Crowd Size Report: Super Concert, Ulaanbaatar

The short answer

≈18,000people on the lawn at one moment of the show, 0:20 in the drone video.
Most likely between 14,700 and 21,900 (80% range).

Why we believe it

The lawn size is measured, not guessed. We traced it on two satellite images and on the video: about 15,200 m². About 13,900 m² of it could hold a standing crowd.
The crowd density is measured where the video is clear. In a straight-down shot of the front pens we checked 572 heads one by one, then measured the whole pens: 3.3 people per m².
The wide shot shows where the crowd is. An AI counter marks the heads it can see. We use those marks for the shape of the crowd and the measured densities for its size.

What this number is not. It is the crowd standing on the lawn at one moment. It does not include people in the VIP stands, at the bars and toilets, in queues, or people who came and left at other times. Total attendance for the evening is therefore higher than this number.

What we had to work with

We received one video: the organizer's edited 4K drone promo, 4 minutes 8 seconds long. We did not have the raw drone files, the flight data, ticket counts or gate scans. This report is based only on what the video and public satellite images show.

That matters for two reasons. First, a promo is edited: shots are short, the camera moves, and most shots are at night. Second, without flight data we had to work out the camera's height and angle ourselves, from objects of known position such as the stage, the runway and the lawn edges.

We used 44 still frames and three continuous sequences, 242 frames in total. Six of them carry the result:

Video time	What it shows	What we used it for
0:20	The whole lawn from the south, at night	Where the crowd is
1:41	Straight down onto the two front pens	Measured density, zone A
0:42	Steep, well-lit view behind the T-shaped stage extension	Measured density, zone D
2:52 and 2:56	Close, low views of the crowd around the sound desk	Estimated density, zone B
0:27.5	The whole field from the south-west	Found a near-empty strip on the west side

Why you cannot simply count the picture

The obvious approach is to point an AI people-counter at the widest shot and read off the number. We did that with eight different counters. They found between 400 and 3,000 people. The reason is simple: at night, from 150 metres away, most heads are a few dark pixels. The counter only finds the heads that catch light.

The clearest proof is the front pens. The video shows them twice: close, straight down at 1:41, and far away in the wide shot at 0:20.

Straight-down view of the front pens at 1:41, with pink dots on detected heads. — **The same place, seen twice.** In the wide shot the AI finds only about one head in seven, even in the best-lit part of the field. Further back, where it is darker, it finds one in twenty or fewer. So the wide shot can tell us where people stand, but not how many.

The same two pens outlined in green in the wide shot at 0:20, with far fewer pink dots. — **The same place, seen twice.** In the wide shot the AI finds only about one head in seven, even in the best-lit part of the field. Further back, where it is darker, it finds one in twenty or fewer. So the wide shot can tell us where people stand, but not how many.

How we counted, in four steps

The method is the one used for large crowds in general: measure the area, measure the density in sample places, and multiply. The AI marks decide how the density changes from place to place.

Measure the lawn

We traced the lawn edge on two independent satellite images and matched it to the video. The lawn is about 15,200 m², roughly two football pitches. We then removed the areas where nobody can stand: the stage, the runway, the T-shaped extension, the sound desk and the tents. That leaves about 13,900 m² for the crowd.

The wide shot with the lawn edge drawn in white and the stage, runway, sound desk and tents outlined in red dashes. — **The lawn edge** in white. Areas where nobody can stand are outlined in red dashes.

Measure the density in sample zones

Density means people per square metre. We measured it in the places where the video gives a clear view, and call those places reference zones.

Zone A, the front pens. The straight-down shot at 1:41 is the clearest view in the video. We placed 15 boxes on a regular grid over both pens, so we could not pick only the busy spots. In each box, three AI counters marked the heads, and we then checked every mark by eye, removing doubles and adding missed heads. The 15 boxes held 572 people, and the AI counters had found 68 to 75% of them. Applying that rate to all the heads they found in the whole frame gives about 3,200 people in the part of the pens this shot covers.

Straight-down view of the pens at 1:41 with 15 white boxes on a grid, each labelled with its head count. — **The 15 sample boxes** at 1:41, each labelled with its checked head count. Each box is about 3 × 3 metres. The stage is at the top. Boxes range from 12 people in the open area on the left to 70 people at the front beside the runway.

To turn a count into a density we need the size of the picture in metres. Three measurements agree: the runway's width compared with its known length (about 3.3 m wide), the gap between the pen fences, and the size of adult heads and shoulders. They give 74 to 90 pixels per metre. The result is 3.3 people per m², between 2.7 and 4.1 across that range. That is packed, as expected at the front of a sold-out K-pop show.

An earlier draft of this report used 60 to 70 pixels per metre, taken from equipment cases whose true size we did not know, and arrived at 15,200 people. The independent reviewer measured the runway differently and flagged the difference. We checked it, and the runway and body-size measurements agree with the reviewer.

Close-up of one 4 by 4 metre box beside the runway, every head marked with a pink dot. — **One box, close up.** The pink dots are the heads the AI counters found (68). We checked each one by eye, removed doubles and hands, and added the dark heads they missed: 70 people. This box at the front barrier beside the runway is the densest of the 15.

Zone D, behind the T-shaped extension. The shots from 0:30 to 0:44 look steeply down on this zone in good light. On 30 frames we compared the head density here with the front pens: it is 0.84 times the pen density, between 0.72 and 0.93. That gives 1.9 to 3.8 people per m². This ratio was measured on an angled view, so it is less certain than the pen density; it is the largest single source of doubt in the result.

Zone B, behind the sound desk. This is the weakest input. No shot looks down on this zone. The close shots at 2:52 and 2:56 are low and angled, so people hide each other. We judged the density at 1.5 to 2.5 people per m² by eye. We say more about this below.

Zone C, the back of the field. Here the crowd is thin and people stand apart, so the AI counters find most of them. We checked their count by hand in sample patches: 360 to 900 people in the back band.

Steep view at 0:42 with six white boxes behind the T-shaped extension.

Low close view at 2:56 of the dense crowd around the white sound-desk tent.

See where the crowd is

In the wide shot at 0:20 the AI counters mark every head they can find. Most heads are missed, but the marks still show the pattern: where the crowd is packed, where it thins out, and where the ground is empty. We averaged 25 frames around 0:20 and used three different counters, so no single frame or counter decides the result.

The wide shot with pink dots on detected heads: dense near the stage and in the middle, sparse at the back. — **What the AI sees at 0:20.** Each pink dot is one detected head. Dots are dense near the stage, where the light is good, and sparse in the dark areas, even where the crowd is packed.

The video also caught one thing the dots alone would miss. At 0:27.5, a strip along the west fence, south of the VIP lounge, is almost empty. We capped the density there at 0.1 to 0.5 people per m².

View at 0:27.5 with a dashed amber outline around a near-empty dark strip in the lower left. — **The near-empty strip** along the west fence at 0:27.5, outlined in amber.

Scale up and add up

We place every AI mark on a ground map of the lawn, using the camera position we worked out. Then, in each reference zone, we compare the marks with the measured density. In the pens, for example, there are about seven people for every head the AI marked. Far behind the stage extension, where it is dark, there are twenty or more.

These correction factors are blended smoothly with distance from the stage and applied to every mark. The result is a density map of the whole lawn. We then fill the parts of the lawn outside the frame from their neighbours, add the VIP lounge at a low density, and apply the cap on the empty west strip.

The wide shot with the four reference zones shaded green and labelled A, B, C and D. — **The four reference zones** as they fall in the wide shot. Everything between them is corrected by blending their factors.

Top-down map of the venue on a satellite image, coloured by crowd density: brightest in the front pens, fading towards the back. — **The final density map** on the satellite image of the venue. Bright colours are dense areas. The front 40 metres hold about 2.5 to 4 people per m², falling to under 0.3 at the back.

From 2,300 heads seen to 18,000 people

This chart shows how the number is built. It starts with the heads the AI actually found in the wide shot and adds each step in turn.

Step	People
Heads the AI saw at 0:20 (average of three counters)	2,314
Correction to the measured densities	+14,490
Dark area on the west side, filled from its neighbours	+834
Parts of the lawn outside the frame	+639
VIP lounge, at 0.5 to 1.2 people per m²	+514
Near-empty strip along the west fence	−761
People on the lawn	18,030

The middle of the 100,000 re-runs described below is 18,000, which is the number we report. The large correction step is not a guess: it is the measured densities from step 2, spread over the lawn with the pattern from step 3.

How sure are we

Every input has some uncertainty: the densities, the camera position, the lawn edge, and how the factors are blended. We re-ran the whole calculation 100,000 times, each time varying every uncertain input within its own range. The results form the range below.

In plain words: the crowd was very likely between 13,900 and 23,100, most likely between 14,700 and 21,900, and our best single estimate is 18,000.

What moves the number most

Some inputs matter more than others. This chart shows how much of the uncertainty comes from each one.

Three inputs carry three quarters of the uncertainty. The densities of zones D and B decide the level in the middle of the field, which holds most of the crowd. The camera position and lawn edge decide how the wide shot maps onto the ground. Zone D rests on a ratio from an angled shot and zone B on a judgement by eye, so these are the first things we would improve with better footage.

Sense checks

Average density

18,000 people on 13,900 m² is 1.3 people per m² on average. That fits a busy, sold-out concert: packed at the front, loose at the back.

The pens in two views

The same pens hold about 3,300 people when counted straight down at 1:41, and about 500 AI heads in the wide shot at 0:20. That one-in-seven ratio is what the model uses at the front of the field.

The steep shots

The best frame of the steep, lit sequence from 0:30 to 0:44 shows about 3,600 heads. It covers roughly the front 40 metres of the lawn, which our map puts at about 8,000 people. Seeing under half of the heads in an angled shot, against 70% straight down, is what we would expect.

A second, independent estimate

A separate reviewer built an estimate from scratch, without seeing ours: dividing the lawn into eight zones and judging each zone's density by hand. They found about 16,900 on a slightly smaller crowd area, or about 18,900 on the same area as ours, with a wide range of 10,200 to 24,900.

How to check this number against your own data

This estimate is independent. It was made from the video alone, without ticket or gate data. The best test is to compare it with what the organizers recorded.

Gate scans or wristbands by time. The number of people inside the venue around the time of the drone shot should be higher than 18,000. The difference covers the VIP stands, bars, toilets and queues.
Total tickets or unique scans. Total attendance for the evening should be higher again, because people arrive and leave at different times. At many concerts the total is 10 to 20% above the crowd at any one moment, but we cannot measure that from this video.
If your figure for the lawn alone is far outside 14,000 to 23,000, we would like to see it. It would point to an input we should re-examine, most likely zones D and B or the camera position.

Check it yourself. Every head the AI found, every sample box and every reference zone can be opened and inspected in the evidence viewer. Hover a dot to see where it lands on the ground and how many people it stands for in the final number.
Open the evidence viewer

Limits, and what would improve the estimate

Zones D and B are the weak points. No shot looks straight down on the middle of the field. Zone D comes from a ratio measured on an angled shot, and zone B is judged by eye. One straight-down photo of the middle would remove about half of the uncertainty.
The camera position is worked out, not recorded. The raw drone files contain the exact height, position and angle for every frame. They would remove the largest source of uncertainty.
The picture scale in the pens comes from the runway, the pen fences and body sizes. The stage plans, with the exact runway width, would narrow the pen density range.
One moment, one lawn. We counted the lawn at 0:20. We did not count the VIP stands, the areas outside the lawn, or turnover during the evening.

Appendix

Glossary

Density: people per square metre. 1 per m² is a loose crowd, 2 is busy, 3 is packed, above 4 is uncomfortable.

AI counter: software trained to find heads in photos. We used three that place a dot on each head (PET, P2PNet, APGCC) and one that estimates a count per small block of the picture (ZIP).

Reference zone: a place where we measured the density ourselves from a clear shot, used to correct the AI counters everywhere else.

Correction factor: the number of real people for each head the AI marked, measured in a reference zone.

80% range: in 80 out of 100 re-runs with the inputs varied, the result fell inside this range.

Method details

Lawn: outline traced on a Sentinel-2 image and on a Planet image from the setup week, matched to the video. Stage and runway positions confirmed on the Planet image.

Camera: position, height, angle and lens fitted to landmarks in the 0:20 frame (lawn corners, stage, runway, tents). Frames from 0:19.5 to 0:20.5 were aligned to 0:20 and their detections averaged.

Counters: run on the full 4K frames after a contrast boost, on a local GPU. On the 1:41 frame the three dot counters found 68 to 75% of the 572 checked heads. The pen density is their whole-frame count divided by that rate and by the frame's crowd area (without the runway corridor and the empty pit at the stage fence).

Picture scale at 1:41: 74 / 82 / 90 pixels per metre, from the runway width (0.12–0.13 of its 27 m length in the 1:35 frame), the gap between the pen fences (8.25 m), and adult shoulder and head sizes.

Reference densities (low / middle / high, people per m²): A 2.7 / 3.3 / 4.1; D 1.9 / 2.8 / 3.8; B 1.5 / 2.0 / 2.5; C 360 / 550 / 900 people in the back band.

Caps: no spot above 6 per m² in the pens or 4.5 elsewhere; excess is spread to neighbouring cells in the same distance band.

Uncertainty: 100,000 Monte Carlo re-runs, each input drawn from its own range; the range reported is the 10th to 90th percentile.

How the number changed during the work

We record every published draft, so the reader can see what changed and why.

Draft	Estimate	What changed
1	15,800	First zone-by-zone estimate
2	16,600	Camera model and density map
3	14,500	Fixes from four reviewers
4	12,800	Lawn area corrected from 20,400 to 15,200 m²
5	16,300	Zone D measured from the unused 0:42 shot; pens hand-counted
6	14,800	Near-empty west strip found; 25-frame averaging
7	15,200	Pens recounted on a 15-box grid; picture scale taken from equipment cases
Final	18,000	Picture scale corrected from the runway, pen fences and body sizes, after the independent reviewer flagged it; pen density from the whole 1:41 frame