The short answer
Treat every crowd number as a range, not a point. Manual density sampling, aerial imagery, computer vision and mobile-signal methods all carry systematic bias and random noise, so two careful teams can honestly disagree by two or three times. Before acting, define the decision at stake, estimate a plausible interval, and test your plan against the worst credible value rather than the single headline figure.
Key takeaways
- Every method of counting an open crowd returns an estimate with bias and noise, so figures must be treated as intervals rather than exact values.
- Error creeps in from difficulty judging 2–4 people per square metre, non-uniform density, numbers changing over time, imagery artifacts, and device penetration limits in wireless methods.
- Match the method to the setting: area-density sampling suits dense static crowds, aerial and computer-vision methods suit open clear spaces, and wireless signals suit large dispersed moving crowds.
- Physical limits anchor decisions: above roughly five people per square metre a crowd stops behaving as individuals and becomes dangerous, so an area times a safe density gives a hard ceiling.
- Always report a count with its method, timestamp and range so independent estimates can be compared on the same basis.
- The most defensible counts combine two sources and calibrate one against the other at known points.
A crowd figure is an interval, not a number
At an event without tickets there is no exact way to count: people arrive and leave, occupy irregular footprints, and the human brain is far better at comparing sizes than at grasping large absolute numbers. Experts agree that accurately counting such a crowd is extremely difficult if not impossible, so every published figure is really an estimate.
Disagreement between estimates is not automatically a sign of bad faith. At a London march in November 2023, organisers reported 800,000 attendees while police put the number near 300,000; the same pattern followed the 2017 US presidential inauguration, where aerial photos and transit data clashed with official claims. The honest takeaway is to define your estimate as a range from the outset.
Where the noise and the bias come from
The dominant error is density judgment. Even experienced observers struggle to distinguish two, three or four people per square metre, and density is rarely uniform: people bunch near stages, screens and focal points and leave gaps elsewhere. Extrapolating a sampled density across a whole perimeter therefore compounds the mistake.
Error also creeps in from misjudging the usable area, and from timing — two counts half an hour apart can differ meaningfully as people arrive, leave or shift zones. Computer vision is degraded by shadows, poor light, weather and occlusion from banners or umbrellas. Wireless methods assume people carry switched-on devices with location enabled. Finally, being inside a crowd biases perception: people overestimate both its size and its emotional charge.
- Difficulty distinguishing 2, 3 and 4 people per m²
- Non-uniform density and clustering at focal points
- Fluctuations in numbers across the event timeline
- Imagery limits: light, shadows, weather, occlusion
- Device dependence of Wi-Fi, Bluetooth and mobile data
Match the method to the setting
The classic approach is the Jacobs method, devised in the 1960s when a journalism professor estimated protest sizes by dividing a plaza into squares, estimating average density, and multiplying by the number of squares. Modern tools do the same arithmetic with better data: aerial imagery and drones, computer-vision head counting, and wireless sensing from Wi-Fi, Bluetooth or mobile towers.
The right tool depends on the scene. Open spaces with clear sight lines favour aerial and vision methods; narrow streets, occlusion and low light undermine them. Static dense crowds suit area-density sampling, while large, dispersed or moving gatherings suit wireless sensing. Static crowds pack tighter than moving ones, which need more space per person. The most reliable results combine methods, for example calibrating wireless data against vision counts at choke points.
Build an uncertainty budget before you decide
Convert a single guess into low and high estimates. Suppose the occupied footprint is 4,000 m², of which 1,500 m² is dense. At about four people per square metre that zone holds roughly 6,000 people, while the remaining 2,500 m² at about one person per square metre adds about 2,500 — an order-of-magnitude total near 8,500. Recomputing with cautious densities turns that number into an interval, say 5,000 to 11,000.
Physics provides an independent sanity check. At four people per square metre movement is already hard; above about five, independent motion becomes difficult; near ten, the crush becomes lethal, as recorded in the Hillsborough disaster. Multiplying the measured area by five to six people per square metre therefore yields a realistic ceiling that no credible estimate should exceed.
Turn an uncertain count into a decision
Separate the decisions. Safety actions should be driven by the worst plausible value, while reporting and routine logistics can use the midpoint of the interval. If your plausible range straddles a threshold such as a zone's planned capacity, act as though the crowd is above it: control entry, open extra exits, and deploy stewards before density rises.
In crowd-risk management the forecast is built in advance from past attendance, ticket sales, publicity and weather, and safe capacity is assessed zone by zone. Define trigger densities and the person authorised to act on them ahead of time rather than while the area is filling. After the event, debrief with police and local authorities to refine assumptions for next time.
Report uncertainty honestly
Never present a single point as if it were exact. Always state the method, the time of measurement and the range, plus the assumptions about density and usable area. This is not weakness: it is what allows independent estimates to be compared on a common basis and what distinguishes methodological differences from deliberate distortion.
Remember that conditions and numbers change through an event and schedules can shift, so re-measure at several moments. Final decisions about people's safety remain a human responsibility, and an interval is the honest basis on which to make them.
Put it into practice
Crowd-count uncertainty checklist
A reusable sequence for turning a noisy headcount into a decision-ready interval, applicable at large events, venues and public gatherings.
- State the decision the number will drive (capacity, staffing, ingress control, reporting) before you measure anything.
- Choose the method by setting: area-density for dense static crowds, aerial or computer vision for open clear spaces, wireless for large dispersed moving crowds, and combine two sources where feasible.
- Measure the usable footprint in m² from a map or plan, excluding stages, barriers and unusable ground.
- Split the area into zones of visibly different density and estimate each zone separately rather than one average.
- Assign each zone a density band, not a single value — for example light ≈1, dense ≈2–4, critical ≈5–6 people per m².
- Compute low and high totals from your density bands; the span between them is your uncertainty range.
- Sanity-check against physics: area × 5–6 people per m² is a realistic ceiling, and confirm your estimate does not exceed it.
- Take measurements at several moments, since numbers rise and fall with arrivals, departures and programme flow.
- Record method, timestamp, area, density assumptions and known biases such as lighting, occlusion or phone penetration.
- For safety decisions use the worst credible value; for reporting use the midpoint of the range.
- At debrief, compare with police or organiser figures only when method and assumptions are stated.
- Pre-agree trigger densities, the response action and the person authorised to act before density approaches a limit.
Questions people ask
Why do police and event organisers so often give very different crowd numbers?
Because no exact count exists and every estimate carries bias and noise. Organisers have an incentive to show scale, police tend to be conservative, but even in good faith different methods measure different things: an area-density judgement, aerial imagery and mobile-tower data each answer a slightly different question. A two-to-three-fold spread is therefore not proof of dishonesty; it usually reflects method limits, the time of the count and the non-uniformity of the crowd.
What is the Jacobs method and is it still used today?
Herbert Jacobs, a journalism professor, devised it in the 1960s while estimating protest sizes: he divided a plaza into equal squares, estimated the average number of people in a sample, and multiplied by the number of squares. Today the same logic runs on aerial photos and drones, and estimates are still built by multiplying measured area by assumed density. Density rules of thumb vary by source — roughly one person per m² when spaced out, two to four when dense, and up to five to six at the limit — so the method yields an interval, not an exact figure.
How accurate are automated and AI crowd counters compared with humans?
Computer-vision and deep-learning counters work well in open spaces with clear sight lines, but degrade with shadows, poor light, bad weather and occlusion from flags or umbrellas, and dense crowds create heavy mutual occlusion. Because of this, researchers increasingly recommend that counting models return a density with an uncertainty range rather than a single number. The most reliable results come from combining methods and calibrating one source against another, not from trusting any single algorithm or human observer.
At what density should I start acting to protect crowd safety?
Use these physical anchors: at about four people per square metre free movement is very difficult; above roughly five, people can no longer move independently and the crowd behaves like a fluid; near ten per square metre a fatal crush can occur, a density documented at Hillsborough. Because density estimates are approximate, set action thresholds with margin — if the plausible range crosses a threshold, respond as though the crowd is above it. Exact thresholds depend on the venue and should be set by the organiser with stewarding and security teams.
Can mobile-phone and Wi-Fi data give an exact headcount?
No. Wireless methods estimate the number of switched-on devices with location enabled within coverage, not the number of people: some attendees carry no phone, many have devices off or without location services, signals overlap, and identifiers may be randomised. These methods are most useful for large, dispersed or moving crowds where aerial imagery is impractical, but they should be calibrated against another source and reported as a range with stated assumptions about device penetration.
How should I present a crowd estimate so it is not misleading?
State three things: the method (area and density, aerial imagery, video counting or wireless data), the time of measurement, and a low-to-high range rather than one point. Add the assumptions you made — which density values you used and what you counted as usable area. Reporting a range with its method is what lets your figure be compared honestly with police or organiser counts and shows what the disagreement is really about.
Sources and further reading
Sources were checked when this page was generated. Confirm changing dates, rules and prices with the original publisher.
- How accurate are crowd counts at large events?Full Fact
- Assess crowd safety risks and identify hazardsHealth and Safety Executive (HSE)
- How do scientists estimate crowd sizes at public events – and why are they often disputed?University of Melbourne
- How Are Crowd Sizes Determined?National Geographic Education
- Evaluating Crowd Density Estimators via Their Uncertainty BoundsarXiv (preprint)
- Как полиции удаётся посчитать количество людей в огромной толпеTechInsider
- Сколько человек было на митинге 5 октября?Factcheck.kg