Jez:
8051 processor or 100 MHz DSP with 32 bit floats and/or 32bit integer MAC?

It would be possible to do this. There would be a transform for the pixels. But after the transform to the flat image there would be a need to fill in missing pixels because the spherical view will not be 1::1 to pixel adjacency in the flat view. One way to evaluate the required transform would be to take pictures of a field of a large white background at depth of field that would be typical for your application. On the background you would draw a grid of black lines. If your field size was say 10'x12' you could use a new white painters drop cloth onto which you drew black magic marker lines about 3 -> 6 inches apart. Then you could take the picture frame with the camera that is looking at the ball at the painters cloth. Using a search algorithm during a calibration phase you could locate each line intersection in the ccd image and store these coordinates in a table. Then at run time you can relocate each pixel from the CCD image to the corresponding flat bitmap image using an interpolation technigue called "4 point interpolation". The equations for this can actually be done in fixed point math and the techniques would be almost analogous to the process used in correction tables for laser scan head galvos.

After you remap all the pixels then you would have to interpolate the actual pixel colors to fill in the resulting flat image where there were missing pixels. (You may need to also smooth and level the colors where a lot of moved pixes crowd too close together. This part could be done as you translate the pixels to the flat. You would simply watch to see if a new pixel location was already occupied and if so average it into the flat picture including interpolation with its neighbors.

Michael Karas