Essays & Theses

A Vision-Based Particle Tracking Velocimetry

A Vision-Based Particle Tracking Velocimetry
of 14
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  A Vision-Based Particle TrackingVelocimetry P article Image Velocimetry (PIV) is a non-intrusive optical technique to measure velocity of flows. It provides the simultaneous visualization of the streamline pattern in unsteady flowsand the quantification of the velocity field over the image plane. To reveal the flow motion,the flow is seeded by small scattering particles. The instantaneous fluid velocities are evaluated byrecording the images of tracers, suspended in the fluid and traversing a light sheet. A PIV systemconsists of seeding particles, illumination unit, image acquisition system, and a computer forimage processing. For industrial applications a classical PIV system is not suitable for its cost,sizes and needs of specialised users and work areas. Moreover, classical PIV are unable to work inreal-time for the huge amount of data and expensive algorithms adopted. In this paper, a studyand the implementation of a new PIV system is described and compared against classical PIVsolutions. The solution proposed is capable of working in real-time and is a Continuous PIV,CPIV, system: with respect to classical PIV, it is composed of a continuous laser light source and aCCD camera. A specifically new image-processing algorithm for velocity estimation andrecognition of correct traces has been developed. It is based on the grey level distribution in theparticle trace image, and indicates those particles moving with an out-of-plane velocity vectorcomponent and provides the measure with a limited error. # 2001 Academic Press A :  Baldassarre 1 , M :  De Lucia 1 , P :  Nesi 2 and F :  Rossi 1 1 Department of Energy Engineering (DEF), University of Florence, ItalyE-mail: 2 Department of Systems and Informatics (DSI), University of Florence, ItalyE-mail:, Introduction The algorithms for motion estimation are relevant forseveral image-based systems. The moving objects/entitiesin the observed scene project their 3D velocity on theimage plane (i.e. apparent velocity). This is usually called‘‘velocity field’’ or ‘‘motion field’’ [1–3]. Most of thetechniques adopted for evaluating velocity fields considerchanges in the image brightness features/areas. For thisreason, the fields obtained are normally called ‘‘opticalflow’’ or ‘‘image flow’’ field. Generally, optical flow fielddiffers from the velocity field, which is a pure geometricconcept. The latter is equal to the optical flow field onlyunder specific physical conditions of illumination,reflectance, moving object texture, etc. [2]. The opticalflow precision with respect to the true velocity field alsodepends on the estimation technique adopted [4]. On theother hand, the estimations of an approximated velocityfield can be enough for many applications.Motion estimation is very useful in important areas:motion-compensated image sequence processing,dynamic scene analysis, and for measuring and trackingsystem [3,5–7]. 1077-2014/01/040145+14 $35.00/0  #  2001 Academic Press Real-Time Imaging  7 , 145–158 (2001)doi:10.1006/rtim.1999.0203, available online at on  Most of the above mentioned applications need toconsume optical flow fields in real-time. Motion estima-tion must be performed in real-time especially for imagesequence coding and robot vision. The implementationof a parallel architecture and/or specific VLSI chip is atechnique to achieve this goal. In the literature, thereexists some motion estimation architectures for motionanalysis in real-time e.g. [8–11].In the literature, several main approaches for motionestimation in a regular grid of the image can beidentified: spatio-temporal filtering-based, block-match-ing, pel-recursive, gradient-based, corner tracking, andline tracking approaches [3,11]. Only some of these approaches can be profitably used for motion estimationof fluid flow, thus for the building of vision-basedvelocimetries instead of using pressure-based tools thatare intrusive. Typically, images of fluid flow do notpresent enough pattern information to estimate theoptical flow with the above mentioned approaches. Forthis reason, seeding particles are typically inserted in thefluid flow. Therefore, the problem is reduced toestimating the velocity of small particles in a sequenceof images or even in single images with long or multipleexposures. For this purpose, corner tracking, pointtracking and line tracking produce more suitablesolutions than gradient based algorithms. More speci-fically, both of these approaches can be used in differentcases, as it will be discussed in the following.Specific applications for velocity estimation of fluidflow by using seeding particles and image processing arecalled Particle Image Velocimetry (PIV). During the last10 years, PIV approaches have received a rapidevolution. They are typically based on non-intrusiveoptical measurement methods to capture flow velocityfields and [12–16]. PIV provides the simultaneousvisualisation of the 2D streamline pattern in unsteadyflows, as well as the quantification of the velocity fieldover an entire plane. With the improvement of imageacquisition (optics and CCD technology) and light andprocessing systems, several kinds of PIV systems werestudied and proposed [12,17]. Different techniques have been defined to study fluidsin different regimes. In the last few years, a particularPIV system has been widely used, to study flows in manydifferent conditions. Hereafter, this is called the  classical PIV   system. Here, a pulsed laser is used as a lightsource, while small seeding particles, and a CCD cameraor a photographic film are used such as a recordingsystem for the classical PIV system [18]. A common lightsource is a Q-switched Nd:Yag laser with a doubledfrequency, so that it is capable of producing successivepulses with a very short time lag between them. Thisapproach is obtained by using cylindrical lenses toexpand the highly collimated laser beam, thus producingan intense light sheet down to a thickness of tens of microns. In a PIV system, the illumination pulses needto be synchronized with the image exposure. This isusually achieved by programmable pulse delay genera-tors that can either stand alone or be hosted by thecomputer performing the image acquisition and analy-sis. In classical PIV measures, it is possible to record oneor two successive images of the small particle tracers.With non-standard cameras, especially developed forPIV, it is possible to acquire two images, each of whichis synchronized with one laser pulse. The duration of theQ-switched Nd:Yag pulse is only a few nanoseconds(typically 10ns or less). Thus, the images obtained showone illuminated spot due to the scattering particle, intwo different positions in the two successive images.Knowing the time lag between the two light pulses andmeasuring the distance between the two positions allowsthe evaluation of the velocity component in theilluminated plane.The detection of the sign of particle velocity vectors isa classical problem of PIV analysis [19,20]. In fact, mostof the proposed solutions are only capable of evaluatingthe magnitude and direction line of the velocity vector,without the sign. To solve this problem, many differentsolutions were proposed for the classical PIVs. One of the most widely used methods is based on theintroduction of a known and fixed shift between thetwo images recorded corresponding to the two flash-lights. This method is typically implemented by usingrotating mirrors, [21], and was studied in severaldifferent conditions by many authors [22,23]. Thissolution can be used with previous systems, by usingthe camera to grab the images coming from the mirrorinstead of directly from the measurement area. Thedisadvantages of this technique are due to the complex-ity of the system alignment and to the configuration.Other techniques that have been defined and implemen-ted to solve the sign ambiguity problem [24] are: electro-optical image shifting [25], multi-color systems [26,27], etc. Electro-optical image shifting methods employdifferently polarized light as an illumination source byway of appropriate bi-refringent crystals this simplifiesthe identification of starting and ending points in thesame image. Multi-color systems work by using two ormore light sources at different frequencies. Imagescoming from the same particle can be ordered in time 146 A.B ALDASSARRE ETAL.  by analysing their color and knowing the light frequencysequence.In a PIV system, the processing methods used to studyvelocity vectors are very important. These can becharacterized on the basis of the seeding density andof the image acquisition method (single frame ormultiple frames). Typically, on the basis of the seedingparticle density, three different partially overlappedcategories can be identified. These are (i) low density,(ii) medium density, and (iii) high density/smokes. (i) Low density In this case, the techniques for direct analysis are themostly used. With these methods, a scanning analysis of the single frame is provided in order to find the singleparticle properties. Then, these properties are correlatedbetween two subsequent particle images [28] (see Fig. 1). Therefore, the problem is shifted to match couples of points due to the two laser pulses e.g. algorithms forparticle tracking discussed above. In these algorithms, itis usually required to remove erroneous velocity vectorsthat can be due to insufficient particle pairs, or for thepresence of so-called  out-of-plane particles  or  motions (when the particles come/go from/out of the imageplane). This is due to the fact that a velocity componentin the orthogonal direction to the light sheet can bringthe particles seeding away from the illuminated area,during one of the two flashes. This problem cannot beavoided since the seeding particles arrive withoutsynchronization with respect to laser pulses. Thus, byonly having the intensity of the particle image, it isimpossible to study a misalignment of this kind of motion for the particles. This means that the particles onthe image, that are due to out-of-planes, do not havetheir counterpart and thus the velocity algorithm canerroneously couple them, creating and estimating wrongvelocity vectors.A variation of the light scattered from the singleparticle can be due to a different size of the particle, asthe G. Mie scattering theory demonstrates. Thisproblem can, from one hand, create confusion andfrom the other can be used for matching the sameparticle in successive images. (ii) Medium density In this case, statistical methods are typically used suchas those based on correlation of the image gray levelmatrixes [28,29]. The images are scanned with rectan-gular windows, and in each window a uniform motionflow is supposed. Thus, the algorithms work on a localarea of the flow image to produce a measure of theaverage local displacement of region particles. Whenone image showing the same scattering particles in twodifferent positions is acquired, autocorrelation is used.Studying the autocorrelation matrix, it is possible tolocate two symmetric peaks, corresponding to the shiftvector in the single window. The symmetry does notallow for recognizing the sign of the particle shift. Inmultiple frame acquisition systems, the cross-correlationis determined for two subsequent images. In this case,the same local areas located in two successive images aretaken (supposing the motion of particles is limited).Only one peak is present in the cross-correlation matrixand it represents the whole shift (module, sign anddirection). Correlation-based processing is a quiterobust technique, but it is time consuming and requiresanalysis of the image by window, thus reducing thespatial resolution of the system. (iii) High density Here, two images are typically considered. Thus,algorithms for optical flow estimation can be profitablewhen applied [1,4,9,11], such as the interference techniques [28]. These methods are based on opticalsystems producing interference patterns from a negativeimage of two equally spaced particle images. Theseparation between two adjacent fringes is proportionalto the displacement of the single particle. With thistechnique, it is possible to estimate the velocity with highprecision, but the needs of a particular optical arrange-ment for visualising and analysing the interferencepattern do not allow real-time analysis.In addition to these techniques, there are thedifferential ones that are useful for flow motion analysiswith very high density seeding or smokes. Differentialtechniques are based on several constraints, whose Figure 1.  Two images related to the same particle taken atdifferent time instants. AV ISION -B ASED P ARTICLE T RACKING V ELOCIMETRY  147  unknown quantities are perspective projection compo-nents of the motion field on the image plane, forexample, the Optical Flow Constraint (OFC), whosecoefficients are partial derivatives of the brightness of the image plane [2,4]. The work described in this paper is mainly focused onmedium and low density seeding. The goal of thedescribed work has been to develop a low-cost, real-time, easy to use, safe and flexible solution for themassive measuring of velocity flow in centrifugalcompressors of Nuovo Pignone GE. They perform themeasures by using Laser-Doppler Velocimetry, LDV. The Proposed Solution In recent years, the need for compact and low costsystems to study the gas flow velocity inside turboma-chinery has become greater. The most importantfeatures that these measuring systems have to provideare; high data rate, computerized acquisition proce-dures, bi-dimensional analysis, realtime processing,small size, user-friendliness, and low cost. Thesecharacteristics have been considered in order to builda new system to study gas flow velocity insideturbomachinery. It is based on an image processingapproach and has proved to be cheaper than otherimage- and non-image-based solutions for the adoptionof a PC-based, low-cost laser solution.Other advantages of the employed method, regardingthe measurements at low velocity, are:(i) the availability of simple and rapid image processingalgorithms;(ii) the possibility of detecting the out-of-planeconditions for the particle seeding, and thusomitting these from the velocity estimation;(iii) the capability of determining the velocity vectorsign.In order to satisfy the above points, the classicalapproach for implementing PIV system has beenrevised. In fact, in a classical system there is aQ-switched Nd:Yag laser as a light source, with a highenergy pulse, a non-standard CCD camera and a PC forgrabbing images and image processing. The mainlimitation, of adopting these systems for industrialapplication (for flow studies at a velocity in the rangeof 100m/s), is in the use of high power pulsed lasers(needing specialized and qualified staff in controlledareas). These typically have a large size and high costs.The solution proposed is composed of (i) a  cw  lowpower laser, (ii) a standard high resolution b/w CCDcamera with a high-speed electronic shutter, and (iii) aframe grabber for the image acquisition and post-processing of the PIV images. A specific new algorithmfor the processing of the images obtained with thisacquisition system has been developed. The algorithm isvery simple and computationally cheaper with respect tothe above mentioned cross or auto-correlation techni-ques. This method is based on the fact that it is possibleto use (as a light source) a continuous wave laser insteadof pulsed lasers. With this approach, the movingparticles produce a continuous line on the image plane.The trace length depends on the exposure time of theCCD camera. We have called this solution ‘‘ContinuousPIV, CPIV.’’The solution is illustrated in Figure 2. It should benoted that the earlier PIV systems were based on theanalysis of images produced with continuous lightsources (streak photography technique). In the past,these techniques did not achieve great success because of their poor accuracy in the measurements based on anestimation of the length of the traces on the image [17].Presently, the image acquisition systems and imageanalysis algorithms have been greatly improved as it isshown in this paper. The main problems faced for thiskind of solution in the past were due to the lack of (i)sufficiently high velocity TV cameras, (ii) specific imageprocessing algorithms for detecting the sign (the inverseof the particles velocity), and (iii) real-time imageprocessing solutions for measuring the velocity.The main limit of this approach continues to be themaximum value of the velocity that can be measured. Itis lower than can be performed by using other moreexpensive systems, such as those described above. Infact, with the increment of velocity, two facts limit thesystem performance. Firstly, at higher velocity theexposure time has to be reduced in order to maintainthe particle inside the measurement area. The maximumlength for single traces depends on the exposure timethat is inversely proportional to the maximum measur-able velocity. Moreover, at higher velocity, a lowerbrightness of the particle trace on the image is obtained.In fact, a particle moving at higher velocity remains infront of each CCD sensor element (i.e. confidentially thepixel) for a shorter time, scattering towards the CCDsensor with less photons. This can be balanced by using 148 A.B ALDASSARRE ETAL.  a higher power laser, knowing that the adoption of veryhigh energy is only possible with a pulsed laser.With a simple CPIV system, is not possible to solvethe sign of the velocity vector. This is not due to the useof a  cw  light source, but to the particular imagerecording technique (single frame). In order to removethe directional ambiguity, two CCD cameras (withdifferent exposure times) have to be used simultaneouslyto acquire the images of the tracing particles. With thissolution, in the grabbed images, each particle produces atrace of differing length for each image. If the exposuretime of the cameras starts at the same time and finishesin a different time instant, the corresponding traces startat the same coordinates in both images, but they stop attwo different coordinates of the image plane. Theposition of the trace ending gives the sign of the velocityvector.In order to analyse the continuous traces of theparticles (see Fig. 3), a particular post-processingalgorithm was defined and implemented. With a  cw light source, it is possible to analyse the grey leveldistribution of the trace image, instead of studying theintensity of the single point particle image. Thealgorithm developed can be considered as a techniquebased on direct analysis of the images (see introduction).It works by scanning the image in order to find singlepixels belonging to a trace, and then by measuring thetrace length (in pixels). The most important differencebetween this technique and the other direct analysistechniques is that here, it is not necessary to find similarparticle images to reconstruct the trace, as in the case of a classical pulsed PIV. Beside this, the processing timecan be reduced, scanning the traces on a grid. Thealgorithm developed presents three steps: (a) Thedetection of traces, (b) The extraction and the measureof each trace, and (c) The classification of each trace inorder to avoid erroneous measures. Figure 2.  (a) Experimental set-up of the CPIV solution (b) Particle trace negative image, as recorded by a CPIV system. Figure 3.  A negative CPIV image showing some particletraces. AV ISION -B ASED P ARTICLE T RACKING V ELOCIMETRY  149
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks