Event people talk about a room’s energy as though it were a mood. If you are trying to measure it with a camera, it is more useful to treat it as a signal — something with a level, a rate of change, and a noise floor.
Level versus change
Two rooms can look identical in a photograph and be completely different to work with.
A room that is holding something — standing, leaning forward, hands up — has a high level and almost no change. A room that is doing something — clapping, jumping, waving — has enormous change and no stable level.
This distinction turns out to be the central engineering problem. If you measure frame-to-frame difference, you see the second room brilliantly and the first room not at all: a crowd holding a pose reads as an empty room. If you measure the shape of the room against a reference, you see the first room but the second one becomes a blur.
You need both, and you need to know which one the moment calls for.
The noise floor
A crowd is never still. People shift, scratch, check phones, lean to talk. That baseline fidget is your noise floor, and it is higher than intuition suggests — a “motionless” audience of four hundred generates a surprising amount of frame change.
Anything you want to detect has to clear that floor. Which means subtle gestures are out. The things that work are the things that are unambiguously bigger than fidget: standing, crouching, arms fully up, moving across a space.
Latency and honesty
The most important property is not accuracy — it is immediacy.
A person who stands up and sees the screen respond within a frame or two believes the system is reading them. The same person, with half a second of lag, does not, and will start testing it rather than playing with it. Once a room decides the thing on screen is not really listening, you cannot win them back in the time available.
We would take a noisy, honest signal over a smooth, late one every time. Smoothing is very tempting and it is where most of the credibility gets lost.