph). The range V
z
dV
z
is only taken into account for 3C-vector
fields.
89
12.2 Peak Ratio or Correlation Value
The peak ratio Q is defined as Q =
P1min
P2min
> 1 , where min is the lowest
value of the correlation plane and P1 and P2 are the peak heights of the
first and second highest correlation peak. Thus the peak ratio calculated
here is a relative ratio, relative to the common correlation background to
compensate e.g. a general image background level or background noise.
correlation
level
shift ( x) D
Peak 1
Peak 4
Peak 2
Peak 3
P1
P2
min
Figure 12.1: Definition of
peak ratio or Q-factor.
If there is no second highest peak, the Q-factor is set to 100. This may hap-
pen, if the allowed vector range is small (a few pixels only) so that only one
peak is inside the selected range, or if there is a strong laser light reflection
together with CCD-chip overloading, so that there is only one strong peak
again.
The correlation value contains the correlation coefficients, which can the-
oretically range in value from -1.0 to 1.0, practically it will typically range
in value from 0.0 to 1.0. Negative values only show up if one of the inten-
sity distributions is the inverse of the other. Correlation values close to 0
indicate that there is no preferred shift direction detection, correlation val-
ues close to 1 indicate that the two intensity distributions are identical and
without noise.
The Q-factor or the correlation value can be used as a postprocessing crite-
ria for eliminating questionable vectors below a ratio-threshold. In general
one would assume the higher the Q-factor or correlation value is the higher
90
12.3 Median Filter
is the confidence in this vector. It should be noted that this procedure is
quite unspecific and will also remove many good vectors. One would use
this postprocessing procedure, for example, when it is necessary to have
as few bad vectors as possible, e.g. when calculating flow averages and
r.m.s. In this case it is better to remove all questionable vectors, because
bad vectors will spoil the average. Accidental removal of good vectors only
means that additional images are required to compute an accurate aver-
age. It should also be noted that bad vectors may also have a high peak
ratio or correlation value.
The Q-factors and correlation values are automatically stored as an addi-
tional information for each vector in the vector field in the corresponding
VC7-file. These values are by definition always > 1. For good PIV data the
Q-factor is in the range of 3-4 and the correlation value approx. 0.7-0.8.
Peak ratios close to 1 mean that there is a high probability that the highest
found peak (= best vector) is a false random peak. The peak ratio and
correlation value give a good (visual) indication for the quality of the vector
field.
Caution: You can obtain a Q-factor< 1 after postprocessing if vectors
corresponding to the highest correlation peak have been removed and re-
placed by vectors corresponding to the 2
nd
, 3
rd
or 4
th
highest correlation
peak (see Local Medial Filter paragraph below).
12.3 Median Filter
The median filter computes a median vector from a group of neighboring
vectors and compares the middle vector with this median vector devi-
ation of the neighboring vectors. The center vector is rejected when it is
outside the allowed range of the average vector deviation of the neighbor
vectors.
91
Figure 12.2: Local
median filter.
Criteria for keeping a vector:
U
median
U
rms
<= U <= U
median
+ U
rms
AND
V
median
V
rms
<= V <= V
median
+ V
rms
with
U
median
= median value of all U components of the neighbor vectors,
V
median
= median value of all V components of the neighbor vectors,
U
rms
= deviation of U components of the neighbor vectors,
V
rms
= deviation of V components of the neighbor vectors.
If the vector field has a 3D-component w, the criteria for w is also checked.
Note that each vector component u, v and w is checked independently.
The median value means that e.g. all u-components are sorted according to
their value, and the middle component is taken. E.g. from 5 vectors sorted
the third vector is taken. When e.g. 6 vectors are sorted, the average of
the third and fourth vector is taken. This procedure has the advantage that
big outliers at either end of the sorted list will not change the median value,
as would be the case when computing a simple average of all components.
The deviation is not computed on all components, rather the two most devi-
ating values are ignored (only done if more than 5 neighbor vectors exist),
such that up to 1 or 2 spurious vectors do not spoil the rms calculated.
The median filter can be applied once or several times after each other (up
to 5).
References
Westerweel J (1993), Digital Particle Image Velocimetry Theory and Appli-
cation, Delft University Press.
Westerweel J (1994), Efficient Detection of Spurious Vectors in Particle Im-
age Velocimetry data Sets", Exp. Fluids, Vol. 16, pp. 236-247
92
12.3 Median Filter
DaVis implements three versions of the median filter.
1. remove & replace
The first one is the basic version remove & replace as described in
the above references. LaVision improves on this algorithm in two
ways. First the algorithm checks whether any of the 8 neighboring
vectors are themselves spurious vectors and does not take them into
consideration for the average and deviation computation. Secondly if
a vector has been identified as false, the other 3 highest correlation
peaks (also stored in the vector buffer) are also checked if they fit in.
If so, one of them is taken instead. This way a very reliable method
for identifying spurious vectors is available even when many spurious
vectors are present (the replaced vectors can be identified by a spe-
cial color depending on the height of the corresponding correlation
peak, see Vector choice in Vector Display chapter.
Note: We recommend to enter factors between 1 and 3, depending
on the turbulence of the vector field that has to be processed.
2. strongly remove & iteratively replace
The second version strongly remove & iteratively replace is a
4-pass regional median filter, described in detail below, which is ca-
pable of rejecting groups of spurious vectors. This is not possible with
the first basic version, since groups of spurious vectors have a high
standard deviation of the velocity vector and therefore often do not
reject spurious vectors surrounded themself by a few false vectors.
This filter is ideally suited whenever it is required that the vector field
should not contain any spurious vectors, even with the drawback that
some good vectors are rejected, too. This is especially important for
the following cases:
Any flow field with many spurious vectors:
Often it is not possible to achieve ideal experimental condition,
where only very few spurious vectors are present. Recording
hundreds or thousands of images and vector fields it is not pos-
sible to individually remove false vectors by manual inspection of
93
each image. Here an automatic robust filter capable of removing
most of the false vectors is required.
Calculating averages, standard deviation or other ensemble quan-
tities:
A single false vector, which is not rejected, will substantially
change the calculated average or other ensemble quantities, be-
cause a false vector usually deviates by a large amount. For
example, if the vector field average at a certain position is V
= 5 pixels with a standard deviation of 2 pixel, then a single
false vector of V = 10 pixel among 100 good vectors will change
the average to 5.05 or a 5% error, and the standard deviation is
2.05, also a 5% error. Derived derivatives like vorticity have even
larger errors. It is better to remove all false vectors if possible,
even when some good vectors are removed also.
Multi-pass windowing with adaptively adjusted window-shift:
When the interrogation window becomes smaller and smaller,
eventually more and more spurious vectors are generated. In or-
der to go to very small interrogation windows, it is essential that
in any intermediate pass the false vectors will be detected, be-
cause any false vector will lead most likely to many false vectors
in the later passes, since the interrogation window will be placed
at the wrong position. A well-designed restrictive postprocess-
ing filter can help to decrease the size of the smallest allowed
interrogation window, thus increasing the resolution of the final
vector field.
3. universal outlier detection
The universal outlier detection is an adaption of the original median
test for the detection of spurious PIV data. This normalizes the median
residual with respect to a robust estimate of the local variation of the
velocity.
Consider a displacement vector denoted by U
0
, its 3 3 neighbor-
hood data, denoted by {U
1
, U
2
, ... , U
8
} and U
m
as the median of
{U
1
, U
2
, ... , U
8
}. A residual r
i
defined as r
i
= |U
i
U
m
| is determined
94
12.3 Median Filter
for each vector {U
i
| i = 1, ... , 8} and the median r
m
of {r
1
, r
2
, ... , r
8
}
is used to normalize th residual of U
0
:
r
0
=
|U
0
U
m
|
r
m
.
This method is quite general in outlier detection and can be used to
process a large variety of inhomogeneous data. In fact, for purely
uniform flow, the normalization factor r
m
tends to zero. This can be
compensated by assuming a minimum normalization level , i.e.
r
0
=
|U
0
U
m
|
r
m
+
,
where may represents the acceptable fluctuation level due to cross
correlation. It was found that a suitable value for is about 0.1 pixel,
which would correspond to the typical rms noise level of the PIV data.
References
Westerweel J, Scarano F (2005) Universal outlier detection for PIV data,
Exp. Fluids, Vol.39, No. 5, pp. 1096-1100.
Vector candidates can be removed if their residual is above the re-
move is residual < threshold and alternate vector candidates can
be inserted if their residual is below the (re)insert if residual >
threshold. The number of neighbors used for the calculation of the
residual can be changed in discrete steps using the filter region se-
lection box from 3 3 up to 11 11. The filter is applied only to
vectors, for which the number of neighbors is above the minimum
number min.# of Vs required. The value (see above) is set to a
constant value of 0.1 pixel.
With the difference that the criteria for removal and insertion are different
the used algorithm is similar to:
Nogueira J, Lecuona A and Rodrguez PA (1997) Data validation,
false vectors correction and derived magnitudes calculation on
PIV data, Measurement Science and Techology, Vol 8, no. 12,
pp. 1493-1501
The median filter is used here because it has some very convenient scaling
properties:
It is independent of the frame of reference. Spurious vectors inside
a moving vortex are removed the same way as inside a stationary
vortex.
95
It is independent of the magnitude of the vectors. The threshold for
removal is the standard deviation times a fixed user defined factor
of usually somewhere between 1 and 3. Larger vectors e.g. due to
larger dt between laser pulses have a correspondingly larger standard
deviation.
It is tolerant (at least to some degree) of surrounding false vectors.
It adapts to changing magnitude of the velocity gradient. In regions
of higher velocity gradients the standard deviation is also higher.
The first pass eliminates all vectors, which do not satisfy the median
criteria as described above. The second pass eliminates all vectors
which do not have a certain number n (3-8) of neighboring vectors
left. Here all vectors without sufficient neighbors are marked first be-
fore they are actually removed at the end of this pass.
These two passes together very restrictively remove questionable
vectors, of course also depending on the threshold factor. Left over
are only those vectors, which satisfy the median criteria and are sur-
rounded by a sufficient number of good neighboring vectors.
The third pass then tries to fill in as many good vectors as possi-
ble. The criteria for filling in is the same median filter criteria as for
removal. Good vectors are filled in again (even if they have been
removed before), if they are inside the allowed range of the median
average standard deviation of all existing neighbor vectors. Again
a certain number of neighbors are required (usually set to 3) to be
able to fill in any new vector. In addition not only the vector corre-
sponding to the highest correlation peak is checked, but also up to
three lower correlation peaks are checked. Since all existing vectors
are assumed to be good ones, their associated local standard devia-
tion is low compared to the very beginning, where there are still some
false vectors in the neighborhood. Therefore the chance of filling in
a spurious vector is decreased. This third pass is repeated until no
more new vectors can be filled in.
The fourth pass finally throws out any groups of vectors with less than
a certain number of vectors. This effectively removes small groups
(e.g. 3 or 4) of spurious vectors which have not been detected in
pass 1 or 2. Note that groups of spurious vectors usually do not grow
much during pass 3.
96
12.3 Median Filter
Example:
Let us assume a local vector field mainly in the x-direction with one vector
(V
x
= 9 pixel, V
y
= 4 pixel) which sticks out diagonally:
Figure 12.3: Vector field with the
corresponding particle image shift in
pixel.
(11/2) (13/0) (13/0)
(12/0) (9/4)
(9/1)
(9/-1) (9/0)
(7/0)
Is the sample vector (9,4) a bad vector assuming a specified factor of 1.4?
The velocity components of all 8 neighboring vectors are arranged accord-
ing to their magnitudes (each component separately). The median of the
sequence is the arithmetic average of the two middle elements.
7, 9, 9, 9. 11, 11, 12, 13 U
median
= 10
-1, 0, 0, 0, 0, 1, 2, 2 V
median
= 0
Now we calculate the r.m.s of neighboring vectors. As we have a number
of neighbors N5 the two highest outliers are removed (7, 13 for U and 2,
2 for V). E.g. the r.m.s for the U component is calculated by
U
rms
=
1
N 1
N
i =1
(U
i
U
median
)
2
We get U
rms
= 1.3 and V
rms
= 0.6. The algorithm sets the r.m.s. at least to
1 pixel. This gives U
rms
= max(1.3, 1) = 1.3 and V
rms
= max(0.6, 1) = 1.
The r.m.s for the neighboring vectors multiplied with the factor 1.4 leads to
the allowed deviation range of each velocity component: dV
x
= 1.31.4 = 1.8,
dV
y
= 1 1.4 = 1.4.
If the center vector deviates more than the factor 1.4 of the r.m.s. from the
median (in any component) it is regarded as a bad vector and removed from
the vector field. As the V
y
-component is outside the range in y-direction the
vector (9,4) is removed as a bad vector though the V
x
-component is inside
the range in the x-direction: (9,4) (10 1.8, 0 1.4).
97
12.4 Remove groups
Especially when a big overlap in the evaluation of particle images is used
it is possible to obtain groups of wrong vectors if you have an error (for
example a reflex) at one position in the image. If this error has a big con-
tribution to the correlation signal you obtain a group of vectors which are
obviously wrong but they have similar vectors in the neighborhood. These
vectors are not removed by the local median filter. The option strongly
remove & iteratively replace in the local median filter has to rely on
groups of vectors with a high confidence. To increase the confidence of a
group of vectors you can choose a number N so that groups of vectors with
less then N vectors are removed.
12.5 Fill-up empty spaces (interpolation)
Once the bad vectors are removed the left over empty spaces can be filled
up with interpolated vectors. In the interpolation an average of all the non
zero neighbor vectors is calculated and inserted. This can be executed
with the Fill-up empty spaces function. This function needs at least two
existing neighbors to calculate an interpolation.
Additionally the Fill-up all option can be used to fill up all empty spaces in
the vector field. This option uses the Fill-up empty spaces function iter-
98
12.6 Smoothing
atively. In a first step those empty spaces are filled up where at least two
neighbors are available. In a second step these interpolated vectors are
used as neighbors for empty spaces where no neighbors have been avail-
able in the first step. This procedure will be repeated until no interpolated
vector is added to the vector field.
Caution: Applying the Fill-up functions we recommend always to com-
pare the original vector field and the postprocessed vector field to make
sure that enough good vectors have been calculated and left over by the
postprocessing for the interpolation. By comparing these vector fields you
can also see easily if the interpolated vectors fit to the flow structure.
Note: We strongly recommend not to use the Fill-up functions for manip-
ulation of the vector field itself. But for the calculation of derivatives like
vorticity, divergence, etc. these functions can be useful as these calcula-
tions are very sensitive to empty spaces in the vector field.
12.6 Smoothing
The final vector field can be smoothed by a simple 33 smoothing filter to
reduce noise. It is possible to apply Smooth 3 3 on a vector buffer with
empty spaces. In this case the average of all non-zero neighbor vectors is
calculated for each empty space. Now this average is taken to the former
empty space when applying to the smoothing filter.
Note: We strongly recommend not to use the Smooth 3 3 function for
manipulation of the vector field itself. But for the calculation of derivatives
like vorticity, divergence, etc. this function can be useful as these calcula-
tions are very sensitive to non homogenous spaces in the vector field.
99
12.7 Make mask permanent
If the Make mask permanent option is selected the mask will be applied
permanently to the vector field result. The mask deletes all vectors in the
outside areas.
12.8 Apply allowable vector range again
If the Apply allowable vector range again flag is set the Allowable
vector range filter is applied again as the last function of the vector post-
processing. This can be necessary if several filter functions are combined
and functions that are applied after the Allowable vector range are in-
serting or interpolating vectors outside the range that has been confined
above.
100
12.9 Troubleshooting
12.9 Troubleshooting
Vector field does not look reasonably
Problem 1: Seeding is not adjusted properly or beam overlap of laser1
and laser2 is not adjusted properly.
Remedy: Take a double frame PIV recording with very short dt, e.g. 1s
and calculate a vector field with basic PIV Settings. You may load prede-
fined operation lists using the open button.
Interrogation window:
Iterations: Single pass
Window size and weight: 64 64 pixel, square 1:1
Overlap window: 0%
Initial Reference Vector Field:
Constant: dx=0, dy=0
Correlation function:
Correlation function: standard I1*I2, no zero padding
Assuming that the particle images could not move within this short pe-
riod of time (depends on area of interest and flow velocity) the calculated
particle shift (in pixel, use SHIFT button of keyboard) should be below the
accuracy of the system (< 0.1pixel). If this is the fact the seeding density
and the adjustment of the light sheet is correct.
Problem 2: Pulse delay dt is not adjusted properly.
Remedy: Check whether the mean particle separation ds (in pixel, use
SHIFT button of keyboard) in the vector field is in the range: 0.1 pixel
< ds <
1
4
Interrogation window size. If necessary adjust the dt.
Problem 3: Vector calculation parameter are not selected properly.
Remedy: Check the selected Correlation mode in the PIV Parameter
and make sure that it corresponds to the raw data. Compare the used PIV
parameter with the recommended PIV parameter. You may load prede-
fined operation lists using the open button.
Interrogation window:
Iterations: Multi pass (decreasing size)
101
Initial interrogation window size: 64 64 pixel, 2 iteration
Final Interrogation window size: 32 32 pixel, 3 iterations
Overlap window: 50%
Initial Reference Vector Field:
Constant: dx=0, dy=0
Correlation function:
Correlation function: normalized (I1-I1avg)*(I2-I2avg)/rms
Multi-Pass:
Relative vector range restriction = reference cell_size/3
Absolute vector range restriction = reference 2 pixel
Overlap: 50%
Problem 4: The gradient of particle separation in one interrogation win-
dow is too high.
Remedy: Change factor of magnification. Look at a smaller area of inter-
est.
No vectors are displayed
Problem 1: Settings for the vector display are not selected properly.
Remedy: Check whether the display mode in the Image Display At-
tributes dialog is set to vector arrows and adjust the Vector Length
Factor if necessary.
Strong Peak Locking
Problem 1: The particle image size does not fit.
Remedy: Check the mean particle image size (recommended 2-3 pixel).
Use particles with suitable diameter or increase particle image size by de-
focusing the camera lens slightly.
102
13 Stereoscopic PIV
13.1 Image correction for planar PIV
For planar PIV usually the camera should be set up that it is looking at 90
to the object plane. Even for this ideal arrangement sometimes it is not
possible to avoid image distortion and a correction is required. A slight dis-
tortion at the rim of the recorded image is always present. If the camera
angle is 90
in the middle of the image it may be only 89
or 85
at the rim
of the image depending on the distance to the object plane and the mag-
nification factor. For other optical systems like endoscopes the distortion is
much higher (fish-eye-view).
cushion distortion no distortion oblique viewing barrel distortion
Figure 13.1: Typical image distortions.
For example, if the field-of-view is 10 cm at a distance of 50 cm the dif-
ference of the viewing angle at the border of the image relative to 90
is
arctan(5/50) = 5.7
. This means that e.g. a velocity displacement in x-

direction of 10 pixel would actually be measured as 9.96 pixel. Since such
a (locally constant) error of 0.4% is rather small, most people simply ignore
it. But e.g. if the field of view is 20 cm at a distance of 20 cm, then the
error increases to above 10% (angle 30
)! So one should at least be aware

of this effect and check the numbers for a given experimental setup.
If the experimental setup requires a viewing angle significantly different
from 90
the correction procedure as described in detail in the DaVis

Imaging Tools manual (#1003012) becomes mandatory to be able to cal-
culate correct results. The goal is to recover the true x- and y-velocity com-
ponents inside the illuminated plane from the distorted image data. One
must be aware that such a correction is only 100% correct if the flow is truly
two-dimensional. In such a case the image correction can fully recover the
103
13 Stereoscopic PIV
a
V cot(a)
z
V
z
meassured:
V cos(a)
z
light sheet
camera
90 viewing
oblique
viewing
Figure 13.2: Systematic error introduced by viewing direction.
correct velocity components in the 2D-plane. For example, lets consider a
viewing angle of 60
(i.e. a 30
deviation from the standard perpendicular

view):
A x-velocity of 10 pixel is measured as 8.66 pixel. Provided that a (lo-
cally varying) viewing angle of 60
has been determined by a proper

calibration procedure as described below, the measured projected velocity
component can be corrected by dividing by sin():
V
x
(true) =
V
x
(measured)
sin()
.
However, if the flow has a z-velocity component perpendicular to the light
sheet, this component will influence the measured x-velocity component as
well. Or, in other words, one is not only measuring the projection of the true
x-velocity component but also a projection of a possible z-component. And
there is no way of separating those two components later on (except when
doing stereo-PIV with two cameras)! If the z-component is equivalent to
5 pixels, then it will be measured as an additional x-component of 2.5 pix-
els, putting a large error on the later reconstructed, true x-component.
Therefore such an experimental setup with oblique viewing angles requires
fully two-dimensional flows. Otherwise one has to live with the fact, that
the computed x-velocities include a projection of the z-component. The
error in V
x
due to a z-component, e.g. for a viewing angle of 60
, is
cos(60
)
V
z
sin(60
)
= cot(60
)V
z
= 0.58V
z
.
Note that the actual procedure in DaVis does not involve any specific x-
or y-angles. Instead a general purpose calibration of the image distortion is
performed, which does not distinguish between x- and y-axes and therefore
works for any camera viewing position tilted in x- and/or y-direction.
104
13.2 Stereoscopic viewing
13.2 Stereoscopic viewing
One big disadvantage of the classical two dimensional PIV method is that
it is only capable of recording the projection of a velocity vector into the
plane of the light sheet (illustrated in figure 13.2). In this case the out of
plane velocity can not be measured as you determine two velocity com-
ponents in a two dimensional domain (2D-2C). The unknown third velocity
component can also affect the measurement of the in plane components
with an unrecoverable error. Because of this the light sheet is arranged
according to the main flow direction, but for highly three dimensional flows
this can cause significant errors.
d
light sheet
d
opt
d
opt 2 d
opt 1
camera 1 camera 2
a
Figure 13.3: Inherent error using 2D PIV.
One approach to recover the complete set of velocity components is an ad-
ditional PIV recording from a different viewing axis using a second camera,
what is generally called stereoscopic PIV (shown in fig. 13.3). Using the two
different viewing angles one can obtain projections of the velocity vector
in two planes. It is necessary to determine the viewing directions of both
cameras in respect to the orientation light sheet position by a calibration.
Since the arrangement of both cameras is known you can extract all three
velocity components form the two projections. Using stereoscopic PIV all
three velocity components can be measured on a two dimensional domain
(2D-3C).
13.3 Camera Arrangement
Depending on the optical access to the area of interest there are different
possibilities to arrange the cameras in a stereoscopic PIV system:
Backward Forward-scattering: In this case both cameras are
mounted on the same side of the light sheet. The calibration can
105
13 Stereoscopic PIV
be done with a two-level calibration plate. As one camera is record-
ing the light scattered in forward direction while the second camera
is recording the scattered light in backward direction the intensities
detected with both cameras with the same f-number of the camera
lens is different (typically a factor 10 between backward- and forward-
direction), so usually the camera viewing in backward scattering di-
rection needs a larger aperture.
lens and Scheimpflug Nd:YAG laser
camera1
a
camera2
z=0
calibration plate position
Forward Forward- or Backward Backward-scattering: In this
case both cameras are mounted on different sides of the light sheet.
For this setup a two-level calibration plate with marks on both sides
is required for the calibration.
Nd:YAG laser Nd:YAG laser
camera1
camera2
z=0
camera1
camera2
z=0
In general the BF-scattering setup is easier to calibrate. The cameras
should be arranged symmetrical as this will increase the accuracy in the
vector calculation. If possible the light sheet should be arrange in a man-
ner that the maximum flow component is in-plane. The angle between the
viewing directions of both cameras is typically about 45
.
References
Westerweel J, Nieuwstadt FTM (1991), Performance tests on 3-dimensional
velocity measurements with a two-camera digital particle-image velocime-
ter , Laser Anemometry Adv. and Appl., Vol. 1, ASME, New York, pp. 349-
355.
Prasad AK, Adrian RJ (1993), Stereoscopic particle image applied to liquid
flows , Exp. Fluids, 15, pp- 49-60.
106
13.4 Scheimpflug criterion
13.4 Scheimpflug criterion
i
m
a
g
e
p
l
a
n
e
object plane
l
e
n
s
p
l
a
n
e
camera 1 camera 2
Figure 13.4: Scheimpflug criterion.
The measurement precision of the out of plane component increases with
the angle between the viewing directions of the two cameras. The mea-
surement precision for in-plane and out-of-plane component is similar if
the angle between viewing directions is 90
. One problem of an oblique

viewing direction is the limited depth of field which can be accommodated
by additionally tilting the image plane in respect to the orientation of the
camera lens and light sheet according to Fig. 13.4.
According to the Scheimpflug criterion the image plane, the lens plane and
the object plane for each cameras has to intersect in a common line. Un-
fortunately this arrangement has the side effect of introducing a strong
perspective distortion. The factor of magnification is no longer constant
across the complete field of view and needs additional calibration.
Figure 13.5: Scheimpflug
adapter.
Scheimpflug
adapter
camera
camera lens
The standard LaVision Scheimpflug adapter can be arranged between
camera mount and camera lens. The separation of camera principal plane
and CCD chip corresponds exactly to the backfocal length of a standard
Nikon camera lens. Using the allen screws the Scheimpflug adapter can be
aligned in arbitrary orientation depending on the camera viewing direction.
107
13 Stereoscopic PIV
The adjustment of the Scheimpflug adapter can be done empirically. Ar-
range a calibration plate in the light sheet position and select a large aper-
ture of the camera lens (i.e. small #f-number). The large aperture will re-
duce the depth of field and makes the alignment more sensitive. Focus the
camera lens in the middle of the area of interest. If the Scheimpflug adapter
is not properly aligned only a narrow vertical band will be in focus. Tilting
the Scheimpflug adapter to the direction according to the Scheimpflug cri-
terion will result in a wider focused range. Proceed in tilting the adapter
when the focused area is increasing. Eventually refocusing and adjustment
of the camera viewing direction becomes necessary. Stop the procedure as
soon as the complete area of interest is focused and select a suitable aper-
ture for image acquisition. A reduction of the aperture will even increase
the focused area as the depth of field increases.
Figure 13.6: Adjustment of
the Scheimpflug adapter.
References
Prasad AK, Jensen K (1995) Scheimpflug stereocamera for particle image
velocimetry to liquid flows, Appl. Optics, 34, pp . 7092-7099.
13.5 Calibration for stereoscopic PIV
The precondition for stereoscopic PIV is an accurate a calibration of the
two cameras. Correcting for off axis viewing is an essential part of stereo-
scopic PIV evaluation and must be done with care to ensure accurate re-
108
13.5 Calibration for stereoscopic PIV
sults. LaVisionemploys a fully empirical calibration method, which does
not involve measuring angles or distances. It has been proved to be highly
accurate and reliable.
For the calibration of stereoscopic systems we recommend to use La-
Vision two-level calibration plates with defined separation of the cali-
bration planes and mark spacing. Avoid a movement of the plate position
within the calibration procedure, as this is often a source of errors. Two-
level plates are available in different dimensions suitable for different field
of views. The plate Type is printed on the bottom right of and may be
selected in the software. This will preselect all relevant parameter.
Figure 13.7: Two-level
calibration plate.
In general you have to differentiate between two stereoscopic setups:
(1) FB-setup (Forward-Backward-Scattering): The cameras are on the
same side of the calibration plane.
light sheet
camera1
two-level calibration plate
camera2
camera1 camera2
Type #
calibration plate
(2) FF- or BB-setup (Forward-Forward-Scattering or Backward-Backward-
Scattering): The cameras are on opposite sides of the calibration plane.
camera1
two-level calibration plate
camera2
camera1
camera2
Type #
light sheet
calibration plate
109
13 Stereoscopic PIV
As vectors only can be calculated in the overlapping area of the im-
ages of both cameras adjust the cameras in order to obtain a maxi-
mum overlap. Make sure the complete images are focused. To avoid
defocused areas due to the oblique viewing direction use small aper-
tures or Scheimpflug-adapters.
Arrange the two-level calibration plate in a manner that the front level
that is closer to camera1 is in the light sheet plane (compare setups
in paragraph 13.3). If you use a flat calibration plate arrange it in the
calibration plane first. This will define the plane z=0. Move it with the
required distance to a second plane.
Make sure the calibration plate is adjusted plan-parallel to the light
sheet. Rotational errors may be corrected using the Self-Calibration
function.
Make sure that the viewing angle to the calibration plane between the
two cameras is higher than 10
. An angle separation of 30
to 45
is
recommended.
The calibration for a stereoscopic setup can be done by the Calibration
dialog. For details on the calibration procedure please refer to the DaVis
Imaging Tools manual (#1003012).
13.6 Vector Field Computation
Once this calibration procedure is done, you can proceed to the stereo-
scopic PIV measurement. Image recording is done using the standard
recording dialog. You need to activate both cameras in the Devices dialog.
The four frames acquired for each recording are stored in a single file. The
PIV-parameter for vector field computation work in exactly the same way
as in conventional 2D-PIV except that the correlation mode needs to be set
to stereo cross-correlation.
Note: In the simple case of the 2D-PIV with image correction there is also
the possibility of correcting the images first and then doing the 2D-PIV eval-
uation. This is in principle also possible for the stereo-PIV case, but it is not
supported in DaVis, because the method described below is more accu-
rate since a previous correction of the images would introduce an additional
error of up to 0.5 pixel.
110
13.6.1 Camera Mode
To change the recording from 2D PIV to stereo PIV open the Devices in the
Recording dialog. Choose the double-frame camera mode.
In the stereo cross-correlation mode the double images of both cameras
are stored in one file in following sequence:
frame0 = camera1 (1
st
exposure),
frame1 = camera1 (2
nd
exposure),
frame2 = camera2 (1
st
exposure),
frame3 = camera2 (2
nd
exposure).
Figure 13.8: Order of frames in
image buffer using stereo cross-
correlation mode.
2
1
Camera 1
Camera 2
1
2
13.6.2 Vector Calculation Parameter
To activate stereo vector calculation select the stereo cross-correlation
correlation mode on the Vector calculation parameter card.
If the stereo cross-correlation mode is selected the image correction is
switched on by default. This is indicated by the enabled Use image cor-
rection option on the Vector calculation parameter card.
111
13 Stereoscopic PIV
In the Multi-pass options card you may specify a maximum value error
for the velocity difference of the vector fields received from camera 1 and
camera 2 in the 3D-vector validation text box. This is the maximum
allowable difference of the calculated particle image shift obtained by cam-
era1 and camera2.
13.6.3 Procedure for Stereo Vector Calculation
Different approaches for stereo vector computation have been proposed
as summarized e.g. by Prasad (2000) or Callaud and David (2004). In all
cases a 2D2C-vector fields are computed for each camera from which by
stereoscopic reconstruction a 2D3C-vector field is computed. One has the
choice of following procedures:
1. Computing the 2D2C-field on a regular grid in the raw images and
using the two interpolated vectors to compute a 3C-vector at regular
world grid positions.
2. Computing the 2D2C-vectors in the raw image at a position corre-
sponding to the correct world position.
3. Dewarping the images first and computing the 2D2C-vectors at the
correct world grid position.
Method 1 has the disadvantage that the vectors are not computed at the
correct world position and due to vector interpolation false or inaccurate
112
vectors affect four final 3C-vectors. Method 2 has the disadvantage that the
size and shape of the interrogation windows differ between the two cam-
eras due to the perspective viewing. For method 3 the computed 2D2C-
vectors are already computed at the world correct position and originate
from the same interrogation window of equal size and shape, but a sub-
pixel interpolation is required during the dewarping, which together with
the sub-pixel interpolation necessary for the multi-pass window deforma-
tion scheme leads to added image degradation. Therefore DaVis is using
a modified method 3 approach where the dewarping and image deforma-
tion is done at once before each step in the multi-pass iterative scheme.
For the first vector computational pass the two frames (t
0
,t
0
+ dt) of each
camera are dewarped and evaluated. This already provides vectors at
the correct position in the world coordinate system. Also the size and
shape of the interrogation windows for both cameras are the same, which
means that the correlation is done on the same particles apart from effects
due to the non-zero thickness of the light sheet. Then a preliminary 3C-
reconstruction is done to remove corresponding vectors in the 2C-vector
fields for which the reconstruction error is too large, e.g. larger than 0.5
or 1 pixel. This very effectively removes spurious vectors since two false
vectors with random directions are rarely correlated. At the end of the
first pass missing vectors are interpolated and the vector field is smoothed
slightly for numerical stability.
The resulting vector field is used as a reference for deforming the inter-
rogation windows in the next pass. Actually not each interrogation win-
dow is deformed individually, but the complete image is deformed at once
with half the displacement in backward direction assigned to the first im-
age at t
0
and the other half in forward direction to the second image at
t
0
+ dt. Image deformation is mathematically the same as window defor-
mation but requires less floating point operations, since e.g. for an overlap
of 75% the same region would be deformed 16 times using window de-
formation. The image deformation is combined with the dewarping of the
original image in one step. Usually after 3 or 4 passes at the final inter-
rogation window size the 2D2C-vector fields have converged sufficiently.
Then the 3C-reconstruction is done which consists of solving a system of
4 linear equations with three unknowns (u,v,w). This is done by using the
normal equation which distributes the error evenly over all three compo-
nents. Computing from (u,v,w) again the (u1,v1)- and (u2,v2)-components
the deviation from the measured (u1,v1) and (u2,v2) can be calculated (re-
construction error). Usually with a good calibration and 2C-vector errors of
113
13 Stereoscopic PIV
14
2D2C
Optional:
Initial Reference Vector Field 2D3C
Mapping Function
of Camera 1
MF
1
Raw Image Camera 1
Reference Vector Field 2D2C
for Camera 1: REF
1
Dewarping
t : MF - * REF
0 1 1
t +dt : MF + * REF
0 1 1
Projection to Camera 1
Vector Validation
Median Filter
Fill-up + Smooth
Correlation
t O t +dt
0 0
X
Mapping Function
of Camera 2
MF
2
Raw Image Camera 2
Reference Vector Field 2D2C
for Camera 2: REF
2
Dewarping
t : MF - * REF
0 2 2
t +dt : MF + * REF
0 2 2
Projection to Camera 2
Correlation
t O t +dt
0 0
X
Vector Validation
Median Filter
Fill-up + Smooth
Vector Validation
3C-Reconstruction
Eliminate 2D2C-Vectors
if Reconstruction Error > Threshold
2D3C
2D2C
2D2C
Final 2D3C Vector Field
+ REF
2
+ REF
1
Fig. 4. Flow chart of Stereo-PIV vector field computation
Figure 13.9: Flow chart of stereo PIV vector field computation.
114
less than 0.1 pixel, the reconstruction error is well below 0.5 pixel. This
can be used as an efficient rejection of false random vectors which usually
produce large reconstruction errors.
The complete flow chart is shown in fig. 13.9. The 2D2C-vector fields are
separately computed for camera 1 and 2. A typical multi-pass scheme
consists, for example, of one or two passes with interrogation window size
of 6464 pixel and 50% overlap followed by 4 passes with 3232 and 75%
overlap. After each pass a 2D3C-reconstruction is done for the purpose of
eliminating 2D2C-vectors with reconstruction errors above some threshold
(e.g. 1 pixel), but the 2D3C-vector field is not used further. Only at the end
the reconstructed 2D3C field is taken and validated e.g. by a median filter.
References
Details on the stereo vector calculation are given in:
Calluaud D, David L (2004), Stereoscopic particle image velocimetry mea-
surements of a flow around a surface-mounted block, Exp Fluids 36, p.33-
61.
115
13 Stereoscopic PIV
116
14 Stereoscopic PIV
Self-Calibration
With the Self-Calibration functionality you have the possibility to adjust
the coordinate system and the camera calibration of a stereoscopic PIV
setup so that the plane z=0 mm will be adjusted exactly in the middle of
the laser light sheet. This feature allows to eliminate potential errors in the
vector calculation due to misalignment of the calibration plate position and
light sheet position. The self-calibration is using the simple assumption
that images taken at the same time by camera 1 and camera 2 should
contain the identical information after the corresponding image correction
has been applied. Usually you find small discrepancies due to deflections
of the calibration plate position. These discrepancies can be determined
and corrected using the Stereo PIV self calibration wizard.
The Self-Calibration is only available after a stereoscopic calibration has
been generated already. The Stereo PIV self calibration wizard can be
opened as follows:
From the Project Manager by a right mouse click on the image set or
recording set and selecting of the Stereo PIV self calibration wizard
option.
From a stereoscopic particle image by a right mouse click on the im-
age and selecting the Send to Stereo PIV self calibration wizard
option in the context menu.
Using the Self-Calibr. button in the tool bar of the Project Manager.
117
14 Stereoscopic PIV Self-Calibration
The wizard will generate the self-calibration in three steps:
Step 1 : setup initial disparity map
Step 2 : calculate disparity map / update calibration
Step 3 : refine new calibration
14.1 Introduction
In the following paragraphs the terms disparity vector, disparity map,
correlation map, sum of correlation and correlation of sum will be
explained.
14.1.1 Disparity Vector
The concept of the disparity vector is very important in the understand-
ing of the self calibration procedure. Suppose a particle in the laser light
sheet (see figure 14.1). After the camera calibration, the z=0 mm plane
(the plane through the origin, perpendicular to the z-axis) is defined by the
surface of the calibration plate. The position and orientation of the laser
light sheet will normally differ more or less from the z=0 mm plane. By
dewarping (or correcting) an image for a certain camera, the images are
transformed in a manner as if they had been captured from a camera view-
ing perpendicular onto the light sheet from a position on the z-axis. To this
end it is assumed that all particles where located in the z=0 mm plane.
If the laser light sheet would be exactly the z=0 mm plane, then the de-
warped images from camera 1 and camera 2 would show the particles at
the same x,y positions in the dewarped images. If however the laser light
sheet is behind the z=0 mm plane (relative to the cameras), like in figure
14.1, then the dewarped position of a particle will be shifted to the left
for the left camera 1 and to the right for the right camera 2. The differ-
ence between the two positions will increase with the distance of the laser
light sheet from the z=0 mm plane. The vector connection the dewarped
positions for camera 1 and camera 2 is called the disparity vector.
14.1.2 Disparity Map
A disparity map is created by separating the dewarped images in small
windows with a certain size and overlap. Then a disparity vector is calcu-
lated for each window, by calculating the shift between the particle patterns
from the dewarped images of camera 1 and camera 2.
118
14.1 Introduction
z
x
light sheet
calibration
plate
disparity
vector
particle
camera1
particle in dewarped
image of camera1
camera2
particle in dewarped
image of camera2
z
x
light sheet
calibration
plate
particle
camera1 camera2
Figure 14.1: The top image shows a Stereo PIV setup where the laser light
sheet is not in the z=0 mm plane. If the images of camera 1 and camera 2
are dewarped (corrected) to the z=0 mm plane, then the particle is mapped
to different positions for camera 1 and camera 2. The vector between this
mapped positions is the disparity vector for this particle. The bottom image
shows a translated and rotated coordinate system to make sure that it coin-
cides with the light sheet position. Here the length of the disparity vector is
reduced to zero.
119
Figure 14.2: A disparity map. In this example the camera 1 was to the right
of camera 2.
14.1.3 Correlation Map
Figure 14.3: A part of a correlation map showing the characteristic elliptical
correlation peaks.
The disparity map is calculated from peak positions of the correlation map.
The correlation map is made up by the two dimensional correlation of the
windows fromthe dewarped images of camera 1 and 2. The disparity vector
of a window corresponds to the location of the correlation peak.
After a successful self calibration procedure, the disparity map should con-
tain only zero vectors (or very small vectors). The correlation map should
show zero centered peaks in this case in each window.
14.2 Requirements for the self-calibration proce-
dure
There are certain requirements for the usage of self-calibration:
You need a stereoscopic PIV setup.
120
14.3 Step 1: setup the initial disparity map
The system needs to be calibrated with a stereo pinhole or polynomial
of 3rd order calibration based on a two-level calibration.
You need a series of about 20 to 100 stereo PIV images from your
experiment (single or double frame images). The quality of the self-
calibration will increase with the used number of images.
To open the PIV self calibration wizard select the set with 20 to 100 PIV im-
ages in the tree view and then press the Self-Calibr. button in the Project
Manager. This will lead you to step 1.
In the first step an initial disparity map is created. The initial disparity map
is used as a starting point for the disparity map calculation in step 2. Four
items are visible in this step:
1. The parameter sub dialog where you enter the cameras in use and
the method of initial disparity map definition.
2. A stereo view for the images of the PIV image set, with an image slider
and a frame selector.
3. A view for the current initial disparity map, which is updated every
time a relevant parameter is changed.
121
4. A sub dialog for user defined point correspondences.
First you need to enter the correct camera numbers for camera 1 and cam-
era 2 in the parameter sub dialog. In most cases this will be number 1 for
camera 1 and number 2 for camera 2.
With the dewarp check box you can select the way in which the stereo
images are displayed. If dewarp is not checked, the raw images are dis-
played. If dewarp is checked, the images are dewarped (corrected) using
the current camera calibration for camera 1 and camera 2. This setting only
changes the view and does not affect the calculation in any way. When the
dewarp option is checked, you can easily estimate the disparity of a parti-
cle or structure in the images by pointing the cursor on the same structure
in both images and subtract the corresponding coordinates (displayed in
the lower right of the DaVis main window, either in units or with the
SHIFT key pressed in pixel).
The initial disparity map can be constructed with two different methods:
With a constant disparity vector for the complete image, or from user de-
fined point correspondences. Use constant shift (pixel) for the former and
user defined point correspondences for the later method. If the expected
disparity is small (0-50 pixel) the easiest way to get the initial disparity map
is to use the constant shift and enter dx=0 and dy=0.
If the expected disparity is big (>50 pixel) you can enter the values for
a map with constant disparity vectors in the parameter sub dialog, e.g.
dx=-260 and dy=0. A quick check for the initial disparity map is to go
to the second step and to compare the image frames in the lower left view
(see below). Use the method described above in the dewarped images to
estimate the disparity vector.
122
The other method is to use user defined point correspondences for the
initial disparity map. You may define up to 15 corresponding points in the
image taken by camera 1 and camera 2. First you have to enable a point
with the corresponding check box. Then you press the button with the dis-
played coordinates. Click a structure with the mouse cursor first in the left
view and then the same structure in the right image view. The selected
point will be displayed in the images and the coordinates in pixel are dis-
played on the points button. The disparity map is updated when the point
has been defined. In a similar manner, define more corresponding points,
until you are satisfied with the initial disparity map.
If you have finished the setup for the initial disparity map, use the go to
next step button to enter step 2.
123
14.4 Step 2: calculate disparity map / update cal-
ibration
After you have set up the initial disparity map, the disparity map is com-
puted from the input images and the camera calibration can be adjusted.
The dialog for the second step includes four parts:
1. The parameter sub dialog where you enter the window size and over-
lap, and determine the range of the input images to be used for the
calculation of the disparity map. Choose the method for the image
summing: sum of correlation or correlation of sum (see below).
2. A view for the input images in the lower left. You can switch between
the view of the first and the second camera with the frame slider.
When you enter the dialog, the input images are dewarped and cor-
rected with the initial disparity map. So if the initial disparity map
would be perfect, then the images of the first and second camera
should fit and no more shift should be visible. However a perfect fit is
not needed for the calculation of the disparity map. A remaining shift
of up to 20 pixel is acceptable.
3. A result display in the lower right. Here you can view the different
results of the calculation process: The initial disparity map, the calcu-
lated disparity map, a vector buffer with the vectors pointing to the
124
14.4 Step 2: calculate disparity map / update calibration
light sheet plane, a vector buffer representing the remaining triangu-
lation distance, and the correlation map.
4. An info text window where information about the recalibration process
will be displayed.
14.4.1 Parameter Setup
First select the desired window size and overlap for the disparity map.
A size of 128 or 256 pixel and an overlap of 50% will be all right in most
applications. For small images (<1k1k) smaller sizes (<128) and for very
big images (>2k2k) larger sizes (>256) may be useful. Adjust size and
overlap so that the resulting disparity map will contain at least 20 vectors
or up to about 150 vectors (e.g. 45 to 1015). Small window sizes may
also be useful, if the disparity map is very inhomogeneous. Then select the
input image range. You should use at least 20 images, better 50 to 100
images. The more images, the better will be the fit of the resulting camera
calibration to the laser light sheet.
A mask can be used to mask out the desired areas of the particle images.
However, the has to be applied to the images in a separate step using the
Group: masking functions and Operation: mask out image in the
Processing prior the self-calibration. For details on the masking procedure
please refer to the Imaging Tools manual (#1003012).
14.4.2 Sum of correlation / correlation of sum and passes
Select the summing mode for the input images and the number of
passes:
Sum of correlation: In this mode a correlation map (see above) is
calculated for each stereo image pair. Then all the correlation maps
are summed and the disparity vectors are calculated from the result-
ing sum of correlation maps. This method is more robust against
noise and will in general give the better correlation maps. It is rela-
tively slow, however. A single pass or two passes are enough using
this summing mode.
Correlation of sum: In this mode, the images for each camera are
summed up first. Then a single correlation map is calculated from the
125
summed stereo image and the disparity vectors are calculated. This
method is less robust against noise, but it is fast, and the image sum
only has to be calculated for the first pass. Use up to ten passes
here.
We recommend to use the first option by default.
14.4.3 Advanced settings
Make sure that normally the Refit external camera parameter check box
in the Advanced settings ... dialog is not checked (compare 14.5.1).
After all parameters have been set, start the calculation with the Calcu-
late new calibration button. The current calibration is not altered by the
calculation.
14.4.4 Info text output
The output during the calculation process in the info text window will be
like the following example:
The average length of the disparity vectors is displayed:
A triangulation is performed for each disparity vector in the disparity map
to calculate the 3D position of a structure. Then the residual triangulation
error is evaluated. To this end, the allowed triangulation error is iterated. It
starts with one pixel, and is increased stepwise by one until at most 30% of
the vectors have a triangulation error smaller than the allowed error.
126
14.4 Step 2: calculate disparity map / update calibration
A plane is fitted to the triangulated points and the plane parameters, rota-
tion angle for x and y axis, the translation in the z-direction and the average
deviation from the plane, are displayed.
Then the new camera calibration parameter are displayed, together with
the changes relative to the old parameter.
14.4.5 Accept / discard new calibration
After the calculation of the new calibration, you should inspect the info text
messages and then accept or discard the new camera calibration. If you
discard, you may leave the dialog, or change some parameter and do the
calculation again. If you accept, you will enter step 3 in which you can
Refine the camera calibration.
You need to select between the following options:
Select the first option to overwrite the calibration of the currently ac-
tive project - be aware that previously recorded images and their de-
rived results might not be consistent with results that will be calcu-
lated with the new calibration.
Select the second option to overwrite the calibration of the currently
active project and scale all previous recordings present in the root
level of your project to be consistent with the new calibration - be
aware that results that have been derived before the changed cal-
127
ibration might not be consistent with results that will be calculated
with the new calibration.
14.5 Step 3: refine new calibration
The fit of the laser light sheet to the z=0 mm plane can in many cases be
improved by one or more Refine steps. The meaning of the calculation
parameter and the advanced settings are the same as in step 2 (see above
for details).
The procedure is the same as in step 2: Enter the desired parameter, do a
Refine, accept or discard the new calibration. After each Refine step you
can inspect the info text output to decide whether another Refine step is
necessary or not. Another Refine step is useful if
The average disparity still decreases from step to step.
The plane parameter ( x-angle, y-angle, z-translation) change signifi-
cantly from step to step.
The new (external or internal) camera calibration parameter (rotation,
translation, focal length, ...) change significantly from step to step.
Note that the disparity vectors in the image above are magnified by a factor
of 50, this is useful to examine the disparity map in detail.
128
14.5 Step 3: refine new calibration
14.5.1 Advanced settings: Refit camera parameter
You can switch between two different modes to adjust the camera calibra-
tion with the check box Refit external camera parameter in the Advanced
settings ... dialog.
1. If Refit external camera parameter is not checked, the camera cal-
ibration system is only rotated and translated, so that the z=0 mm
plane will coincide with the laser light sheet. The relative camera po-
sitions of camera 1 and 2 and the internal camera parameter (like
focal length etc.) of the pinhole models for both cameras are not
changed.
2. If Refit external camera parameter is checked, the position and ori-
entation of camera 1 and camera 2 are refitted independently, to fit
to the new z=0 plane. In this way the relative positions of the cam-
eras according to the pinhole model may change. You may also select
internal parameters to be refitted.
14.5.2 When should internal camera parameters be refit-
ted?
The internal camera parameters can not be refitted if the calibration was
calculated using the polynomial fit. Normally the camera parameters
129
should not be refitted. The refit of external or internal camera parame-
ter is useful, if the vector buffer representing the remaining triangulation
distance (see above) or the initial disparity map shows a systematic pattern
(e.g. all vectors point up or down, or all vectors point away from a central
point) even after several Refine steps. This systematic patterns indicate,
that the initial camera calibration was not perfect or that the physical cam-
era setup has been altered after the calibration (cameras refocused, or a
small displacement of the cameras). In such cases, the camera calibra-
tion can benefit from a refit of external or internal parameter. As a rule of
thumb, you should first try to remove a systematic deviation pattern by a
refit of the external parameter only. If the pattern persists, then try to fit
the internal parameter as well.
14.6 Warning
14.6.1 Old results become invalid
After the self calibration procedure, the camera calibration is modified. So
all old results that have been calculated using the calibration before the
self calibration will not fit to the results after the self calibration. So it is
best to first do the self calibration and then calculate other results from the
recorded data.
14.6.2 Original scaling pixel/millimeter is changed
The self calibration procedure may change the original scaling from pixel to
millimeter in the raw images. As the vector calculation with activated Use
image correction option will use the Calibration.set in the Properties of
the corresponding project to determine the conversion factor from particle
image shift in pixel to velocity in m/s this is not critical for the velocity
calculation.
130
14.7 Calibration History
14.7 Calibration History
You may at any time load an old calibration from the Calibration History.
In the Project Manager you may find previous calculated calibrations under
Properties Calibration History. These calibrations have either been gen-
erated using the calibration dialog, using recordings of a calibration plate or
has been imported from other projects. To import and old calibration, use
the Import Calibration context menu of the Properties folder (DaVis
7.1 and higher).
References
Details on the self calibration are given in:
Wieneke B (2004), Stereo-PIV using Self-Calibration on Particle Images,
Experiments in Fluids, submitted.
Wieneke B (2003), Stereo-PIV using Self-Calibration on Particle Images,
5th International Symposium on Particle Image Velocimetry (Busan), Paper
3106.
Hartley R, Zissermann A (2000), Multiple View Geometry in Computer Vi-
sion, Cambridge University Press, UK, ISBN 0521623049.
131
132
15 Proper Orthogonal
Decomposition (POD)
15.1 Introduction
The Proper Orthogonal Decomposition (POD), also known as Karhunen-
Leve expansion, is a method to decompose a set of vector fields (2D or
3D) into a set of empirical eigenfields, which describes the dominant be-
havior or dynamics of a given problem. The POD can examine a series of
input vector fields of a certain flow condition, each at a different instant
in time. These input vector fields are used to form an eigenvalue problem
that is solved to determine a set of optimal basis functions for representing
the flow field. Each eigenfield has an associated energy value. This value
reflects the fraction of the overall energy in the input vector fields, that is
represented by this eigenfield. The eigenfields are numbered by decreas-
ing energy values, so that the first eigenfield is the most important one.
After a POD, any vector field can be approximated by a superposition of
the most important eigenfields.
References
Details on the POD can be found in following monographs:
Leve M (1955), Probability Theory, D. Van Nostrand, New York, 1955.
Holmes P, Lumley J, Berkooz G (1996), Turbulence, Coherent Structures, Dy-
namical Systems and Symmetry, Cambridge University Press, Cambridge,
UK, 1996.
133
15 Proper Orthogonal Decomposition (POD)
15.2 POD Wizard step 1: Decomposition
For POD a set of vector fields in a DaVis PIV project is needed. To enter
the POD wizard select a vector field set in a PIV project. After that the POD
toolbar icon is enabled. Then press the POD icon in the toolbar to open the
POD wizard. The wizard opens and the source vector set is displayed in a
window manager.
134
The first step of the POD wizard has four subdialogs:
1. the data source,
2. the window manager,
3. the decomposition settings,
4. the result.
15.2.1 The data source
In the data source subdialog, the file name of the input vector field set is
displayed. With a mouse click on the "i" icon, additional informations about
the input set can be obtained. The maximum range and the number of
images in the set is displayed as well. Select the vector fields to be used
for POD with the selected range and increment.
15.2.2 The window manager
Configure the layout and the contents of the data windows in the bottom
part of the POD wizard with the window manager. Select the number of
displayed data windows: 1x1 up to 4x4 windows. Then click on a data
window to activated it. The active window is indicated by a blue bar on the
top. Then select the desired data to be displayed from the drop down list.
You may also optimize the size and resolution of all displayed windows at
once with the corresponding window manager icons.
15.2.3 The decomposition subdialog
In the decomposition subdialog, you can select a storage name for the
POD results. Use the Default button to restore the default storage name.
Open the settings dialog with the Settings ... button (see below). Finally
start the POD with the Start decomposition button.
135
When the decomposition is finished, the results are stored and can be dis-
played in the data area of the window manager. The following results are
calculated:
1. The mean vector field
2. The spatial eigenmodes: a set of orthogonal vector fields
3. The temporal modes: a time plot for the evolution of the spatial
eigenmodes over time
4. The energy curve for all eigenmodes, scaled in percent.
5. The sum of energies.
15.2.4 Settings
In the settings dialog, you can check the Use double precision ckeckbox,
to do the POD with a higher precision for the numerical calculation. The
required memory is higher for double precision. If the required memory for
double precision is too high (more than the physical RAM of the computer),
switch back to single precision. The required memory for POD is displayed
in this subdialog.
136
15.2.5 Results
After the decomposition, the name of the result set is displayed as the
Current result. If the storage name already exists, a number postfix is
appended for each new decomposition, so that no result is overwritten.
An earlier result can be loaded using the Load icon of the current result
subdialog.
137
15.3 POD Wizard step 2: Approximation
After the decomposition, you can use the go to next step button in the
toolbar to reach the second step of POD: Approximation. Here you can
calculate a vector set that approximates the original vector set using a
subset of the spatial eigenmodes.
You can also do an approximation for a vector set different from the decom-
position set:
1. Do a decomposition on the desired decomposition vector set and
leave the POD wizard by using the Close button in the toolbar.
2. Select the desired vector set for approximation in the project view and
enter the POD wizard again.
3. Use the Load icon in the current result subdialog, to load the de-
composition results.
4. Enter step 2 with the Next step button in the toolbar.
There are four subdialogs in the approximation step:
1. the data source,
2. the window manager,
3. the approximation settings,
138
15.3 POD Wizard step 2: Approximation
4. the result.
15.3.1 The data source
The data source subdialog is the same as in the first step. The file name of
the input vector field set is displayed. With a mouse click on the "i" icon,
additional informations about the input set can be obtained. The maximum
range and the number of images in the set is displayed as well. Select the
vector fields to be used for approximation with the selected range and
increment.
15.3.2 The window manager
Like in the first step, the displayed data can be configured with the window
manager.
15.3.3 The Approximation subdialog
In the approximation subdialog, you can select a storage name for the ap-
proximation results. Use the Default button to restore the default storage
name. Open the settings dialog with the Settings ... button (see below)
to set the order of the approximation. Finally start the approximation with
the Start Approximation button.
When the approximation is finished, the results are stored and can be dis-
played in the data area of the window manager. The following results are
calculated:
1. The approximated vector fields
2. The residual vector fields
3. The temporal modes
139
15.3.4 Settings
In the Settings dialog, you can set the Approximation order. The spatial
eigenmodes number one to number order are used to approximate the
input vector field. The higher the order, the better is the approximation, an
the smaller are the vectors in the residual vector fields.
15.3.5 Results
After the approximation, the name of the result set is displayed as the
Current result. If the storage name already exists, a number postfix is
appended for each new approximation, so that no result is overwritten.
An earlier approximation result can be loaded using the Load icon of the
current result subdialog.
140
16 Display Options
For 2D View and 3D view of vector display please refer to the DaVis
manual.
Figure 16.1: Stereoscopic vector field in 2D and 3D view.
This chapter gives additional comments to Correlation Map, Scatter Plot
and Probability Density Function display settings of vector fields.
16.1 Correlation Map & Vector Edit
To open the Correlation map dialog use the right mouse button on an
image and select the Send to Correlation map option.
141
16 Display Options
On the top of the dialog you find the Correlation parameter and the Vec-
tor component values, on the bottom left the particle image that is used
as the source for the correlation and on the bottom right the correlation
function at the certain position determined by the cursor on the particle
image.
When the cursor is moved on the particle image while the Switch mouse
selection on/off flag is off the correlation is recalculated automatically
for the specified Correlation window size at the Correlation window
position in pixel determined by the cursor position on the particle image
and the result of the correlation is updated. The left mouse button can be
used to set the Switch mouse selection on/off flag, now the position of
the correlation window is fixed and by moving the cursor on the correlation
the correlation height is determined and shown in the status line together
with the cursor position (scaled or in pixel coordinates when you keep the
SHIFT button on the keyboard pressed).
Switch mouse selection on/off: When the flag is off the correla-
tion map is updated automatically when the cursor is moved on the
particle image.
Correlation window position: Shows the center position of the in-
terrogation window depending on the cursor position on the particle
image.
142
16.1 Correlation Map & Vector Edit
Correlation window size: Determines the correlation window size
for the currently executed correlation.
Vector field was created using image correction: Determines
if the calculation of the correlation is taking image correction into
account or not.
Vector component values: Shows the position of the 1st, 2nd, 3rd
and 4th choice vector in [mm] for x- and y-component.
disabled: not used for correlation map.
If you open the Correlation map dialog based on a particle image the
location of the four highest correlation peaks are displayed in the Vector
component values textboxes. If you open the Correlation map dialog
based on a vector field you may access the check boxes in order to enable
a vector that corresponds to a different correlation peak or to disable the
vector at the is position manually.
In this case you will find also further Vector edit option on the right. You
can use them for following purposes:
Undo current vector: undo the modifications for the last vector that has
been changed manually.
Undo whole vector field: Undo the modifications for all vectors that
have been changed manually.
143
16 Display Options
Smooth current vector: Smooth the vector at the selected position over
a neighborhood of 33, 55 or 77 vectors.
Save all changes to file: Save the current vector field including all man-
ual modifications to a file on hard disk.
16.2 Scatter Plot
Using the right mouse button on a vector field you may obtain the Scatter
Plot option in the pop-up window. The dialog offers a visualization of the
velocity distribution of a vector field. Each velocity vector is indicated by a
dot in a two dimensional domain, e.g. Vx/Vy, Vx/Vz or Vy/Vz.
Figure 16.2: Vector field in 2D view and and as Scatter Plot.
This feature can be used to identify outliers in the vector field more easily
and to find a suitable Allowed vector range for the vector postprocessing.
In order to zoom in/out or to change the color use the buttons in the icon
bar. To move the scatter plot origin press the left mouse button with the
cursor on the scatter plot and move the mouse while the mouse button is
still pressed. Using the right mouse button on a Scatter plot and selecting
Data and display properties... you can access the display properties on
the Scatter Plot file card:
144
16.3 Probability Density Function
Data Range: Specifies the data range for the scatter plot function.
This can be either the complete vector field or one of the rectangles
1-8. When a rectangle is selected the Rect. button becomes active.
Using this button the software will ask you to declare a rectangle on
the corresponding vector field and is using this rectangle as input
range for the new scatter plot.
Calculating: Specifies the vector components Vx/Vy, Vx/Vz or Vy/Vz
to be drawn in units [pixel] or [m/s] in the scatter plot window.
Grid: Can be used to show a grid on the scatter plot for better lo-
calization of the velocity values. The grid size is set automatically to
integer values and is changing with the zoom, this can not be changed
manually.
Zoom: Manual selection of the zoom. Can be changed by moving
the slider. Value 1.00 indicates the optimal selection (just all vectors
inside the display window), when the zoom is changed the display is
updated automatically.
Number of vectors in scatter plot: Shows the number of vectors
in the scatter plot.
Using the right mouse button on a vector field you may obtain the Prob-
ability Density Function (PDF) option in the pop-up window. The PDF
computes a histogram of a vector field where the Vx, Vy and/or Vz compo-
nents are separated in certain velocity intervals.
The PDF curve can be used to check whether the velocity field computation
has a bias toward integer velocity values, which is called peak locking ef-
145
16 Display Options
fect. Peak locking can occur, for example, when the used seeding particles
are too small and produce particle images on the CCD of less than one pixel
in diameter.
1 2 3 4 5 6
1 2 3 4 5 6
A particle image size of 2-3 pixel guarantees a small bias to integer values.
A particle image size smaller than 1 pixel causes a strong bias to integer values.
Figure 16.3: Peak locking effect for small and big particle image size.
Using the right mouse button on a PDF and selecting Data and display
properties... you can access the display properties on the PDF file card:
Data Range: Specifies the data range for the PDF function. This can
be either the complete vector field or or one of the rectangles 1-8.
When a rectangle is selected the Rect. button becomes active. Using
this button the software will ask you to declare a rectangle on the
corresponding vector field and is using this rectangle as input range
for the new PDF.
Calculating: Specifies the velocity component for the calculation of
the PDF. Either a single velocity component Vx, Vy, Vz or (for better
statistics) the two components Vx,Vy can be selected.
146
Complete V: Histogram of the corresponding velocity component for
all values from minimum to maximum.
V mod. 1: Histogram of the decimal places of the corresponding
velocity component (cuts of the integer values, e.g. Vx=2.13 px
Vx=0.13 px.
V mod. 0.5: Works like V mod 1, velocities corresponding to a sep-
aration V > 0.5px are mapped to a value 1-V for a better estimation
of the peak locking effect.
Peak lock: This value allows to estimate the peak locking effect, a
value of:
0: indicates no peak locking effect,
1: indicates a strong peak locking effect,
< 0.1 indicates that the peak locking effect is acceptable.
The peak lock value is calculated from the center of mass of the
V mod.0.5 histogram: peak lock = 4 (0.25 centerofmass).
Since there is more or less a equipartition of the decimal places
the center of mass should be close to 0.25, when there is a high
peak locking effect due to small particles etc. the center of mass
is shifted to zero.
Slot resolution: Using the slider one can determine the slot resolu-
tion for display. The width of one bin can be selected within a range
of 1/30 px-3 px.
Number of vectors: Shows how many vectors were used to calcu-
late PDF.
Number of bins: Shows how many bins are used for display (number
of bins = reciprocal value of slot resolution).
In order to show the PDF in bars like in fig. 16.4 you need to select the
corresponding options in the Profile Display in the Display Settings.
147
16 Display Options
Figure 16.4: PDF showing low and high peak locking effect.
148
17 Vector related Expert
Settings
For advanced users there are some options available that help to address
special requirements or make DaVis legacy options available. In both
cases LaVision does not support these features officially, but you are
free to use them on your own risk.
Please note that for reasons of stability, robustness or precision of the re-
sults some specific default settings have been chosen. In some specific
cases using other options are required or may help to improve the result.
In general it is recommended to use the default options.
17.1 How to open the dialog
These non default options are globally valid and cannot be set in the PIV
processing parameter card settings. To open the dialog go to the project
menu and select the entry vector related expert settings...
A warning is displayed that changing settings is at your own risk.
Accept the warning to go ahead, which opens the following dialog:
149
17 Vector related Expert Settings
On the left side you see the default selection of all options. If any setting is
changed to a non default state a warning message Non default settings
appears at the bottom of the window to indicate that at least one option
is not the standard setting. To set all parameters to their default values
press the Reset button which will set all settings to default (in this case
the warning disappears).
17.2 What can be set in the dialog
Following is a list of all available options and their effect:
changing behavior (result will differ)
Changing one or more of the options will result in different vector results!
use default values (Lanczos 10) for high accuracy mode: default is
on. If switched off in the PIV parameter card you may select from dif-
ferent interpolation methods (bicubic spline, Whittaker, Lanczos) and
the filter length used for pixelwise deformed interrogation windows
when accessing to float pixel positions in images.
use bicubic vector interpolation: default is off. If switched on bicubic
interpolation is used when accessing to float grid positions in vector
fields for pixelwise deformed interrogation windows.
disable vector scale reset on entering project: default is off. If switched
on upon entering a PIV project the vector scale is not set to m/s (or
150
17.2 What can be set in the dialog
mm in Strain type projects) but becomes selectable in the PIV pa-
rameter card gives the choice between pixel , mm and m/s vector
component scaling.
take out rotation for extract scalar field (e.g. vector background):
default is default, i.e. never for PIV type projects and always for
Strain type projects. If switched to always this takes out the locally
present rotation in the vector field before calculating the gradients
E
xx
, E
xy
, E
yx
, E
yy
. if switched to never this is done not at any time.
use default restriction for peak search: is the default. If switched to
peak restriction the legacy mode from DaVis 7 is used where you
can specify that the peak search is restricted to a region relative to
the specified window size or to the absolute shift in pixel.
additional features
All these options will give you some extra functionality not available in the
standard PIV project.
register extra Batch OPs in PIV projects: default is off. If switched
on some extra processing operation become available in the process-
ing dialog. At the moment this is a processing operation to put the fo-
cus quality, the dynamic range and other overlays to already recorded
images. For a description of those values see the chapter Recording
on page 17.
show also auto correlation modes: default is off. If switched on ad-
ditionally to the cross correlation, sequential cross correlation and
stereo cross correlation modes for the vector calculation the auto
correlation, sequential auto correlation and stereo auto correlation
modes become available where two laser shots are captured in one
frame. Nowadays this technique is rarely used anymore.
show global correlation value for reconstructed image using vector result:
default is off. If switched on after each PIV calculation the original
images are shift corrected with the result vector field and the global
correlation value (over all pixels) c = (
I
1
I
2
)/(
I
1
I
1
I
2
I
2
is cal-
culated and displayed as overlay on the images. A value of 1 means
the images contain identical intensity values, a value of 0 means that
the correlation is totally random. A typical value achieved is 0.8. This
may be used to compare different PIV settings. Those that produce a
151
17 Vector related Expert Settings
vector field that leads to a higher global correlation values is able to
match patterns more accurately.
draw adaptive PIV ellipses in result buffer: default is off. If switched
on when using adaptive PIV overlay graphics are drawn to the result
vector field with the specified number of ellipses that are used to
weight the correlation functions and enhance the correlation in high
gradient velocity fields. Additionally you can increase size and line
width of the ellipses to make them visible more easily. This is normally
used to produce bitmaps for papers.
FFT by FFTW: default is FFT by FFTW which uses a very fast FFT library
licenced by LaVision. If switched to FFT by numerical recipies
a standard non optimized FFT implementation is used.
Do not forget to reset the settings to the default settings once the
modified settings are no longer required!
152
18 Customer Service
If you have a technical problem or questions regarding hardware or soft-
ware which is not adequately addressed in the documentation please con-
tact your local representative or LaVision service directly.
In order to speed up your request please include following information:
The order number of your system (see section 18.1).
The number of the used dongle (see section 18.1).
A short description of the problem.
Your customer settings file (see section 18.2).
A log.txt if you have a reproducible software problem (see sec-
tion 18.3).
An information on the used operating system and service pack.
You can contact service at LaVision GmbH by:
email: service@lavision.de
phone +49 551 9004 229
Alternatively you may submit your problem using the service request inter-
face on the LaVision homepage under the URL www.lavision.com.
18.1 Order and Dongle Number
To be able to find information on the delivered hardware components and
customer details in the database your order number is required. This num-
ber can be found in the DaVis/Help/About menu or on the original DaVis
installation CD (see Fig. 18.1).
In the About DaVis dialog you find the order number in the 3rd line.
The five digits behind the leading X are the order number followed by the
DaVis version and the release date.
The dongle number is required to exclude license problems. This number
is is written on the hardware key. The number of the used dongle can also
be found using the DaVis/Help/About menu and on the DaVis CD.
153
18 Customer Service
order #
dongle #
Figure 18.1: Order no. in DaVis/Help and on installation CD.
Please include the order number and the dongle number in your service
requests.
18.2 Customer Settings
To be able to reproduce your problem it could be essential to know the
exact hardware setup and software parameter in DaVis. All currently
used parameter can be extracted using the DaVis/Help/Create service
file for LaVision Support menu.
After you have selected this the system will write back all values for the
relevant variables to an lsf-file. This procedure will take a while indicated
by a Busy dialog. The lsf-file will also contain the current settings of the
hardware setup, acquisition setup and processing operation lists.
The lsf-file will be written automatically to a

davis/service subdirectory.
The name contains the order number and dongle number that is ex-
tracted from your software (#ordernumber_donglenumber.lsf). Send this
154
18.3 Log.txt
lsf-file as attachment to the description of you problem by email to ser-
vice@lavision.de.
18.3 Log.txt
If you have a reproducible software problem in DaVis please generate a
log-file. This file contains all functions you have called and all error mes-
sages that have been displayed after you have activated the log. Please
proceed as follows:
Start DaVis and use the Macro/Clear Log file menu.
Enable the Log each CL function call (slow) entry in the Macro
menu. This feature is active if you see a flag left next to the entry.
Every time you click on this entry its status is changed.
Try to reproduce your problem, e.g. until an error message is dis-
played.
A log.txt file has been generated in the DaVis main directory.
Send this text file attached to your email.
Disable the Macro/Log each CL function call (slow). This function
is deactivated if you dont see a flag next to the entry.
155
18 Customer Service
18.4 Export data set for support
Some problems can only be reproduced on images that contain particular
information or artifacts. For error analysis it can be necessary to provide
exemplary data that needs to be extracted from the corresponding project.
Depending on the project type, number of used cameras and error the cor-
responding calibration and derivative data can be required as well.
A convenient way to extract the data from the project is the Extract data
set for support option that can be selected via right mouse click on the
corresponding data set in the Project Manager.
In the Export data set for support dialog the Data source, the Maxi-
mum range and the number of Images is specified. For the Image se-
lection you need to select form following options:
156
18.4 Export data set for support
one image: This will extract one image only. You need to select the num-
ber of the image.
whole set: This will extract the complete data set. Please note that de-
pending on the number of images the resulting zip-file will be very
large.
specify range: This will extract the specified range. You need to select
minimum and maximum image number.
In addition you may activate the Calibration option and the corresponding
cameras. This will make sure that the calibration for these camera will be
included.
Using the OK button you need to specify the location and file name for the
zip-file that contains the selected data.
Note: Files with a size of more than 20 MB should not be send by
email. LaVision can provide a password protected location on a FTP
server for data exchange with read and write access. Please contact
service@lavision.de for details. For the upload your local user account
needs to have enabled FTP functionality. Please check with your system
administrator.
157
18 Customer Service
18.5 Camera Activity Logging
The camera support libraries (camera DLLs) used by DaVis have a built-
in logging function which might be useful for camera activity monitoring
and debugging. If you experience problems with camera behavior and con-
tact LaVision support, you may be asked to generate a log file during
camera operation and send it by email.
You can use the camera activity logging by choosing the appropriate item in
the DaVis Help menus. Open the following modal dialog window which
serves for starting and stopping the logging function, specifying logging
options and viewing the resulting log output. If you create a lsf-service file
after camera logging (see 18.2), the camera log data will be included.
In the top line of the dialog the Current logging status is displayed. If
this is not the OFF status, a number is displayed. Using the following ra-
dio buttons you can choose if you want to Log each CL function call +
programmed output, i.e. the complete communication between DaVis
and the camera DLL, or to Log only programmed output (PrintLog).
Programmed output lines are the result of calls to the PrintLog function
inside the DLL or, from DaVis, results of a CL function call like DevSetStr-
Par(DEV_CAMERA_1, CAMPAR_DEBUG, output string). If you choose Log
each DLL function call, each Get() and Set() function call from DaVis
to the DLL is logged together with the value of the parameter passed re-
spectively the result returned by the DLL.
Using the four checkboxes you can specify the output format and writing
mode:
Show execution times (differences): If this option is checked, each
output line is preceded by the time passed since the preceding output
line (microseconds).
158
18.5 Camera Activity Logging
Show absolute time stamps : If this option is checked, each output line
is additionally preceded by an absolute time counter value (microsec-
onds). This counter can be reset by calling the TimerReset() function
inside the DLL
Close log file after each output line (for TAIL): This options must be
checked if a TAIL program is used for watching the log file, or if you
expect a possible crash of DaVis.
Disadvantage: the logging is much slower (minimum time for an out-
put line about 0.5 millisecond). If this option is unchecked, the log file
remains open so you can view it only after switching the logging off.
Append to log file, dont clear it: If this option is unchecked, the out-
put CameraLog.txt is cleared (i.e. deleted) and a new logging is
started. If the option is checked, new logging output is appended to
the existing file.
On the bottom of the dialog, the syntax of a CL function call is displayed
which sets the logging (debugging) to the state displayed in the dialog win-
dow.
The Logging On button closes the dialog and activates the logging. After
this, the activity of all cameras is logged in the CameraLog.txt file. Unless
you specify the 3rd option above, the file remains open for writing so you
cannot view (read) it.
By clicking the Logging OFF, Show log file button you stop camera ac-
tivity logging (if it was active). The resulting CameraLog.txt file is closed
and immediately displayed in a editor (default: ConText).
Logging is automatically activated when DaVis initializes the cameras
during startup (or after clicking Initialize in the Hardware Setup dialog).
You can see initialization details, some camera properties or possibly error
messages in the CameraLog.txt.
NOTE: The Camera activity logging is only supported for cameras that are
used via camera DLL. This is NOT the case for the ImagerFamily (FlowMas-
ter3S, ImagerIntense, NanoStar) in DaVis 7.2.
159
18 Customer Service
18.6 Shipment of defective items
If some items need to be returned to LaVision GmbH for service or
repair please contact the LaVision service to obtain a RMA (Return Ma-
terial Authorization) number. Fill out the Shipment Cover Letter that is
available in DaVis/Help directory, in the Help folder on the DaVis in-
stallation CD or at LaVision service. List all items with SN and a short
description of the problem. Place the Shipment Cover Letter in the box
with the item(s) being returned. Return the authorized item(s) per shipping
instructions.
Shipping instructions:
Be sure to obtain a RMA number.
Include the Shipment Cover Letter.
Ship only the items that are authorized.
Use the original boxes to avoid damages during transportation.
Remove cooling water from the laser!
Use antistatic bags for computer boards!
Ship returned items to:
LaVision GmbH
Anna-Vandenhoeck-Ring 19
D-37081 Goettingen
Germany
Note: Shipments received by LaVision without a RMA number may be
refused.
160
LaVision GmbH
Anna-Vandenhoeck-Ring 19
D-37081 Goettingen, Germany
www.lavision.com
Email: sales@lavision.com
Tel.: +49(0)551-9004-0
Fax: +49(0)551-9004-100
LaVision
We count on Photons
LaVisionUK Ltd
Downsview House/ Grove Technology Park
Grove, Oxon, Ox 12 9FF/ UK
www.lavisionuk.com
Email: sales@lavision.com
Tel.: +44-(0)-870-997-6532
Fax: +44-(0)-870-762-6252
LaVision, Inc.
211 W. Michigan Ave., Suite 100
Ypsilanti, MI 48197, USA
www.lavisioninc.com
Email: sales@lavisioninc.com
Phone: +1(0)734-485-0913
Fax: +1(0)240-465-4306
M
a
n
u
a
l
N
o
1
0
0
3
0
0
5
-
F
e
b
r
u
a
r
y
1
4
,
2
0
1
1

FlowMaster D80

Uploaded by

Copyright:

Available Formats

FlowMaster D80

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

FlowMaster D80

Uploaded by

Copyright:

Available Formats

LaVision

along the line through the center of the

and the value perpendicular to this direction dV

sinc(x) sinc(x/a) a < x < a, x = 0

in the middle of the image it may be only 89

. This means that e.g. a velocity displacement in x-

)! So one should at least be aware

the correction procedure as described in detail in the DaVis

deviation from the standard perpendicular

has been determined by a proper

. One problem of an oblique

You might also like