We ended up with two detection modes rather than one, because a still crowd and a moving crowd are genuinely different measurement problems.
Movement
Compare each frame to the one before it and measure how much changed.
It is simple, needs no reference, and survives a room whose lighting drifts slowly. It reads jumping, waving, clapping and dancing beautifully, and it responds the instant anything happens.
Its failure is absolute rather than graceful: a crowd that goes still reads as an empty room. Not “reads as low” — reads as nothing. Any game where the interesting state is holding a position is impossible in this mode.
Posture
Capture a reference frame of the empty room, then subtract it from the live picture. What remains is the shape of the people.
Because it measures a state rather than a change, it holds. A room that stands up and stays standing produces a high, stable reading for as long as they hold it. That is what makes stand-and-crouch work at all, and it is the mode most of what we want to build depends on.
The cost is fragility. The reference has to match the current conditions, which means the moment the lighting changes, the subtraction produces nonsense. It also requires capturing a reference before the room fills, which is a real operational constraint on a show day.
Why not just use posture for everything
Because posture cannot tell the difference between a room standing still and a room going absolutely berserk while occupying the same silhouette. All the energy is invisible to it.
A game about enthusiasm wants movement. A game about position wants posture. Trying to build one mode that does both produces something mediocre at each — we tried, and the compromise was worse than either.
The operational consequence
Each title has to declare which mode it needs, and that declaration changes the setup. Movement titles are close to plug-in-and-go. Posture titles need an empty-room capture, a proper calibration with the actual audience, and a lighting state that does not move.
We would rather be explicit about that per title than pretend the setup is uniform.