1. get a photo, the higher the better
2. take a sample of the density. Basicaly, count heads for a small area of the photo.
3. extrapolate: how many similarly sized areas are there here? multiply the number of sample spaces by the count of the sample. Presto, a rough estimate of the crowd size.
Wired's Dangerroom discusses the science and basically agrees with me, but they add one step I never considered, and come up with a maximum value for a given space.
Basically, a person comfortably takes 5 square feet of space, and can be packed into half that space if they really really need to. I imagine Brisbane's Southbank during RiverFire, and I can see 2.5 square feet as quite a luxury.
Anyway, that's the science of it: Count, extrapolate, and crush the bastards with hard facts using Google satellite views. I knew most of that already, did you?