The test set-up for the measurements

Several of such set-ups have been used to determine accuracies in various directions and to analyse the effect of various variables. In this case we especially want to study the difference in results in the X and Y direction.

The number of pixels (in X direction) of the CCD sensor do not play a very important role in this story. If the number of pixels is less, the bandwidth is (useally) limited too. So the number of pixels available for reconstruction, mainly depends on the clockrate of the framegrabber. In principle one can say; "The more, the better", since the accuracy of the reconstruction is a function of the root of the amount of data available. There are systemetical errors, where a large amount of data does not help however.

There is a report from Gerald Dicker (NIKHEF), describing the influence of air turbulence and temperature gradients. You can get the .PDF (340 kb).

• The video bandwidth differs from one system to an other (systematic).
• The frame grabber locks its own pixelclock to the horizontal syncpulses. The quality of PLL in the grabber determines the stability of this clock from one line to the next. This affects the fit quality and the stability from one picture to the next (statistical).
• Settings of the grabber:
• Is the synclevel properly adjusted to the video signal to be digitised (statistical).
• Which part of each video line is digitised? There is no definition of which part of the active video line should be visible or, in this case, should be digitised (systematical).

Here you can find the CCIR definition (44 kbyte postscript). The standard also prescribes a gamma correction of 0.5. This means that when the illumination doubles, the electrical signal increases with the square root of two. Some (more expensive) cameras offer the option to switch the gamma correction off.

Here you can read more about how the gamma correction affects the accuracy.

Neither the start of the digitising, nor the stop are fixed by any definition. So we need a reference in the digitised data, independent of the video signals. By using a high quality frame grabber pictures were taken, starting before the active pixels on the CCD sensor untill the black area at the end of the line.

The black pixels at the start are used by the camera itself to dynamically determine the electrical black level. We can use them to determine the start (and end) of the active pixels. The mechanical size of the active area of the CCD sensor is known, so now we know the scale of the pixture, independent of the various clocks involved. We also know where the image of the projected mask is in respect to the CCD sensor. This has been checked by disturbing the video signal by means of a capacitive load on the 75 ohm line. On the monitor, as well as in the digitised data, the mask seemes to shift to the right. But also the black edges shift along with the picture. So the picture can still be accurately reconstructed in respect to the CCD edges. The picture below shows the start of the video line. One can clearly see the transition from intrinsic black to a white spot in the projected mask.

Below the end of a video line is shown, where the white is followed by an intrinsic black pixel. Depending on the camera, one might not actually 'see' black pixels, but a clamped video signal. Since this is locked to the internal camera clock, this makes little difference for the principle.

The black and white transitions of the projected mask stretch out over several CCD pixels. Once differentiated, this gives a nice gaussian curve, which can be fitted very well. The transitions, introduced by the intrinsic black pixels, are much steeper however. This is shown here, with a timescale of 20 ns/div..

The risetime of the transition is app. 20 ns. The clock of the framegrabber is in the range of 14 MHz to 20 MHz maximum. So samples are taken every 50 ns, at the best. Therefor the CCD's black edge cannot be determined very accurately. One needs to know to what value the shape rises, by looking at the next pixel. Knowing the shape of the transition also helps. As a rule of thumb we say that the CCD sensor edge can be determined with an accuracy of app. one third of a pixel.

Here the fact that several hundreds of these transitions are available in every picture does not help much. There is too little noise in these samples to improve the accuracy by means of statistics. 

But there are solutions.

You can mail  suggestions for (or questions about) this page to: Henk Groenstege.

august 1995