I suspect the real problem is just that those images are so noisy - in the demo with two people moving around, for instance, it's almost impossible to tell at t

I suspect the real problem is just that those images are so noisy - in the demo with two people moving around, for instance, it's almost impossible to tell at times whether there's one, two, or more blobs. Certainly it wouldn't be very reliable from a single frame, you'd have to use the movement information. Personally, I wouldn't waste my time on that, that's a rather straightforward image processing problem that could be made much easier by improving the output quality of the machine, which I imagine would be necessary anyways for this to see any real use in the field.