Lorem says:

Well, from a purely theoretical point of view, split the visible crowds into the dense sections and the tails to hopefully remove some of the density variation. Then, figure out the areas of each of the two categories, and then apply the typical technique for such things (divide area occupied into small sections, determine an average number of people in each section, and multiply by the number of sections occupied, while using the appropriate distributions for the estimates).
Of course, even just figuring out the areas would be rather labour-intensive and if the crowd density varies too much, the estimate could still prove imprecise. Taking samples from their heads and tail-ends (all the crowds seem vaguely teardrop-shaped) may help a bit.