Experiments in Fluids 10, 181-193 (1991)

Experimentsin Fluids
9 Springer-Verlag 1991

Digital particle image velocimetry

C. E. Willert and M. Gharib

Dept. of Applied Mechanics and Engineering Sciences, R-0t l, University of California, San Diego, La Jolla, CA 92093, USA
Abstract. Digital particle image velocimetry (DPIV) is the digital

counterpart of conventional laser speckle velocitmetry (LSV) and particle image velocimetry (PIV) techniques. In this novel, two-dimensional technique, digitally recorded video images are analyzed computationally, removing both the photographic and opto-mechanical processing steps inherent to PIV and LSV. The directional ambiguity generally associated with PIV and LSV is resolved by implementing local spatial cross-correlations between two sequential single-exposedparticle images. The images are recorded at: video rate (30 Hz or slower) which currently limits the application of the technique to low speed flows until digital, high resolution video systems with higher framing rates become more economically feasible. Sequential imaging makes it possible to study unsteady phenomena like the temporal evolution of a vortex ring described in this paper. The spatial velocity measurements are compared with data obtained by direct measurement of the separation of individual particle pairs. Recovered velocity data are used to compute the spatial and temporal vorticity distribution and the circulation of the vortex ring.

1 Introduction

Conventional two-dimensional particle tracing methods suffer from complications that arise from the tracking of individual particles when obtaining velocity field information. While methods such as PIV and LSV inherently do not require the tracking of individual particles, they offer a higher temporal and spatial resolution of the instantaneous flow field. The LSV and PIV techniques rely on the images of double (or multiple) exposed seed particles within a thin sheet of light in the flow. Although the techniques are essentially the same, LSV uses a particle concentration so dense that individual particles are no longer distinguishable on the image plane; that is, they appear as speckle. However, as Adrian (1984) argues, in most practical cases low seeding densities exist, which results in the imaging of individual particles, and the technique is then referred to as PIV. In both cases, the photographically recorded transparency images are mechanically interrogated at many locations with a probing laser. A Fourier transformed image of the illuminated region is then obtained by passing the diffracted laser beam through a converging lens and projecting it onto a

screen. If the region contains double exposed speckle or particle images, a Young's fringe pattern will be generated. The spacing and angle of these interference fringes was shown by Burch and Tokarski (1968) to correspond directly to the displacement of the speckle patterns, or, as in the case of PIV, to the displacement of the particle images. Measurement of the fringe spacing and angle is either performed directly using image processing techniques or by digitally Fourier transforming the pattern. The latter process yields a spatial autocorrelation of the fringe pattern from which a displacement magnitude and angle may be estimated. An alternative to the Young's fringe method of image interrogation is the faster optical correlation technique suggested by Coupland and Halliwell (1988). LSV was first demonstrated on fluid flows by Barker and Fourney (1977) with a low Reynolds number Poiseuille flow in a glass tube and by Grousson and Mallick (1977) in a similar experiment. Transient Brnard convection was studied using LSV by Simpkins and Dudderar (1978). Other studies using LSV include the axisymmetric vortex roll-up and pairing events of a jet in air (Meynart 1983), vortex rings in water (Meynart and Lourenco 1984) and the convection driven flow in a liquid cooled from above (Meynart et al. 1987). PIV has also been applied in numerous fluids experiments such as the turbulent wake of a cylinder (Kompenhans and Reichmuth 1987; Lourenco and Krothapalli 1987), the flow around cavitation bubbles (Vogel and Lauterborn 1988), the flow behind an impulsively started airfoil (Lourenco et al. 1986), and jets in water (Adrian 1986 a). An excellent review of these two spatial velocimetry techniques and similar methods is given by Adrian (1986b). Reliance on opto-mechanical techniques to obtain velocity data makes PIV and LSV time consuming and tedious, especially when the photographic processes are included. Furthermore, the separation of the exposed particles on the image imposes a lower limit on recoverable velocity vectors. This means that overlapping particles in a low velocity region of the flow cannot yield velocity vectors unless an image shifting technique is used (Adrian 1986a; Landreth and

182 Adrian 1988). To account for this spatial shift the recorded transparency image must be interrogated with a proportionally larger beam which reduces the spatial resolution. Because PIV and LSV rely on single images for the flow velocity measurements, phase information is lost. This causes a directional ambiguity for each recovered velocity vector in the field, especially in flows of spatially varying direction. Although the angle and magnitude of the vector is known, its polarity is not. To avoid this problem, image shifting (Adrian 1986a; Landreth and Adrian 1988) may be applied during the image capture. Color coding (Zhang 1986) has been suggested as well. Unfortunately, these techniques further complicate the experimental setup and subsequent data analysis. In an effort to remove the complex optical analysis processes, Blackwelder et al. (1988) digitized photographic records of a particulate flow and performed the analysis digitally by cross-correlating small sections of two images. In an attempt to eliminate some of the complications associated with LSV and PIV, some video-based techniques have been suggested and applied. Cho (1989) formulated the equations associated with the digital counterpart of the Young's fringe phenomenon and applied them to a numerically simulated model flow. A technique using frame-toframe tracer cross-correlation of video images has been applied by Kimura and Takamori (1986) to obtain the unsteady flow field behind a cylinder. Both techniques have the advantage of using two separate images in which particle images that normally would overlap one another in a double exposed image are now distinguishable. Adrian (1988) has suggested a similar technique but no details about its nature have yet been published. In this paper the authors will present a digital approach to resolve many of the problems associated with the conventional LSV and PIV methods and demonstrate the technique's feasibility by mapping the temporal evolution of a vortex ring. Essentially a digital counterpart to PIV, digital particle image velocimetry (DPIV) emphasizes the use of digitally recorded video images rather than photographic transparencies. All analysis is performed computationally, thereby removing the need for photographic processing and optical interrogation of the images. The use of separate images of single-exposed particle images removes the directional ambiguity in the recovered data by preserving the phase of the images. This also allows the displacement measurement of partially overlapping particle images. Rather than using a digital counterpart of the Young's fringe technique of local image interrogation as suggested by Cho (1989), a local spatial cross-correlation is performed which yields translational displacement data directly. Although the method described is similar in principle to the one described by Kimura and Takamori (1986), the methodology of obtaining the cross-correlation and the actual displacement vector is different. Their technique only allowed for a resolution of one picture element (pixel) in identifying the peak correlation, whereas the largest displacements measured were on the order of ten pixels. Consequently, the relative

Experiments in Fluids 10 (1991) error in their displacement measurements is significant. In the technique presented here, the peak correlation is obtained at sub-pixel accuracy which lowers the relative errors in the measurements. This is a major requirement for the calculation of secondary quantities through spatial differentiation or integration of the flow field data. From a technological point of view, the current maximum of 30 Hz (video rate) for the image acquisition rate limits the application of the technique to low speed flows until high speed, high resolution digital video systems (512 by 512 pixels and greater) become more readily available and less costly. It should be noted that the emphasis of DPIV is to use video related technology in an effort to improve on the image acquisition and processing aspects with respect to the PIV techniques. With respect to spatial resolution and application to high speed flows the current state of PIV offers better capabilities for the time being. On the other hand, DPIV's current limitations are more of a technological nature rather than a conceptual one. The first of the following eight sections addresses the mathematics associated with the formation of the images and the cross-correlation process.

2 The techniques of deconvolution and cross-correlation in DPIV

A digitized video image may be considered to be a two-dimensional signal field analogous to a digital time series in one dimension. Many one-dimensional signal processing techniques can be readily extended to two dimensions (e.g. Rosenfeld and Kak 1976; Pratt 1984). In DPIV two sequential digital images are subsampled at one particular area via an interrogation window (Fig. 1). Within these image samples an average spatial shift of particles may be observed from one sample to its counterpart in the other image, provided a flow is present in the illuminated plane. This spatial shift may be described quite simply with a linear digital signal processing model shown in Fig. 2. One of the sampled regionsf (m, n) may be considered the input to a system whose output g(m, n) corresponds to the sampled region of another image taken a time At later. The system itself consists of two components, a spatial displacement function s(m, n) (also known as the system's impulse response) and an additive noise process d (m, n). This noise

I
D f(m,n)
Image 1 to

g(m,n)
Image 2 t o , At

s'(

Estimated fields of displacement funct ons

Fig. 1. Conceptual arrangement of frame-to-frame subsampling associated with digital particle image velocimetry

C. E. Willert and M. Gharib: Digital particle image velocimetry

Input image {Image 1) Image transfer function (Spatiat shift ) Output image (Image 2 )

183 crete Dirac delta functions do not allow sub-pixel measurements because more than one non-zero point is required to interpolate between pixels. In the technique of Kimura and Takamori (1986) the correlation peak could only be resolved with an uncertainty of one pixel. The method of choice in finding the displacement function s(m, n) is the statistical technique of spatial cross-correlation. The discrete cross-correlation function tklg(m, n) of the sampled regions f (m, n) and g(m, n) is given by the expected value E: q~s0(m, n) = E [ f (n, m), g (m, n)] (4 a)

f(m,n) F(u,v)

g'(m,n) ~ "1 $(u'V) GI--T~"~ T d(m,n)

D(u,v)
Additive noise process

g(m,n) :- G(u,v)

Fig. 2. Digital signal processing model describing the functional relationship of two successiveframes of displaced particles; capitalized functions are the Fourier transforms of the functions in lower case and represent the spatial frequency domain process is a direct result of particles moving off the edges of the sampling region, particles disappearing through threedimensional motions in the laser sheet, and the total number of particles present in the window. Of course the original samples f (m, n) and g (m, n) may be noisy as well. The major task in DPIV is the estimation of the spatial shifting function s (m, n) but the presence of noise d (m, n) complicates the matters. A description of how the output sample g (m, n) relates to the input sample f (m, n) can be given mathematically as: g (m, n)=[f(m, n)* s (m, n)]+d(m, n) (1) where * denotes the spatial convolution of the two functions f (m, n) and s (m, n) and can be rewritten for this discrete case
as

~, f(k,l)9(k+m,l+n) (aso(m,n)_ k=-ool=-~ ~ f(k,l)~, ~ O(k,l)

k=-ool=-~ k=-ool=-oo

(4b)

A high cross-correlation value near 1 is observed where many particle images match up with their corresponding spatially shifted partners, small cross-correlation peaks may be observed when individual particle images match up with other particle images. The highest correlation peak is considered to represent the best match of particle images between the functions f(m, n) and g(m, n). The maximum of this cross-correlation peak coincides with the location of the displacement delta function 5 ( m - i, n-j). To illustrate this, Eq. (4) may be rewritten in absence of the noise effects d(m, n) as: q~Io(m, n)= E [f (n, m), O(m, n)] (5 a) (5 b) (5 c)

9 (m, n) Lk=-oo/=oo

s ( k - m, l - n) f (k, 1) + d (m, n)

(2)

dpyo(m, n) = E [f (m, n),f(m, n)* s (m, n)] 49.ro(m, n)= q~f y (m, n)* s (m, n)

The displacement function s(m, n) for this discrete case is nothing more than a Dirac delta function 6 (m-i, n - j ) displaced from the origin by i,j units and this displacement corresponds directly to the average displacement of the partides in the sampled region. Division of the displacement by the time scale At between the capture of the images gives the average velocity in the sampled region. One conceivable way of recovering the spatial displacement function, or delta function s(m, n), is via a deconvolution technique assuming that the noise effects are negligible. Deconvolution may be reduced to division in the frequency domain by use of the convolution property

g(m,n)~f(m,n)*s(m,n)<->G(u,v)~F(u,v)S(u,v) (3) where G(u, v), F(u, v)and S(u, v) are the discrete Fourier
transforms of the corresponding functions in lower case letters. An approximation to S(u, v) can be obtained with Eq. (3) if the effects of d (m, n) can be neglected, and inversely transforming S (u, v) recovers the desired displacement function s (m, n). Although simple in implementation, this method is very sensitive to the effects of the noise d(m, n) which hides the delta function between the spurious peaks of the noise. Also, as will be described later, the displacement function needs to be measured to sub-pixel accuracy to achieve sufficient resolution in the particle displacement values. Dis-

where ~bll (m, n) is the autocorrelation function of the input functionf(m, n). Equation (5 c) may be interpreted as follows: the location of the individual particls is a stationary random process such that the input function f(m, n) correlates with itself only at the origin. The convolution with the displacement functionf(m, n) in Eq. (5c) moves the correlation peak away from the origin by the average spatial displacement of particles in that region. Including the effects of noise results in a decrease in height of the cross-correlation peak relative to the background. Spatial displacement gradients within the sampling regions cause a broadening of the peak. Figures 3 and 4 illustrate graphically the cross-correlation technique applied to two different image samples. In Fig. 3 Dirac delta functions are used to represent particle images within 32 by 32 square sampling regions of two sequential images. The cross-correlation between these regions shows a large spike (delta function) displaced from the origin corresponding to the spatial shift of particles images from one region to the next. The smaller spikes are a result of particle images correlating with other particle images and not with themselves. Figure 4 is a sample section of particle images recorded with the image acquisition system which is described in Sect. 5. These samples contain a considerable

184

Experiments in Fluids 10 (1991)

30

30

30 2O

30 20

30

30

20 ~

20

Shi~ 0 ~

0 ,,

Sh,7 0 ~

~,

Fig. 3a-e. Cross-correlation estimate e of image pair a, b which consist of randomly positioned particles represented as Dirac delta functions; the relative spatial shift is - 4 pixels along the Y-axis yet uniform background noise level as is evidenced around the borders of the plots. The cross-correlation between these two samples results in a broader cross-correlation peak than in plot (Fig. 3 c) but allows sub-pixel measurement using an appropriate centroiding technique. As a result of the broader particle images, the relative background noise level in the cross-correlation plot (Fig. 4c) is much higher than in plot (Fig. 3 c). This noise is not only a result of particle image mismatch but can also be attributed to particle images leaving and entering the sampling region from one sample to the next. The images formed by the particles on the imaging array are best described by the radius a of their Gaussian-like intensity profile I(x):
I (x) = I o exp [ - ( x

Fig. 4a-r Same as Fig. 3, but for samples taken from recorded images (particle displacement ~ 8 pixels along Y-axis)

3 Computational implementation of D P I V

2a 2

--Xo)2] j

(6)

A radius a of I pixel means that 68.3% of light scattered by the particle is imaged onto a circular area of 1 pixel radius. The intensity each pixed records is the spatial integral of Eq. (6) over the pixel area. The resulting distribution of the intensity readings has a Gaussian shape as well.

Figure 5 describes the actual implementation of the D P I V technique. The most important feature in the authors' version of the technique is the use of fast Fourier transforms (FFT's) to simplify and significantly speed up the cross-correlation process. Rather than performing a sum over all the elements of the sampled region for each element as in Eq. (4b), the operation can be reduced to a complex conjugate multiplication of each corresponding pair of Fourier coefficients. This new set of coefficients is then inversely transformed to obtain the cross-correlation q~Io' Because of the FFT's periodicity in space there is no need for normalization of ckyo as in the linear case of Eq. (4 b). The normalization coefficients are constants for the finite sampling domain. The Nyquist sampling criterion associated with the discrete Fourier transforms limits the maximum recoverable spatial displacement in any sampling direction to half the window size in that direction. In reality, even this displacement is often too large for the technique to work properly, since the signal to noise ratio in the cross-correlation decreases with increasing spatial shift. That means the number

C. E. Willert and M. Gharib: Digital particle image velocimetry

System input Image subsampting at position ( i,j } Operations in spatial. frequency domain l Cross- I "'"' I - ~ correlation~ CrossConversion corretation to vetocity peak detection System output

185

~ v } ~
Image (to) I ~

i6,(m,n)~dx,dy
=1 I I =i~,~;iL~il---~gata rite

Image2

1~'Ig(,,,.~I--j~G(,.v)

I ~"(,%)= .I---q r'~' I

Fig. 5. Numerical processing flow-chart of DPIV process; cross-correlation estimates are calculated by complex conjugate multiplication in the spatial frequency domain

of particle image pairings decreases in the sampling region while more particle images are unpaired. Given a window side of length N, we found approximately a third of this length (N/3) to be an adequate limit for recoverability of the displacement vector. This directly corresponds to a similar criterion in PIV and LSV where half of the interrogation beam diameter is the upper limit for displacement recovery (Attrian 1986b; Meynart and Lourenco 1984). The cross-correlation peak is initially determined by finding the highest value in the two-dimensional array of correlation values. Around this element a parabolic or exponential curve fit in both the horizontal and vertical direction yields the approximate location of the correlation peak to sub-pixel accuracy. A three-point curve fit was chosen rather than the more traditional method of center-of-mass centroiding because the latter requires a properly selected cutoff threshold to obtain reasonable correlation estimates. A series of experiments that go beyond the scope of this paper monitored the performance of both centroiding techniques. They showed that the three-point exponential curve fit was able to estimate the correlation peak with a much lower error than the center-of-mass centroiding technique. This phenomenon can be attributed to the roughly Gaussian shape of the cross-correlation peak itself. The current image digitization speed of 30 Hz limits the highest detectable particle velocity magnitude, given a certain size N by M of the sampling window. A maximum measurable displacement of particles in the flow can be determined easily by multiplying the pixel displacement by the magnification factor between the image plane and the object plane. Dividing the displacement by the time constant fit between the capture of images (1/30 of a second), a maximum detectable flow velocity can be determined. The aspect ratio of the sampling window can also be changed to match a large scale flow. For flows with a mean cross-image convection velocity the sampling regions can be shifted digitally with respect to each other, offsetting much of the convective motion. The Young's fringe method of PIV and LSV image interrogation does not permit such a respective offset of the sampling regions. Given a transparency in which image shifting has been used, such an offset would be advantageous in order to retain the spatial resolution that could be achieved without a shift. As a consequence, the shifted image has to be interrogated with a beam that is larger proportional to the applied shift.

4 Resolution and uncertainty limits

A way to assess the limits of resolution and the level of uncertainty of the DPIV technique is to simulate a flow field via the translation or rotation of a random set of points. These simulated particle images could either be generated numerically or imaged with the camera from a screen. Several numerical as well as optical resolution experiments similar to those described by Browand and Plocher (1985) were performed by the authors and have shown that the technique is capable of resolving velocity vectors in the range of serveral hundredths of a pixel depending on a variety of parameters. Some of the factors entering into the resolving error are particle image size, the size of the interrogation window, local velocity gradients, the number of particles within the sampling window, intrumentation performance characteristics (camera and digitizer), quantization effects, and computational errors (rounding, truncation). One specific series of tests allowed the authors to directly estimate the level of uncertainty in the measurements for the ideal case of linear displacement. A random pattern of black dots on a white background was mounted on a linear translation stage and was imaged by the image acquisition system that is decribed in the following section. First, several images were captured with no relative displacement, inverted to yield bright "particle images" on a dark background, and then processed by the DPIV code. Processing was performed using a 32 by 32 pixel sized window which was chosen for the analysis of the images recorded in the vortex ring experiment (Sect. 5) The statistical uncertainty in the measurements (Fig. 6 a) reflects the best performance of the authors' DPIV system, taking into account the effects of noise in the acquisition system as well as numerical noise generated by particle images being truncated on the edges of the interrogation window. Image pairs were then shifted mechanically with respect to each other over the entire of possible displacements in the 32 by 32 pixel window. These results are compiled in Fig. 6 b and show that the lowest obtainable uncertainty is approximately 0.01 pixels. As the displacements increase, the associated uncertainties increase as well, but not in a linear fashion; the relative error actually decreases with increasing displacements. Also increased seeding densities decrease the level of uncertainty, apparently because more particle pairs are taken into account. The lowest level of uncertainty can be considered to be the displacement

186
0.050 0.045
--

Experiments in Fluids 10 (1991)

1.0
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

[] Vertical fluctuation o Horizontal fluctuation 0.8

~0.6

/////. ,--'" .......

//

0.040 0.035

.....................

bY_ g
1:3

0.030 0,025
o. EJ "'"--..... "o.

% 0.020 0.015 0.010 0.005 0

&

~
~

0.4
.
i

i ,,' i i i i

.
'

.
,\ 'x

.
,/
/

.
i

/ /"

,// /

.
- -

~',,,

z 0.2

"...,.
"'""o"-"-" ._-0 .......................... 0 ......~ .......................................... [3

......

162 Window 322 Window 642 Window

rq.................. - ~ " " ".............. E]

0
80

32

64

96

128

192

256

Spatial wavelength ( Pixets ) 10 20 30 40 50 60 Seeding density (Dots/Window) 70

0.16 0.14 0.12 0.10 = o 0.08 '~ 5: cr 0.06 0.04

Fig. 7. Spatial wavelength response of DPIV for various window sizes

Seeding density for 322 Window

[] O A +

61 Particles I Particles
/ /

ii
.~ /,,~

24 Particles 36 Particles (> 46 Particles 74 Particles

J3"" /

.._e ....

O""

",/

/,[ O

f',,[] ..-- / @

,."
.....

B/

0 ~. ....... ................... _~,

..-

.<=..~

,'Zli

.........

0.02 0
1 0 -2

......... .......... " ...,<,<-'~. ..... ~ ...... ~

10 ~ 100 DispLacement ( Pixe[s )

101

Fig. 6a and b. Measurement uncertainties of DPIV with respect to uniform spatial shift and seeding density: a no spatial shift, i.e. system noise only: b nonzero spatial shift

detection limit for the above configuration. The associated dynamic range, the ratio of maximum and minimum of measurable displacements, is on the order of a 1,000-1. Further numerically based simulations showed that the 8 bit quantization of the digitizer limits the lowest uncertainty obtained by a 32 by 32 pixel window to 2/1,000 of a pixel. These simulations generated and cross-correlated samples of randomly distributed particle images with of varying radius a (Eq. 6). The discrepancy between the numerical simulation result and the lowest measured value of 0.01 pixels is thought to be not only due to recording noise but also to a nonlinear intensity response of the imaging array. This means that a recording of the Gaussian-like intensity distribution of the particle images does not retain the Gaussian shape. Another important aspect of DPIV is its spatial resolution which is limited directly by the particle seeding density. In other words, the mean particle separation limits the best possible spatial resolution for any field measurements technique involving tracers. The size of the interrogation window has no direct influence on spatial resolution because the displacement estimate obtained from the sampled region reflects an average of the displacement of all particle images in that region (i.e. moving average). The movement of the window over the image continuously changes the ensemble

of particle images from which the average displacements are computed. Its effect is that of a low-pass spatial filter; that is, wavelenghts of less than the window's size are increasingly suppressed. The spatial wavelength response of several windows is shown in Fig. 7. The step size of the sampling window should be limited to the average particle spacing. This yields a lower spatial wavelength cutoff of twice the step size known as the Nyquist sampling criterion. The above observations suggest that a smaller window increases the spatial wavelength bandwidth, or, alternatively, increases the upper spatial frequency cutoff. Although this is the case, a repetition of the experiments that led to Fig. 6 with a 16 by 16 point interrogation window showed an increase in the uncertainties by a factor of approximately two. Effectively, the dynamic range is reduced by factor of four, while the spatial wavelength range is only improved by a factor of two. The reason for the increased level of uncertainty is thought to be a direct results of the greater probability of particle images being truncated by the edges of the smaller window; that is, the ratio of truncated particle images to untruncated ones increases as the window becomes smaller. Also, the number of unpaired particle images versus particle image pairs increases for the same seeding density and displacement. The majority of the error between the true displacement and the measured displacement is a result of two factors. First, the estimation of the correlation peak through interpolation from three points in either direction introduces a certain amount of error. Second, the randomly spaced particle images within the sampling region can only yield displacement vectors for their respective locations. This means that especially if the seeding density is low, the average displacement vector obtained within the sampling region is only unbiased if there is no velocity gradient present, that is all particles move with the same velocity. In all other cases the error due to this phenomenon will increase and will be observed in the form of a broadened correlation peak. A higher seeding density, meaning the presence of more particles within the sampling region, reduces this error. Increased displacement gradients increase the biasing error. The phenomenon

C. E. Willert and M. Gharib: Digital particle image velocimetry is especially noticeable when derivative quantities like vorticity are computed from the displacement data. Another source of error is the fact that DPIV is not capable of taking the curvature of the particle trajectory between two successive images into account. 5 Experimental setup The DPIV technique was first implemented in a water facility because the seeding of the flow is less complicated. A low Reynolds number vortex ring (Re=U~,b~DTub./V~I,050) was chosen, not only for its predominately in-plane motion, but also for its unsteady formation period and low velocity (Vm,x~ 3.3 cm/s at the edge). The vortex ring was generated by a tube of 3.00 cm internal diameter which was coned on the outside to a sharp edge at 15 ~ to the center line. A constant pressure piston, controlled through a timing device, pushed fluid from the tube into a reservoir. Formation of the vortex ring described here took 1.83 s during which the piston traveled 5.7 cm. The seeding particles were neutrally buoyant 80 lam phosphorescent polymer spheres at a seeding density of 10-20 particles in a 32 by 32 pixel sampling window. Based on Eq. (6) the images formed by the particles on the imaging array have a radius a that is on the order of 1-1.5 pixels. Illumination was provided by a 5 W continuous wave Argon-ion laser whose beam was spread to a 1 mm thick sheet using a cylindrical lens. The laser was shuttered at a 30 Hz rate with a pulse duration of 4 ms to minimize streaking effects of faster moving particles and yet provide enough light for the recording of the images. Shuttering was necessary because similar to photographic film the imaging array of the camera integrates over the time period between frame capture (1/30 s). If available, a pulsed laser would have been used for its high energy content over short pulse durations. The image acquisition system consisted of a Hitachi C1H charge coupled device (CCD) color camera of which only the green signal was sampled at 8 bits per pixel by a Vicom image digitizer and processor (Fig. 8). Digitized images have an active area of 512 pixels horizontally by 480 pixels vertically, which, in the case of this experiment, corresponds to a field a view of approximately 8.3 cm by 7.8 cm, or a scaling factor of 6.13 pixels/mm. These images were written to a real-time digital video disk at a framing rate of 30 Hz. This disk is capable of recording 750 images, or 15 s worth, and can then be downloaded and archived on magnetic tape as conventional binary data files. In this experiment 300 images were stored, covering the initial roll-up of the vortex ring, its separation from the tube, and its laminar phase while traveling through the image field until it was 10 cm from the nozzle. 6 Velocity field measurement results All DPIV results shown here for the vortex ring experiment were obtained with the code described in Sect. 3. The samLaser
(5Watt Argon ion )

187

Chopper
~ ~ CyLindrical-

Light sheet and

seeded flow ~.~ ~-

Ch~;i:i:t
II
SUN

3/160 I I

i_,zlvideo

Fig. 8. Schematic of the illumination and imaging system for the DPIV image generation

512 480 448 416 384 352320288256224O n~ ;;iiii.. ;i;;:ZZZ ''*bhtlll

I'~llltttttt,,,, ................

: .... : : : : : : : : : : : > , ~ > > > > > > > ~ # J , , : .................... . . . . . . . . . . .----......,,....-~:.~....:::,..,~~:::::: ................ ...........

192

................... :: ~: : .: - ~ _ .~ ~ -~ ................... , . - 5 _ ..................... : : : : , . . ~

. . . . . . . . . . . . . .

<

................... .................... ... ........................ ~.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

160 128 96 64 32

.............. ................

,~IIIIItI/II___~N\\\\\\\\ ,,,tttll?tttt,__X\\\\\\~\~''"
,ttf[[ll[Ir~r

\\\~111~

. . . . . . . . . . . . . . . . . .

:: : : : : : : : : : :

::::::I:;~:::::;;;;

i ii i l

10 Pixets 64 128 192 256 320 CoLumns (Pixets) 384 448 512

Fig. 9. Sample, unprocessed data set of displacement vectors obtained by the DPIV technique; the nozzle center of the vortex generator is located to the right at row = 276, column = 0; the reference vector at the bottom indicates 10 pixels of displacement

pling window had a size of 32 by 32 pixels (N by M) and was overlapped 75% with the next, to obtain a grid of approximately 3,500 equally spaced (8 pixels or N/4 by M/4) displacement vectors for each image pair. Although this overlap does not make adjacent estimates truly independent from one another, they still reflect individual measurements because each sample incorporates a different set of particle images, given a sufficiently high seeding density. In analogy to time series analysis, the spatially shifted cross-correlation is similar to a moving average. Of the 3,500 measurements only 240 can be viewed as truly independent.

188
5.0cm/s > 3.270

Experiments in Fluids 10 (1991)

2.770
2.270 3 1.770

I -. -.:.:. . . ... ..... . . . . . . . . . . . ./ Z./ l / I l l . . z

/lzzzz,'

l~'~l/llttltt\\~\.

_,, $It1~ . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . .

.~

I
v

1.270 0.768 0.268 -0.232

iiii!i!!! -oi
. . . . . . . . . . . . .

!!!!!!!iiiiiiii!!!
- 0.732 -1.230
-

1.730

- 2.230 - 2.730
~\~\\\\\l]((fll/~N\\lil/ll ~\\\\\X XX\~'zleZ/ldlll/ t~k\\\\\~'~\tt+~/t//illltl,

-3.230 - 3.730 0 1 2 3 x(cm) L

I I I I

............
. . . . . . . . . . . . . . . .

5 ~-

Fig. 11. Isovelocity contours (velocity magnitudes) of flow field shown in Fig. 10
i

- 4

~J

A 5 6 7 8 xlcm) ;Fig. 10. Converted data set of Fig. 9 after the application of a 3 x 3 convolution low pass filtering operation to remove high frequency jitter caused by low particle density; time reference is + 2.24 s after the end of the formation perid; the nozzle center is located at origin (0,0); the reference vector at the top indicates a velocity of 5 cm/s

s12I
A80 . . . .

,t

i ' 84

:-

. . . .

AA8 f A16-:

",

"

"

'-

"

',',

, .... " _ ",, .. . . . .

",,, ". ""'--'-,

,',", '., ,\ ~

' ,' , ,

_'~'" J,,

,' "t

,',, '

," .' .'. ....

,'

"

"

Due to the high quality of the captured images, no image processing such as thresholding was required prior to the analysis. A typical set displacement d a t a is shown in Fig. 9 as a vector plot for t = 2 . 2 4 s after the formation period of the vortex ring. Close inspection of this d a t a indicates an a n o m a l o u s behavior in regions of high velocity gradients, i.e., toward the core of the vortex ring. Velocity gradients and rotational motions near the center of the core result in a loss or disortion of the cross-correlation peak. To circumvent this problem a smaller sampling window should be used. In this realization, however, the low seeding density of the original images did not permit this without a frequent loss of correlation due to a lack of particles in some areas. Velocity vectors in the data which had magnitudes and directions greatly in difference with their eight neighbors were considered misrepresentation of the actual flow and were removed interactively. In all d a t a analyzed at most 1 0 - 2 0 of the 3,500 vectors for each d a t a set had to be reinterpolated. Since the resulting velocity field has a grid spacing of 8 by 8 pixels (N/4 by M/4), but sampling occurred with a 32 by 32 pixels (N by M ) sampling windows, fluctuations at this lower spacing cannot be resolved. Therefore each set of 3 by 3 d a t a vectors constitutes a set of separate measurements of essentially the same region in the flow. Convolution filtering of the d a t a with a 3 by 3 kernel therefore is possible without loss of information. The kernel chosen here replaces the center d a t a vector with the average of the 3 by 3 d a t a vector set. This filtering operation acts as a low pass spatial filter

oo, 352: 320

288

-'-~'~-~--"-~ ~ " ~

,," ' - . .: .
::--

-" -.
-

" . .
"

--_.-~-~ :-_ . , ' - . ~ . ~ . .

- ---:-

~'~,-~.-'+--~ ~ s _ .

9~

256

"

'--

---

~ - ~ ' ~ ~ ' ~ "

"'-

"

':

. . . . . . . . 192 r ! " ,,"

~.--" ~---m,~.~-~--z

--'-

" . . . ..
.9

160 ~ 128-

, "

-->'> ~.~,~---,~.~ \~"~-,. .-. -" ' i .i //I~" ." ~ "&"~" ~,\ . ... " ,'/' '/" ~ _~-"~',~" t '.," , , ' ,s', , , ' ; , " - ~ " , , ", , ,', ,, , ' '~ .... ": , ,1. . . . ' ', ,

- , ,,/.

" , - .

32
0 0

"
6'A

: . : i:

'

. . . . .

10 P i x e t s 128 192 Col.umns 256 320 38/+ A 88 512

( Pixel.s )

Fig. 12. Displacement data retrieved for time step t = + 2.24s using an interactive particle matching technique; errors are of the order of 1/5-I/10 of a pixel; the reference vector at the bottom represents 10 pixels of displacement

which removes the high frequency jitter associated with the different location estimates of the peak correlation. A representation of data processed in this manner is given in Fig. 10 from which the p r o p a g a t i o n velocity of the vortex ring has been subtracted. Reinterpolation of d a t a vectors was not necessary in this realization. Figure 11 is the associated velocity magnitude contour plot.

C. E. Willert and M. Gharib: Digital particle image velocimetry

512 A80 /+A8 A16 38/+ 352 320 .................................................. 320"
........................................................

189

312 3O/., 296288280272in

::::::::::::::::::::::::::::::::::::::::: : :::.I';; ';"::: " : ..... !: ! : - : : i ! ' [ : ' ; ; ~ 5 ; ; i "'::::: ! 7 : : ' : : : : : : :: ................... ~-~-~7,,~-':;,.:.: . . . . . . . : : : : : : : : : : : : : : { T i T ~ : ' ( ! T i : { i i i : ! i i : { : i !7{7!!7L'.i'7)i77!77; ! ) 7 j 7 ! L.':: : : 7 : : : : : : : J:: : : : ; , : : : : : . . : : : ; , ; : : : ' t : : : : :::: ::;: .:.-:::.:
,

!i:!

:{ : .

288- : : : . : : 256

' :::

:11:::::

264256. 248. 240232224216\'..\ \"..\ t / i ..-~.--~.. -~

. . . . . . . . . . . . .

.................................. iiii:~! :!!!;!!!!i!!i!ii!!]i]i!

;::;:; :::~L-.;:::I i:.

224 ! ! L ; . i i i i : . : . ! ! ! ! ] ] i i i i i i ; : ! ! /
rid.
o

.~ 192 !~!!:!.::~ii i.!!7.:!!:.i!;!!iii.!!ii:!i: :ii: i!iii!:: !!!!!!!!:

i,j;;;i:;':~;~',;:.; :.-:~;~'.',k':~::

160 128 96
6/,

......................
....

............................

: ; ; : ; -; 3 .,'. .: :. .. :.: .- :. :. .. :;::

:::K, , , . .=,-:======== . . . . ====. :====,== . : : : : : : t . . . . . . . . . . . . . . . . . . . . :::: : : : : : : : : : ::: :

::::; :::;

;:'-::

::::::::;:::::::::

================================================

32 10 Pixets 0 ' f~Z. ' 189 ' 162 ' 256 320 Cotumns (Pixets) 384 ' 4Z.8 ' 512

208200192184 56

t tt

1,,

11/t,
>

_.-i.

0.1 Pi

7'* ' 88 ' lk'

1~0' 1A6' ~A2' 1s

lk

Oolumns (Pixets)

Fig. 13. Algebraic difference vectors obtained by subtracting an interpolated version of Fig. 12 from the data displayed in Fig. 9; the large vector at the bottom represents 10 pixels of displacement

Fig. 15. Enlargement of the stagnation flow zone between the vortex ring (right) and the nozzle (left) at time t = + 8.35 s; the large vector at the bottom represents a displacement of one tenth of a pixel

320 312 304 296 288 280 272

/ / / / / / / / / / / i

i
3.270
2.770

///'//'/~/'
~ / ~ / / / / / /

/////
/ / / / / I I
I 2.270 1.770 1.270 0.768 0.268.

~ i i 7 ~ / / / / / / / ////////~///i"

~x 26A.
256
o

.1,.1t

/.1.1/,

...%~.L..~//

248. 2A0232. 224 216208 200 ~' 1 / 2 Pixe[ ,

" ~ - 0.232 ~, - 0.732- 1.230-

..~..-~.A/...%..%A .7..~' . 7 . . 7 /

- 1.730

392

~;8' A2A '~0'~;8' '

A72 '~;8' '

504 '

- 2.230- 2.730-3.230- 3.730-

Cotumns(Pixets)

4 x(cm) '

5 i,

Fig. 14. Enlargement of far field flow at the beginning of the vortex ring rollup period (t= - 1 . 0 0 s); the large vector at the bottom represents 89 pixel of displacement

Fig. 16. Streamlines obtained by integration of the flow field of Fig. t0 using the stream function relations for axisymmetric flows; the difference between the contour levels is 0.25 cm2/s

7 Error analysis

Verfication of the displacement measurements proved to be the most difficult task, because the relatively large size of the seeding particles did not permit simultaneous laser D o p p l e r measurments during the single realization of the unsteady flow experiment. Rather, it was decided to use the images

themselves as a d a t a base for determining displacements of individual particles. Particle pairs were matched interactivel y on a frame by frame basis and their displacements could be measured to an accuracy of one fifth to one tenth of a pixel. This r a n d o m l y spaced d a t a (Fig. 12) was interpolated to a regularly spaced grid using a Gaussian weighting function (Agui and Jimenez 1987). The algebraic difference of the

190

Experiments in Fluids 10 (1991)

1
-1 -2 --

1.~

-1

*
2 4 6

--

c 0 f 0 Fig. 17a-f. Vorticity distribution over time calculated from the velocity fields by finite differences (vertical scale is compressed by a factor of 5)

vector fields obtained with each method is shown in Fig. 13. Near the core these difference vectors become larger in magnitude and r a n d o m in their orientation, indicating inaccurate velocity measurements of the D P I V method in these regions. The magnitude of the mean difference vector is 0.076 pixels and shows that there is essentially no systematic error in the form of a bias. A calculation of the standard deviation of these difference vectors yielded a magnitude of 0.263 pixels. This value reflects the level of uncertainty in the measurements and can be associated with the accuracy limitations of the interactive measurement technique. In Sect. 4 it was shown that the uncertainty of D P I V itself was much lower.

F o r the case of the vortex ring experiment, this limitation made a quantitative assessment of the true displacement measurement error of D P I V impossible. An illustration of the wide dynamic range capability of D P I V is given in Fig. 14. This, still unfiltered, vector plot was obtained from the far field of the vortex ring during its formation. The associated displacements are on the order of 1/201/10 of a pixel which corresponds to 0.2 to 0.5 mm/s. An enlargement of the stagnation flow zone between the vortex ring and the nozzle at t = + 8.35 s is shown in Fig. 15. M o tions here are more r a n d o m than in the previous figure but they do not necessarily reflect errors in the displacement mea-

C. E. Willert and M. Gharib: Digital particle image velocimetry surements because a flow of this type typically is three-dimensional. In this region the velocity ranges from 0.050.5 mm/s.
\

191
14

12
10

o
o O A A A

8 Fluid dynamical quantities obtained from the velocity data

The recovered velocity data can be used to numerically recover derivative quantities like vorticity, strain, and continuity, and integral quantities such as the stream function and circulation. Three of these, vorticity, circulation, and the stream function are described here. Because the vortex ring is axisymmetric, the stream function has to be integrated with respect to its axis of symmetry using the stream function equations for axisymmetric flow. The results of the path-independent calculation is shown in Fig. 16 and represents the average of a radial and axial numerical integration of the velocity field in Fig. 10. The vorticity distribution was computed using finite differences of four velocity terms in each spatial direction which is a second order approximation to the derivative quantity. Figure 17 shows the evolution of the vortical field from the inception of the vortex until its passage from the field of view. The first two plots (Fig. 17 a and b) clearly show the vorticity in the exiting boundary layer being carried into the vortex core. As the vortex ring propagates through the image field, the vorticity peaks shrink; wheras, the footprint of the high vorticity region increases as expected in a viscous flow. An indication of the associated decay rate is visible in Fig. 18, which is a plot of the absolute value of the peak vorticity for both cores over time. Peak vorticity is associated with the highest velocity gradients in the flow, and, since DPIV is a spatially averaging technique, the actual values of vorticity most likely are higher. Furthermore, the spread in the measurements of peak vorticity is a good indicator on the performance of DPIV in highly vortical regions. The vorticity distribution of Fig. 17 exhibits some "lumpiness", especially around the regions of high vorticity or regions that exhibit larger velocity gradients. These "lumps" are not typical of a laminar vortex ring, but rather are a direct result of the lower seeding density in some regions as opposed to others. This spatial lack of particles yields the same velocity displacement (velocity) vector over several overlapping sampling regions until other particle pairs are incorporated by the sampling windows. This effect results in a biased measurement of the real average displacement vector, and can only be resolved through a higher seeding density. The effect is amplified through three-dimensional motions. The relative height of these "lumps" may be used to obtain a measure on how seeding density affects the displacement measurement in regions with velocity gradients. Integration of the volume contained underneath the vorticity distribution yields circulation or by application of Stokes' Law this operation reduces to a line integral of the dot product of velocity and the incremental path element on

~8 u
>o
o s

~B
s

~o
&

6 4 2

,o

2 3 4 5 6 Time(s) Fig. 18. Absolute value of peak vorticity (in core center) over time while vortex cores are within field of view; core above symmetry axis (o), core below symmetry axis (A)

-2

-1

10-

6-

~ 2~

e
....

.o_ 0 . . . . . . , - - - - - ' - ' - - - - , - - - - - - ' ' - - - - ' - , - - - - ' - ' - - - ' - ' - - - " =

-~-2
~
. Q @

-1~

,Q 0

i
Time (s)

Fig. 19. Circulation obtained by separating the flow fields along the axis of symmetry and integrating around the areas' boundaries; upper half (o), lower half (n), entire imaged flow field (A)
a closed contour. In Fig. 19, the velocity data fields were separated along the axis of symmetry and circulation for each half, as well as the entire field, was obtained via integration around the perimeters. In the plot, the rightmost two and the leftmost three sets of data points cannot be considered values representative of the total vortex ring circulation because the recirculating region of the flow is only partially visible. The intermediate region exhibits a decay of circulation which is due to the viscous diffusion of vorticity. The maximum obtained circulation is 9.66 cm2/s which is within 10% of the value obtained by multiplying the square of the mean piston velocity by half of the formation time At: V2.be At Fest - ~ ~ 8.87 (7)

The discrepancy can be attributed to the fact that the velocity at the edge of the vortex tube is actually a little higher than the mean piston velocity. This observation was already reported by Didden (1979) and Glezer (1981). A measurement of more interest to current research in vortex rings is the distribution of circulation around the core

192 1098/ .../~ ./../

..if .i

Experiments in Fluids 10 (1991) of higher resolution imagers the spatial resolution currently obtained only by LSV and PIV will be achieved with D P I V as well. Based on the size of the interrogation area with respect to the image size, the spatial resolution of LSV and P I V are, on average, still a factor fo four times better than the D P I V system described here. Currently we are conducting experiments on the dynamics of the vortex ring with a higher particle seeding density which increases the spatial resolution. This enables vorticity measurements in the core region with better accuracy. The results of the current efforts will be reported in the future.

.. . . . ,,,,- / , / /'// ,,// /// /

-~-~--~-S 2 ~ : ~

~ 7~.
~6

~, s'6/..~3L5 2 . . . . . . . 1 .... /// -'7/. r .

//(/
,,, /
. . . . ........... t = * 0.L7 s ---t = + 2.50 s t = + A.SA S

'

.~sS , O0 0.2 O.A 0.6 (3.8 I'.0 1'.2 l'.A I'.6 I'.8 2.0 Radius(cm)
Fig. 20. Circulation obtained by circular integration around the core centers (points of highest vorticity) for 3 different time steps; straight line at F = 9.66 cmZ/s is the maximum measured circulation

Acknowledgements
This work was sponsored by the Defense Advanced Research Project Agency ACMP through the URI Program under contract no. DARPA/ONR N00014-86-K-0758. The authours would like to thank Mr. A. Weigand for his assistance in the vortex ring experiment.

centers, a sample of which is given in Fig. 20. Integration for the upper core was performed on circular paths a r o u n d the points of highest vorticity for three instances in the temporal evolution of the vortex ring. All three curves show the same trend with their slopes decreasing over time. The straight portion of he plots is very distinctive of the potential core of the vortex ring, which eventually smoothes to the asymptotic limit of circulation associated with either half of the core. The initial curved portion of the curves for radii below 0.2 cm can also be attributed to the smoothing effects of the D P I V technique, as well as its neglect of curvature effects. A true potential core exhibits linearity all the way to the center of the core. A detailed study of the core would require the core to cover more area in the image plane which is seeded at a higher density and sampled with a smaller interrogation window.

References
Adrian, R. J. 1984: Scattering particle characteristics and their effect on pulsed laser measurements of fluid flow: speckle velocimetry vs. particle velocimetry. Appl. Opt. 23, 1690-1691 Adrian, R. J. 1986a: Image shifting technique to resolve directional ambiguity in double-pulsed velocimetry. Appl. Opt. 25, 38553858 Adrian, R. J. 1986b: Multi-point optical measurements of simultaneous vectors in unsteady flow - a review. Int. J. Heat Fluid Flow 7, 127 145 Adrian, R. J. 1988: Invited lecture: imaging methods in experimental fluid velocimetry. Paper presented at the Americal Physical Society's 41st Annual Meeting of the Division of Fluid Dynamics, Buffalo, New York Agui, J. C.; Jimenez, J. 1987: On the performance of particle tracking. J. Fluid Mech. 185, 447-468 Barker, D. B.; Fourney, M. E. 1977: Measuring fluid velocities with speckle patterns. Opt. Lett. 1, 135-137 Blackwelder, R. F.; Utami, T.; Ueno, T. 1988: A novel technique for velocity field extraction from particulate visualization. Paper presented at the American Physical Society's 41st Annual Meeting of the Division of Fluid Dynamics, Buffalo, New York Browand, F. K.; Plocher, D. A. 1985: Image processing for sediment transport. In: Proc. Int. Association of Hydraulic Research, Melbourne, Australia Burch, J. M.; Tokarski, J. M. J. 1968: Production of multiple beam fringes from photographic scatters. Optica Acta 15, 101 111 Cho, Y.-C. 1989: Digital image velocimetry. Appl. Opt. 28, 740-748 Coupland, J. M.; Halliwell, N. A. 1988: Particle image velocimetry: rapid transparency analysis using optical correlation. Appl. Opt. 27, 1919-1921 Didden, N. 1979: On the formation of vortex rings: rolling-up and production of circulation. Z. Angew. Math. Phys. 30, 101-116 Glezer, A. 1981: An experimental study of a turbulent vortex ring. Ph.D. Thesis, GALCIT, California Institute of Technology, Pasadena, USA Grousson, R.; Mallick, S. 1977: Study of flow patterns in a fluid by scattered laser light. Appl. Opt. 16, 2334 2336 Kimura, 14 Takamori, T. 1986: Image processing of flow around a circular cylinder by using correlation technique. In: Flow visualization IV (ed. Veret, C.). pp. 221 226. Washington: Hemisphere Kompenhans, J.; Reichmuth, J. 1987: Two-dimenstional flow field measurements in wind tunnels by means of particle image re-

9 Conclusions and future work

Ease of implementation of the technique has p r o m p t e d the authors to fully implement D P I V in their l a b o r a t o r y complementing the single point laser D o p p l e r and hot wire capabilities already available. Currently the processing time for generating a d a t a set as in Fig. 9 is 2 h and 10 min on a S U N 3/160 16 M H z computer. This time can be reduced to the order of several minutes through the use of an array processor since the code itself is highly vectorized (i.e., repetitive). The implementation of an array processor was imminent at the time of the submittal of this article. The real strength of D P I V is the fact that images can be captured at real time, viewed, and then processed without delay. This allows for real-time modification of the seeding density, illumination and other experimental parameters. The current state of PIV and LSV does not permit this because of their reliance on p h o t o g r a p h i c methods. It should be further noted that the only real limitations of the technique are of purely a technological nature. With the arrival

C. E. Willert and M. Gharib: Digital particle image velocimetry locimetry. In: Proc. Int. Congr. Appl. Lasers and Electro-Optics, San Diego, USA Landreth, C. C.; Adrian, R. J. 1988: Electrooptical image shifting for particle image velocimetry. Appl. Opt. 27, 4216 4220 Lourenco, L. M.; Krothapalli, A. 1987: The role of photographic parameters in laser speckle or particle image displacement ve~ locimetry. Exp. Fluids 5, 29-32 Lourenco, L. M.; Krothapalli, A.; Buchlin, J. M.; Riethmuller, M. L. 1986: Noninvasive experimental technique for the measurement of unsteady velocity fields. AIAA J. 24, 1715-1717 Meynart, R. 1983: Speckle velocimetry study of vortex paring in a low-Re unexcited jet. Phys. Fluids 26, 2074-2079 Meynart, R, Lourenco, L. M., 1984: Laser speckle velocimetry in fluid dynamics applications. In: Flow visualization and digital image processing. Lecture Series 1984-03. Rhode Saint Genese, Belgium: von Karman Inst. for Fluid Dynamics Meynart, R.; Simpkins, P. G.; Dudderar, T. D. 1987: Speckle measurements of convection in a liquid cooled from above. J. Fluid Mech. 182, 235-254

193 Pratt, W. K. 1984: Digital image processing and analysis. In: Flow visualization and digital image processing. Lecture Series 198403. Rhode Saint Genese, Belgium: von Karman Inst. for Fluid Dynamics Rosenfeld, A.; Kak, A. C. 1976: Digital picture processing. Vol. 1 New York: Academic Press Simpkins, P. G.; Dudderar, T. D. 1978: Laser speckle measurement of transient B6nard convection. J. Fluid. Mech. 89, 665-671 Vogel, A.; Lauterborn, W. 1988: Time resolved particle image velocimetry used in the investigation of cavitation bubble dynamics. Appl. Opt. 27, 1869 1876 Zhan, Y. W. 1986: Optical measurement of velocity and acceleration with color coding. Optics and Lasers 18, 255-258

Received February 21, 1990