Here’s a partial answer to my own question:

The CMUcam2 can track colors at 50fps (a very respectable rate, more than low enough for musical timing) if the resolution is set low enough; meaning a vertical resolution of 65 (I assume this means horizontally 85). If the thing being tracked is, say, the musician himself, then the left, top and bottom-most extensions of the person in the field of view create a rectangular area, which is what is being sent as data and discrete zones are possible only in a limited manner. Although this is not a very good resolution for MIDI, it can be interpolated with [slide] to emulate a better resolution.

Frame differencing is possible at 26fps at a resolution of 8×8.

I could not find any information leading me to believe that there is any extra latency in the system; the baud rates available are more than enough to use for a good MIDI translation.

I’ve gotten used to precise definition of discrete zones and responses using 320×240 resolution with Jitter and various motion-tracking schemes, so it seems a shame to limit the possibilities, but it would be nice to have a stand-alone vision-sensor which can send out MIDI and leave the computers to do other tasks.