Three-dimensional Particle Tracking Velocimetry (3D-PTV, or in short PTV) is the experimental technique that allows for time-resolved fluid flow and/or additive phase (particles, bubbles, droplets, colloids: objects) measurements in three dimensions. Time resolution is the built-in feature since we record the motion at the speed that enables 'tracking of objects' in time. For tracking, we need time evolution, or simply object locations in time and space. In order to measure flow field we need multiple objects (we shall call them particles, that's why the system is Particle Tracking Velocimetry, though not all use particles as tracer objects). So, prerequisites:

a) multiple particles

b) space resolution and three-dimensional field of view

c) time resolution

If the flow is fast enough and the time and space (length) scales are small - our requirements turn out to be too demanding: have many particles and very fast time sampling. This is because we 'hit' the (small number) 'trackability parameter': the distances that particles do in a single time step should be smaller than the inter-particle distance. Very fast and complex turbulent flows, those with high Reynolds numbers, all have this problem. The solutions are:

1. measure only what you can measure - e.g. particle accelerations are easier to measure very fast since the only information needed is the single particle trajectory if the experiments are repeated to collect statistics. This is one of the things that Bodenschatz's group have accomplished in Cornell and MPI.

2. very small field of view, e.g. Zeiss et al. work

3. Acoustics or laser doppler measurements - not really multi-particle methods, but some properties could be achieved, e.g. the work in Lyon.

4. ETH Zurich - low to moderate Reynolds numbers, but well resolved in space and time.

Probably some other great examples are forgotten, so I apologize. But this is not an overview of the work done, this is an introduction to the main point of this article: why do we need real time image processing.

The obvious solution to the problem is to have very high speed digital cameras and very strong light. This is possible: there are cameras like Photron or Phantom that can do up to tens of thousands of images per second and there are solid state diode pumped lasers or halogen/xenon lamps that can illuminate the particles for this short exposure times. The bottleneck in this case is the on-board memory that those cameras have - the recording time is very short - tens or parts of the second. The other way of doing it is to stream the data to the external storage. This way is achieved so far at the ETH Zurich and Tel Aviv University: 500 frames per second camera of 1.3 Megapixels generates 625 Mb/sec data rate that is possible to sream using the Full CameraLink interface to the special framegrabbers (currently of IO Industries Inc. Canada, others are also possible, e.g. BitFlow) and to the hard drives organized in RAIDs. Thus, 12 hard drives can store in real time the images at 500 fps, for very long time. One single experiment of 20 minutes is not a problem and the post-processing becomes the next (solved, but) problem. The images are reconstructed off-line (taken out of RAIDs) and transferred to another hard drive. Then our 3D-PTV software (ETH Zurich, see http://ptvwiki.netcipia.net) processes those to extract the particle locations, match them between the cameras and generate 3D clouds that are tracked in time.

So, we propose to develop the Real Time Image Processing on-camera (or on-framegrabber) that does the following: instead of streaming and storing images (that are usually bright spots and dark background), we develop an image processing algorithms on hardware (based on FPGA) that removes the background, identifies the objects (blobs or spots) and stores only the relevant information to the hard-drive (simple, usual SATA-II or SCSI will do just fine). The rest is done on software. However, the streaming and storage needs go down by the factor of 10-1000 (depending on the sophistication of the algorithm and density of particles in images). At 500 fps, we need to stream about 5-50Mb/s instead of 650 Mb/sec. Just to note that 50Mb/s can pass by GigE interface (a simple internet cable and the NIC card that every laptop or PC have). So, the system is by far less bulky (see images on http://www.eng.tau.ac.il/efdl of the existing 500 fps system with 48 hard drives), quite fast and simpler.

But, we would like to have more: 500 fps is not enough for high Re number flows. It's enough to get Re_\lambda of 200 - 300. Some say it's enough to get the qualitative picture of turbulence, some doubt it. In any case, sometimes the mean flow is so high that we need faster imaging. Therefore, we want to develop some algorithms that will be able to work on any future imaging devices. The point is that in any case one doesn't need the original images, if TRACKING is concerned.