Hi Vigs,
Sounds more like a question, than a suggestion, so I moved it.
Eight years ago I did this little example here, my English was pretty rusty at that time, sorry about this.
http://www.cineversity.com/vidplaytut/mograph_step_crowd_scenes
I would suggest to use a similar set up, but perhaps with Polygon-objects (e.g., plane) and with an image sequence as main information.
It might be simple to photograph people in different (sequential) poses with a greenscreen and get the alfa channels in that way. Shot the images from several sides, keep the light direction in mind. For larger scene the Cloner>Sort/Effector Modify clone comes to mind. For large scenes I would try to keep the CPU workload low, so Blend as an option seems pretty expensive (CPU wise) for thousands of figures. Lots of options, but each project has perhaps a specific requirement. Share more about your target and we can get closer.
The random effector might help here a lot. Or the standard Particle Emitter in the Falloff of an Effector. Let me know if you like to have examples/sketches.
There was a presentation at (?) Siggraph/NAB a while ago where I saw that idea as well, but I can’t find it. (Search terms Stadium, Crowd, etc)
So, let me share one more idea, perhaps you can define the camera position early on, so the crowd in the distance could be done by rendering out a small crowd and projecting it on a plane. Etc.
If you have a sketch or more detailed questions, let me know.
All the best
Sassi