Professional Documents
Culture Documents
BACHELOR OF TECHNOLOGY
in
MECHANICAL ENGINNEERING
by
Supervisor(s)
Dr. BALAJI SUBRAMANIAN
We declare that this written submission represents our ideas in our own words
and where others’ ideas or words have been included, we have adequately cited and
referenced the original sources. we also declare that we have adhered to all principles of
academic honesty and integrity and have not misrepresented or fabricated or falsified
any idea/data/fact/source in our submission to the best of our knowledge. We understand
that any violation of the above will be cause for disciplinary action by the Institute and
can also evoke penal action from the sources which have thus not been properly cited or
from whom proper permission has not been taken when needed.
We would like to express to deepest appreciation and a sense of gratitude to our guide ,
Dr. Balaji Subramanian, department of Mechanical Engineering, Indian institute
technology Tirupati, along him we would also like to thank Shanthan Kumar Reddy, MS
Scholar.
i
ABSTRACT
KEYWORDS: PIV
This report highlights the approach and the progress in accomplishing the project.
Our project mainly deals with one of the flow visualization techniques Particle Image
Velocimetry (PIV). PIV can be divided into two parts, Image acquisition and Image
analysis. Our work involves only the image analysis part of the PIV. A basic working
code to visualize flow with an accuracy upto 2 decimals has been developed and validated
it with images from PIVLAB.
ii
TABLE OF CONTENTS
ACKNOWLEDGEMENTS i
ABSTRACT ii
LIST OF FIGURES iv
ABBREVIATIONS v
1 INTRODUCTION 1
1.1 Objectives and Scope . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Literature Review 2
2.1 Cross Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 FFT Cross correlation . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 5 point Gaussian surface fit . . . . . . . . . . . . . . . . . . . . . . 3
3 Description 5
3.1 Mathematical Formulation . . . . . . . . . . . . . . . . . . . . . . 5
3.2 Experimental Setup Materials Used . . . . . . . . . . . . . . . . . 6
3.3 Procedure, Techniques and Methodologies,etc . . . . . . . . . . . . 7
3.3.1 Window Shifting . . . . . . . . . . . . . . . . . . . . . . . 8
iv
ABBREVIATIONS
v
CHAPTER 1
INTRODUCTION
Particle Image Velocimetry (PIV) is the standard technique which is used for flow mea-
surement in both fundamental research and for industrial applications. Its applications
are getting widened day by day. It has a vast scope from aerodynamic to biological
applications. It has become more intriguing in the recent years as volumetric velocity
measurements are made possible. In-house code helps us to have control over it and
customize it to our test conditions. Our objectives are to develop and validate an efficient
MATLAB code for PIV.
CHAPTER 2
Literature Review
In PIV, the main objective is to match the patterns in two images. This is done by a
technique called Cross Correlation. Basically, in signal processing, cross correlation is
where you take two signals and produce a third signal (correlated signal). This method is
a way to objectively compare different time series and allows you to see how two signals
match and where the best match occurs. In case of a 2-d matrix i.e., image, performing
2-d cross correlation on two images gives a correlation matrix.
After correlating both the matrices, a peak occurs in the correlation matrix C when
there is a best match between the matrices.
2.2 FFT Cross correlation
The standard cross correlation takes a lot of computational time(N 4 steps). So, FFT cross
correlation([N×log2 N]2 )which significantly reduces computational time is used as an
alternative. In FFT cross correlation the images are converted into frequency domain and
then complex conjugate multiplication is performed followed by conversion to spatial
domain to acquire correlation matrix. The entire FFT cross correlation method can be
represented as in figure 2.1.
Usually, when we get a peak in correlation matrix and find its position to calculate
the displacement, the displacement will be accurate only up to a single pixel. This is
called Peak Locking effect. To achieve higher accuracies or to avoid this effect, 5 point
Gaussian surface fit is used. In which the correlation curve around the peak is always
assumed to be a Gaussian surface. Given by the equation 1, And then we take four points
around the peak where z2, z3, z4, z5 gives values of correlation matrix coefficients and
(x2, y2), (x3, y3), (x4, y4), (x5, y5) are their respective coordinates on matrix plane from
the origin O as shown in the figure 2.1. Similarly, the peak is given by (x1, y1, z1). Now
3
we solve for the 5 coefficients (c0, c1, c2, c3, c4) of the equation 1 by substituting the
above 5 points in equation 1. Once we get these coefficients, the 3D curve equation is
found. We later use differential calculus to find the true peak.
4
CHAPTER 3
Description
The mathematical formulation and the procedure we followed to attain the objective are
discussed here.
As we see in the above figure, suppose initially at time ’t’ in image G1 (x,y) a small
window of size M×N is given as GM . And when the time is t+∆t, the same window on
image G2 (x,y) is given by GS , with a shift of m pixels in x direction and n pixels in y
direction the total shift is represented as S where S is the resultant of m and n vectors.
Then the windows GM and GS are given as
To compare the pattern in g1 (i,j) and g2 (i+m,j+n) we use a method called Minimum
Quadratic Difference (MQD) which is given by equation 2
1
D(m,n) = ∑M−1 ∑N−1 2
j=0 [g1 (i,j)-g2 (i+m,j+n)] → MinimumQuadraticDi f f erence(MQD)
M.N i=0
Figure 3.1: Surface plot of the MQD function
By using equation 2 we find MQD matrix(D(m,n)) from g1 and g2 matrices which will
have minimum peak when both the matrices match perfectly as shown in the figure 3.1.
We can also find the shift S from the location of the peak in terms of m and n, as S is the
resultant of m and n.
On further expanding the equation 2 and separating the only term with both g1 and g2
we get equation 2.1,
1 M−1 N−1
∑M−1 N−1
i=0 ∑ j=0 [g1 (i,j)-g2 (i+m,j+n)] = [∑ ∑ j=0 g1 (i,j)2 - M×N × D(m, n)+∑M−1
2 i=0 i=0
∑N−1 2
j=0 g2 (i+m,j+n) ]
Here, LHS is called the Correlation Term, and the matrix formed from this term is
called correlation matrix.
In case of correlation matrix, we get a maximum peak because we made the negative
term positive.
As our work only deals with image analysis part, MATLAB software is used.
6
3.3 Procedure, Techniques and Methodologies,etc
Initially, two images are read and stored in a 3D array format. Later, we convert them
into gray scale images i.e., into 2D arrays. Then, some pre-processing techniques like
contrast adjustment and noise removal are performed. The de-noised images are now
padded in such a way that the size of the images are divisible by 64 so that we can
conveniently break the image into 128 or 64 or 32 sized windows with 50% overlap
in both x and y directions. Now the image 1 is divided into 32 sized windows with
50% overlap and the image 2 is divided into 64 sized windows with 75% overlap in
such a way that every 32 sized window in image 1 except for the windows at the border
has a 64 sized window in image 2 with the same center as the corresponding 32 sized
window. Now, the 32 sized windows are zero padded into 64 sized windows and FFT
cross correlation is performed with corresponding 64 sized windows in image 2 to get
the shifts.
This is done because sometimes the patterns might leave the windows. So, to increase
the search area a 64 sized window is taken around a 32 sized window as shown in figure
3.2. This technique is called Correlation Tracking.
7
If the RMS errors for the methods FFT correlation method, MQD method, and correlation
tracking method are compared from the figure 3.3, it seems that MQD method is the
best method. However, it takes a lot of computational time. So, the next best alternative
appears to be Correlation Tracking, as the error is continuously increasing in FFT
correlation method. But, for displacements very close to zero FFT correlation has the
Figure 3.3: Evaluation error for ideal PIV recordings by using different algorithms with
a 64x64-pixel interrogation window
least error. So, if the measuring method is somehow made a null measurement technique
then FFT correlation is the best method both in terms of computational time and accuracy.
The technique which is used to convert the measurement into a null type is called
Window Shifting.
Window shifting is an iterative process in which the shifts calculated in the previous
iteration are used in the present iteration to shift the windows in image 2 by their
respective shifts in previous iteration. As a result the shift in each iteration keeps
reducing and converges to zero, making it a null type measurement. For example, in the
figure 3.4, the previous shift is SW S so the window is shifted by SW S to get the new shift
S’ and the total shift is given by S = S’ + SW S
8
Figure 3.4: Window Shifting
However, for the first iteration, which gives the first estimate of shifts, window
shifting cannot be done as there is no previous iteration. This is the reason correlation
tracking is done for the first iteration as discussed at the start of this chapter.
For the second iteration both the padded images are divided into 32 sized windows
with 50% overlap in both directions. Now the shifts from previous iteration corre-
sponding to each 32 sized windows are used to shift the windows on the second image
accordingly. Then again FFT correlation is done and the peaks are located followed by a
5 point Gaussian surface fit to get more accurate peak locations and shifts corresponding
to second iteration. This process is repeated until the shifts converge to zero. And the
shifts of every iteration are added to find the total shifts. Finally, these total shifts are
plotted to visualize the flow.
9
CHAPTER 4
For a synthetic image of pure translation of 5.23 pixels, the output is shown in figure
4.1(a). An almost same output is yielded.
For a synthetic image of a Rankine vortex, the output is shown in figure 4.2(a). A
similar output is yielded with few error vectors.
A high RMS is observed at 5.5 compared to other points as correlation tracking is used
in iteration (correlation tracking is known to have high errors at 0.5 pixel displacement).
11
CHAPTER 5
A basic working PIV code is achieved. However, still effects like shear and rotation are
to be accounted for. The code works for any set of images with no shear and rotation
effects. An accuracy up to 2 decimals is achieved. Future work includes implementing
window deformation to account for shear effects and also central difference method will
be used to calculate the velocities.
REFERENCES
1. https://observablehq.com/@lemonnish/cross-correlation-of-2-matrices
2. http://publish.illinois.edu/piv15conference/scope/
3. Steven T. Wereley (Spring, 2014). [Fundamentals of Particle Image Velocimetry].
https://nanohub.org/resources/20295series.
4. L. Gui el al. A comparative study of the MQD method and several correlation-based
PIV (2000). Iowa City, USA.
5. Markus Raffel et al. Particle Image Velocimetry A Practical Guide. Germany, 2018‘
6. https://www.photonics.com/Articles/Particle_Image_Velocimetry_Basics_Developments/a25121
13