Introduction of Particle Image

Velocimetry

Ken Kiger
Burgers Program For Fluid Dynamics
Turbulence School
College Park, Maryland, May 24-27
Slides largely generated by
J. Westerweel & C. Poelma of Technical University of Delft
Adapted by K. Kiger

Introduction
Particle Image Velocimetry (PIV):
Imaging of tracer particles, calculate displacement: local fluid velocity
Frame 1: t = t0

Measurement
section
Light sheet
optics
Frame 2: t = t0 + t

Twin Nd:YAG
laser

CCD camera

Introduction
Particle Image Velocimetry (PIV)
992

1004

32

32

divide image pair in

interrogation regions
small region:
~ uniform motion
compute displacement
repeat !!!

Introduction
Particle Image Velocimetry (PIV):
Instantaneous measurement of 2 components in a plane
conventional methods
(HWA, LDV)

particle image velocimetry

whole-field method

non-intrusive (seeding

instantaneous flow field

single-point measurement
traversing of flow domain
time consuming
only turbulence statistics

Introduction
Particle Image Velocimetry (PIV):
Instantaneous measurement of 2 components in a plane
instantaneous vorticity field

particle image velocimetry

whole-field method

non-intrusive (seeding

instantaneous flow field

Example: coherent structures

Example: coherent structures

Turbulent pipe flow
Re = 5300
10085 vectors

hairpin
vortex

Example: coherent structures

Van Doorne, et al.

Overview

PIV components:
- tracer particles
- light source
- light sheet optics
- camera

Hardware (imaging)

- measurement settings
- interrogation
- post-processing

Software (image analysis)

Tracer particles
Assumptions:
- homogeneously distributed
- follow flow perfectly
- uniform displacement within interrogation region
Criteria:
-easily visible
-particles should not influence fluid flow!

small, volume fraction < 10-4

Image density

low image density

NI << 1
particle tracking velocimetry
high image density
NI >> 1
particle image velocimetry

Evaluation at higher density

High NI : no longer possible/desirable to follow individual tracer particles
Possible matches

Sum of all possibilities

Assumption:
uniform flow
in interrogation area

Particle can be matched with a number of candidates

Repeat process for other particles, sum up: wrong combinations will
lead to noise, but true displacement will dominate

Statistical estimate of particle

motion

Statistical correlations used to find

average particle displacement

1-d image @ t=t0

Bx By

(I (k,l) I )(I (k + i,l + j) I )

a

R(i, j) =

1
2

2
2

( Ia (k,l) Ia ) ( Ib (k + i,l + j) Ib )
k=1
l=1

k=1 l=1
Bx By

k=1 l=1

1
Ia
Bx B y

Bx By

Bx

1-d image @ t=t1

By

I
k 1 l 1

(k , l )

Cross-correlation

1-D cross-correlation example

Ia(x) normalized
Ib(x+x) normalized
Ia(x)*Ib(x+x)

Ia

Ib

Bx

(I (k) I )( I (k + i) I )
a

R(i) =

k=1

Bx
B x
2
2
2

( Ia (k) Ia ) ( Ib (k + i) Ib )
k=1

k=1

Finding the maximum displacement

-Shift 2nd window with respect to the first
Typically 16x16 or 32x32 pixels
- Calculate match

Bx By

(I (k,l) I )(I (k + i,l + j) I )

a

Good indicator: R(i, j) =

- Repeat to find best estimate

k=1 l=1

1
2

2
2

( Ia (k,l) Ia ) ( Ib (k + i,l + j) Ib )
k=1
l=1

k=1 l=1
Bx By

1
Ia
Bx B y

Bx By

Bx

By

I
k 1 l 1

(k , l )

Finding the maximum displacement

-Shift 2nd window with respect to the first
Typically 16x16 or 32x32 pixels
- Calculate match

Bx By

(I (k,l) I )(I (k + i,l + j) I )

a

Good indicator: R(i, j) =

- Repeat to find best estimate

k=1 l=1

1
2

2
2

( Ia (k,l) Ia ) ( Ib (k + i,l + j) Ib )
k=1
l=1

k=1 l=1
Bx By

1
Ia
Bx B y

Bx By

Bx

By

I
k 1 l 1

(k , l )

Bad match: sum of product of intensities low

Finding the maximum displacement

-Shift 2nd window with respect to the first
Typically 16x16 or 32x32 pixels
- Calculate match

Bx By

(I (k,l) I )(I (k + i,l + j) I )

a

Good indicator: R(i, j) =

- Repeat to find best estimate

k=1 l=1

1
2

2
2

( Ia (k,l) Ia ) ( Ib (k + i,l + j) Ib )
k=1
l=1

k=1 l=1
Bx By

1
Ia
Bx B y

Bx By

Bx

By

I
k 1 l 1

(k , l )

Good match: sum of product of intensities high

Can be implemented as 2D FFT for digitized data

o

Impose periodic conditions on interrogation regioncauses bias error if not treated

properly.

Cross-correlation
This shifting method can formally be expressed as a cross-correlation:

R (s) I1 x I 2 x s dx
- I1 and I2 are interrogation areas (sub-windows) of the total frames
- x is interrogation location
- s is the shift between the images
Backbone of PIV:
-cross-correlation of interrogation areas
-find location of displacement peak

Cross-correlation

peak: mean
displacement

RC
correlation
of the mean

RF

correlation of mean &

random fluctuations

RD
correlation due
to displacement

Influence of NI
RD (s D ) ~ N I

NI = 5

Cz0 2
NI
DI
2
M0

NI = 10

C
z0
DI
M0

particle concentration
light sheet thickness
int. area size
magnification

NI = 25

More particles: better signal-to-noise ratio

Unambiguous detection of peak from noise:
NI=10 (average), minimum of 4 per area in 95% of areas
(number of tracer particles is a Poisson distribution)

Influence of NI

NI = 5

NI = 10

NI = 25

PTV: 1 particle used for velocity estimate; error e

PIV: error ~ e/sqrt(NI)

Influence of in-plane displacement

X,Y-Displacement
<
quarter of window size

X / DI = 0.00
FI = 1.00

0.28
0.64

0.56
0.36

RD (s D ) ~ N I FI FI (X , Y ) 1
DI

0.85
0.16

Y
DI

Influence of out-of-plane displacement

Z-Displacement
<
quarter of light sheet thickness
(z0)

Z / z0 = 0.00
FO = 1.00

0.25
0.75

RD (s D ) ~ N I FI FO

0.50
0.50
z
FO ( z ) 1
z0

0.75
0.25

Influence of gradients
Displacement differences <
3-5% of int. area size, DI

a M0ut

Displacement differences <

Particle image size, d

a / DI = 0.00
a / d = 0.00

0.05
0.50

RD (s D ) ~ N I FI FO F
R.D. Keane & R.J. Adrian

0.10
1.00

0.15
1.50

F ( a) exp( a 2 / d2 )

PIV Design rules

image density

NI >10

in-plane motion

|X| < DI

out-of-plane motion |z| < z0

spatial gradients

M0ut < d

Obtained by Keane & Adrian (1993) using synthetic data

Window shifting
in-plane motion

|X| < DI

strongly limits dynamic range of PIV

large window size: too much spatial averaging

Window shifting
in-plane motion

|X| < DI

strongly limits dynamic range of PIV

small window size: too much in-plane pair loss

Window shifting
in-plane motion

|X| < DI

strongly limits dynamic range of PIV

Multi-pass approach:
start with large windows,
use this result as pre-shift
for smaller windows
No more in-plane pair loss limitations!

Window shifting: Example

Grid turbulence
fixed windows

windows at same location

matched windows

windows at 7px downstream

Window shifting: Example

Vortex ring, decreasing window sizes

Raffel,
Willert and
Kompenhans

Sub-pixel accuracy
Maximum in the correlation plane: single-pixel resolution of displacement?

But the peak contains a lot more information!

Gaussian particle images Gaussian correlation peak (but smeared)

Sub-pixel accuracy

Fractional displacement can be obtained using

the distribution of gray values around maximum

three-point estimators
peak centroid

R1 R1

R1 R0 R1

parabolic peak fit

Gaussian peak fit

R1 R1

2 R1 R1 2 R0
ln R1 ln R1

2 ln R1 ln R1 2 ln R0

balance

normalization

Peak locking

zig-zag structure, sudden kinks in the flow

Sub-pixel Interpolation Errors

Can exhibit peak locking

Interpolation of peak is biased towards
a symmetric data distribution (Integer
and 1/2 integer peak locations)
Polynomials exhibit strong locking
when particle diameter is small
Gaussian is most commonly used
Splines are very robust, but expensive
to calculate

See Particle Image Velocimetry, by Raffel,

Willert, and Kompenhans, Springer-Verlag,
1998.

Interpolation error (pixels)

particle image size

noise in data (seeding density, camera
noise)
shear rate

0.2

Pixel
Radiu
s 0.5
R=

Polynomial fit
Gaussian fit

0.1
0

-0.1
-0.2
-0.3
-0.5

0.3
Interpolation error (pixels)

Accuracy depends on:

0.2

-0.25
0
0.25
Actual peak location (pixels)

0.5

Polynomial fit
Gaussian fit

0.1

R = 1.0

-0.1
-0.2
-0.3
-0.5

0.3
Interpolation error (pixels)

0.3

0.2

-0.25
0
0.25
Actual peak location (pixels)

0.5

Polynomial fit
Gaussian fit

0.1

R = 1.5

-0.1
-0.2
-0.3
-0.5

-0.25
0
0.25
Actual peak location (pixels)

0.5

Peak locking
Histogram of velocities in a turbulent flow

Gaussian peak fit

centroid
Even with Gaussian peak fit:
particle image size too small peak locking
(Consider a point particle sampled by discrete pixels)

Sub-pixel accuracy
optimal resolution:
particle image size: ~2 px
Smaller: particle no longer resolved
Larger: random noise increase

three-point estimators:
Peak centroid
Parabolic peak fit
Gaussian peak fit
...

total
error

bias errors

random errors

d / dr

Theoretical: 0.01 0.05 px

In practice 0.05-0.1 px
Main difference: sensitivity to peak locking
or pixel lock-in, bias towards integer displacements

displacement measurement
error

window matching
fixed windows

matched windows

FI ~ 0.75

FI ~ 1

velocity pdf
measurement
error

u2

signal

u2

C2

noise

4C2u2

u2 / C2

SNR

1 / 4C2

application example:
grid-generated turbulence
X = 7 px
fixed windows

u/U = 2.5%
matched windows

Data Validation

article

lab

Spurious or Bad vectors

Spurious vectors
Three main causes:
-

insufficient particle-image pairs

in-plane loss-of-pairs, out-of-plane loss-of-pairs

gradients

Effect of tracer density

N
NII=80
=20
=45
=11
= 153
I=

Remedies

increase NI
practical limitations:
optical transparency of the fluid
two-phase effects
image saturation / speckle

detection, removal & replacement

keep finite NI ( ~ 0.05 )
data loss is small
signal loss occurs in isolated points
data recovery by interpolation

Detection methods
human perception
peak height
amount of correlated signal
peak detectability
peak height relative to noise
lower limit for SNR
residual vector analysis
fluctuation of displacement
multiplication of correlation planes
fluid mechanics
continuity
fuzzy logic & neural nets

Residual analysis

evaluate fluctuation of measured velocity residual

ideally: Uref = true velocity
reference values:
Uref = global mean velocity

r U U ref

comparable to 2D-histogram analysis

does not take local coherent motion into account
probably only works in homogeneous turbulence
Uref = local (33) mean velocity

takes local coherent motion into account

very sensitive to outliers in the local neighborhood
Uref = local (33) median velocity
almost identical statistical properties as local mean
Strongly suppressed sensitivity to outliers in heighborhood

Example of residual test

4

2.3
5

2.2

3.0

3.7

3.1

3.2

2.4

3.5

2.7

Mean
2.9

RMS
0.53
2.2

2.3

2.4

2.7

3.0

3.1

3.2

3.5

3.7
Mean
3.7

2.3

9.7

3.0

3.7

3.1

3.2

2.4

3.5

2.7
RMS
2.29

2.3

2.4

2.7

3.0

3.1

3.2

3.5

3.7

Standard mean and r.m.s. are very sensitive to bad data

contamination need robust measure of fluctuation

9.7

Median test
1

1 - Calculate reference velocity: median of 8 neighbors

uref = median( u1,u2 ,...u8 )

2 calculate residuals:

ri = ui uref

3 Normalize target residual by:u median(r

i) +
0 uref
*
r0 =
median(ri ) +

Westerweel & Scarano, 2005

4 Robust measure found for:

0.1andr0* > 2

Interpolation

Bilinear interpolation satisfies continuity

For 5% bad vectors, 80% of the vectors are isolated
Bad vector can be recovered without any problems

N.B.: interpolation biases statistics (power spectra, correlation function)

Better not to replace bad vectors (use e.g. slotting method)

Overlapping windows
Method to increase data yield:
Allow overlap between adjacent interrogation areas
a

Motivation: particle pairs near edges

contribute less to correlation result;
Shift window so they are in the center:
additional, relatively uncorrelated result

50% is very common, but beware of oversampling

A Generic PIV program

Data acquisition
Reduce non-uniformity
of illumination; Reflections

Image pre-processing
PIV cross-correlation

Median test,
Search window
Vorticity, interpolation
of missing vectors, etc.

Laser control,
Camera settings, etc.

Vector validation
Post-processing

Pre-shift; Decreasing
window sizes

PIV software
Free
PIVware: command line, linux (Westerweel)
JPIV: Java version of PIVware (Vennemann)
MatPiv: Matlab PIV toolbox (Cambridge, Sveen)
URAPIV: Matlab PIV toolbox (Gurka and Liberzon)
DigiFlow (Cambridge), PIV Sleuth (UIUC), MPIV, GPIV, CIV, OSIV,
Commercial
PIVtec
TSI
Dantec
LaVision
Oxford Lasers/ILA

PIVview
Insight
Flowmap
DaVis
VidPIV

Particle Motion: tracer particle

Equation of motion for spherical particle:

dv p

dv

1 du
d
u
p

g
mp
3D u v p m f

mp 1
mf

dt
2 dt
dt
dt

Inertia

Where

Viscous drag

Added mass

Pressure
gradient

buoyancy

D 3
m p = p
,theparticlemass
6
D 3
m f = f
,fluidmassofsamevolumeasparticle
6
D = particlediameter
= fluidviscosity
r
u= fluidvelocity
r
vp = particlevelocity
g = fluiddensity
p = particlematerialdensity

Neglect: non-linear drag (only really needed for high-speed flows),

Basset history term (higher order effect)

Simple thought experiment

Lets see how a particle responds to a step change in
velocity
Only consider viscous drag
r
dvp
r r
mp
= 3D( u vp )
dt

0fort

<0
u =
0
Ufort

dv p
18
1
=
U

v
=
(
)
(U v p )
p
dt
p D2
p

vp+

1
1
vp = U
p
p

v p = (v p ) particular + (v p ) homogeneous = U + C1 exp( t p )

v p = U [1 exp( t p )]

u, vp
U

Particle Transfer Function

Useful to examine steady-state particle response to 1-D

oscillating flow of arbitrary sum of frequencies
Represent u as an infinite sum of harmonic functions
Neglect gravity (DC response, not transient)

u( t ) =

du( t )
=
dt

( ) exp(it ) d

v p (t) =

( ) exp(it ) d

dt

( ) exp(it ) d

( ) exp(it ) d

2m p dv p
m f m p du
dv p m f m p du
= 2( u v p ) +

+
2
3D dt
m p 3D dt
dt m p 3D dt

p D2
2 18 p i exp(it )d =
0

dv p ( t )

2(

p ) exp(it )d +

2iSt p = 2( f p ) + iSt [ 3 f p ]

St =

f p D2
i [ 3 f p ] exp(it ) d
p 18

p
f

p D2
18

f
p

Particle Transfer Function

This can be rearranged:

? ?1 2
?A 2 + B 2 ?1 2
vp
pp
r = ? ? = ? 2
?
u ? f f ? ? A + 1 ?

?Im( / ) ?
?A(1 B) ?
p
f
?= tan1? 2
tan ?
?
?
?
A
+
B
Re

/
?
?
(
)
p
f
?
?
1

2
St 2
3
B
2
A

Examine

Liquid in air, gas in water, plastic in water

Liquid particles in Air

Liquid drops in air require St~0.3 for 95% fidelity (1 m ~ 15 kHz)

Size relatively unimportant for near-neutrally buoyant particles
Bubbles are a poor choice: always overrespond unless quite small