Camera placement for crowd-controlled games: height, angle, and the 28% problem
Camera placement is the single biggest determinant of whether a camera-driven crowd game works. It matters more than the camera, more than the lens, and considerably more than the software. This post is the specifics, including the position that gave us the most trouble.
The short version
- High, looking slightly down at the audience
- Far enough back to see the whole seating area
- Reasonably level with the crowd rather than steeply angled
- Locked exposure, always
- Analysis cropped to where people actually are
Why height matters more than anything else
A posture-based crowd game works by measuring the difference between two states — say, crouching and standing. Everything downstream depends on how far apart those two states look to the camera.
From a high mount, a standing person occupies noticeably more of the frame than a crouching one, and the top of their silhouette moves a long way. The two states are clearly separable, so the software has a wide range to map the game onto, and small wobbles stay small.
From a low mount near the front row, the geometry works against you. Rows occlude each other. A crouching person is partly hidden behind the row in front, and a standing person is partly hidden behind the row in front too. The change in visible area between the two is far smaller than intuition suggests.
The 28% problem
We hit this properly during testing, and it is worth publishing because you would otherwise discover it yourself, at a worse moment.
On a low, close camera, we measured crouch at around 27% and stand at around 55% of the analysed range. That is a working window of roughly 28 percentage points to represent the entire difference between the room’s two extremes.
The consequence is not that it fails. It is worse than that: it works, badly. Calibration becomes extremely touchy. A small, involuntary shift in the crowd — people settling, someone shuffling — swings the game across a large part of its range. The room feels like the game is twitchy and unresponsive to what they are actually doing, because it is.
And no filter fixes it. Smoothing a narrow signal gives you a narrow, smooth signal. The information was never captured. The fix is physical: get the camera up, get it more level, and crop the analysis to the seating area so empty floor and ceiling are not diluting the measurement.
There is one more improvement worth knowing about, which is a software matter rather than a mounting one: weighting the top of the silhouette rather than the overall centroid. Crouching changes head height dramatically and total body area only modestly, so tracking the top edge recovers much of the range that a centroid-based measure throws away.
Lock the exposure. Every time.
Camera-based crowd reading assumes the only thing changing in the picture is the audience. Auto-exposure breaks that assumption completely.
When the lighting desk pushes a look, the frame brightness changes. In movement mode, the software sees a large frame-to-frame change and reads it as the entire room moving at once. In posture mode, the empty-room reference no longer matches the current exposure, and the subtraction produces nonsense.
The practical result is that your lighting operator is playing your game, and neither of you knows it. Lock exposure, white balance and focus before calibrating. If the camera cannot lock exposure, use a different camera.
Light on the audience
Obvious once said, routinely forgotten in planning: a camera cannot see a crowd in the dark.
Plenty of shows put everything into the stage and leave the audience in near-blackout. That is a lighting design decision that quietly removes the possibility of a camera-driven crowd game, so it needs to be raised with the LD early rather than discovered at focus. You do not need much — enough for shapes to be distinguishable from the background is enough — but you need it, and you need it to be stable for the duration.
Rake, balconies and awkward rooms
A raked auditorium is the easy case: everyone is visible, occlusion is minimal, and a mount at the back of the room at head height or above sees the whole audience cleanly.
A flat floor is harder, because rows hide each other. Get higher than you think you need to.
Balconies are the case people forget. If a third of your audience is on a level the camera cannot see, that third cannot play, and — more importantly — their non-participation drags the measured average down in a way that makes the game feel unresponsive to the people who are playing. Either frame both levels, or knowingly analyse only one and tell the room which.
Test in the actual room
Every point above is a generalisation about geometry, and your room is a specific instance of it. Twenty minutes with the actual camera in the actual position, with the actual lighting state, will tell you more than any amount of planning.
The setup docs cover the calibration sequence, and how it works explains the two detection modes and when each applies.