What problem does it solve?
Crowd risk builds where many people converge on a narrow space: an entrance, a staircase, a circumambulation area, a station platform. A control room can have hundreds of camera views and no way to count heads on each of them, and a head count from the floor arrives after the crowd has formed. Operators tend to react to a crush they can see rather than to the density that came before it.
Counting is the established half of the technology: a model estimates how many people are in a zone and in which direction they move. Flagging a group moving against the stream, a sudden gathering or a sudden dispersal is the newer half, and the one the Paris trial found harder to assess. The output of either lands on a screen or an alert in the operations room. Public evidence comes from two very different settings.
The Saudi authorities report large scale operation at Hajj and publish no accuracy figures. The evaluation committee for the French trial at the Paris 2024 Games, as quoted in the press, reported transport operators finding density analysis satisfactory with reservations and crowd movement detection hard to assess.
How does it work?
- Place and calibrate the cameras. Counting accuracy depends on camera height and angle. In the Paris trial the software struggled to count tightly packed people because of how high the cameras were mounted. Some deployments add thermal imaging.
- Estimate density per zone. The model converts video into a people count or density value for each zone, near real time, and totals zones into a site figure that planners can compare with an expected attendance.
- Track direction and speed. Flow models follow movement to show a stream, a bottleneck or a group moving against the stream, as the Baseer platform at the Grand Mosque does.
- Compare with a threshold set by safety planners. Each zone has a density level at which staff should act, and the system raises an alert or a dashboard color when a zone approaches it. Some systems also forecast where a surge will form.
- Show the operator, who decides. The alert links to the camera view. The operator checks it and chooses the response: more stewards, a closed gate, a one way system or an announcement.
- Review afterwards. Alerts, operator verdicts and any incident the system missed are logged, so thresholds are tuned from the site's own data.
- Audience
- Employee facing
- Autonomy
- Assist
- Adoption
- Emerging
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Risk and loss reduction, Speed and cycle time, Customer experience.
Indicative value
A convention and exhibition center with 60 cameras on public areas and about 120 event days a year
USD 4800 to USD 84,000
Annual control room watching time cost avoided per year
How this is calculated
Formula: eventDays * hoursPerDay * operatorsOnCrowdScreens * shareOfWatchTimeSaved * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Event days per year eventDays, days per year | 100 | 140 | The reference venue. |
| Hours per day with crowd areas watched hoursPerDay, hours per day | 8 | 12 | Editorial assumption, replace with your own opening hours. |
| Operators watching crowd areas operatorsOnCrowdScreens, operators | 2 | 4 | Editorial assumption, replace with your own control room roster. |
| Share of watching time replaced by automatic counts and alerts shareOfWatchTimeSaved, fraction of watching time | 0.1 | 0.25 | Editorial assumption, replace with your own. No public deployment reports a time saving, so this is kept low. Operators keep checking every alert, so this is time redeployed to other work more than a smaller roster. |
| Fully loaded cost of a control room operator costPerHour, USD per hour | 30 | 50 | Editorial assumption, replace with your own fully loaded cost. |
What it leaves out: Operator time only. It leaves out the value of an incident avoided or caught earlier, the cost of cameras, licenses and integration, and the staff time spent on false alerts.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Saudi Data and Artificial Intelligence Authority (SDAIA)
Saudi Arabia · Government and public sector · 2025
Baseer is a computer vision platform that the Saudi Data and Artificial Intelligence Authority built with the Ministry of Interior for the Grand Mosque in Makkah and, for Hajj 1447 (2026), the entrances of both holy mosques. Reporting from the Ramadan 2025 Umrah season says it tracks crowds, detects movement against the flow of pilgrims and helps find missing persons. In May 2026 SDAIA's spokesman said it uses computer vision and large language models and gives field commanders immediate analysis, and SDAIA told Al Arabiya English that it and Sawaher combine computer vision, thermal imaging and AI analytics to monitor crowd density and movement in real time. No detection or response figure has been published.
No outcome disclosed.
Saudi Data and Artificial Intelligence Authority (SDAIA)
Saudi Arabia · Government and public sector · 2025
Sawaher is an AI platform that the Saudi Data and Artificial Intelligence Authority built with the Ministry of Interior. It analyzes live broadcasts, recordings and images from smart cameras at the holy sites and the main land and air entry points, and uses density and crowd counting algorithms to support Hajj security and to estimate pilgrim numbers. For Hajj 1446 (2025) SDAIA reported more than 5,000 cameras integrated across more than 80 locations, with results shown to operations rooms and monitoring centers. In May 2026 its spokesman described the system again for Hajj 1447. No accuracy, incident or response time figure has been published.
No outcome disclosed.
Ministry of the Interior (France)
France · Government and public sector · 2024
Under the 2023 law for the Paris Olympic and Paralympic Games, France trialed AI analysis of public space and transport camera feeds for eight authorized situations, among them crowd movements and crowd density, abandoned objects, intrusion and weapons. An evaluation committee delivered its report to the Interior Ministry on 14 January 2025. As quoted by the French press, it found uneven technical performance.
Intrusion detection into restricted areas was the one use with unanimous support, while 62 percent of the 270 abandoned object alerts were false positives. Transport operators rated density analysis satisfactory with reservations about tightly packed groups, and crowd movement was hard to assess because the software confused groups walking together with real surges. Only one software was tested and about half of the authorized uses were never tried.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Camera coverage of entrances, queues, corridors and gathering areas with usable sightlines at the height they are mounted
- A zone map with capacity figures and the density at which each zone needs action
- Historical attendance and incident records to tune thresholds
Systems to integrate
- Video management system or camera streams
- Operations room displays and alerting tools
- Ticketing or access control counts, to compare with the camera figures
- Public address, signage and radio systems used for the response
Complexity: Medium
The counting and flow model is bought from a video analytics vendor or built on a computer vision platform the venue licenses or develops itself. The venue's work is camera placement and coverage, a written density threshold for every zone, the link from alerts to the operations room, and a response plan for each alert level that the operators have rehearsed.
- 1
Start with the zones where crowds form first
Pick the entrances, queues or corridors where crowding has caused a problem before, and run the model beside the operators for a full event before any alert is used to act.
- 2
Count against a manual reference
Take manual counts at set times in each zone and compare them with the model's figure, including tightly packed groups and the busiest minutes. The Paris trial found counting harder when people stood close together and cameras were high, so test those conditions on your own cameras.
- 3
Write the response plan for each alert level
Decide in advance who is told, what they check and what they may do at each density level, and rehearse it. An alert nobody has a plan for only adds to the noise.
- 4
Treat movement detection as a second stage
Density counts are the easier part to validate. Flow alerts such as movement against the stream need their own test cases, because the Paris committee reported the software sometimes mistaking groups walking together for a crowd movement.
- 5
Review alert quality on a schedule
Sample confirmed alerts, false alerts and missed incidents each month, and retune thresholds from that record rather than after an incident.
Guardrails
- Every alert goes to a trained operator who checks the camera view before any response that affects visitors
- The system counts and measures movement; identifying or tracking named individuals is a separate capability that needs its own legal basis and assessment
- Each alert is logged with the zone, the figure, the camera view and the operator's verdict
- A manual fallback, such as stewards and manual counts, stays in place for every zone the system covers
KPIs to instrument
- Count error against manual counts, by zone and by crowd level
- Confirmed alerts against false alerts, by alert type
- Time from alert to the first response action
- Incidents the system missed and found another way
Human in the loop
An operator in the control room reviews every alert and chooses the response. The safety manager owns the thresholds and the false alert rate, and reviews both after each event and whenever the vendor changes the model or the cameras change.
Common failure modes
- Movement confused with groups walking together
- The French evaluation committee, as quoted in the press, said the processing can mistake groups of people moving the same way without particular haste for a crowd movement, and that it is hard to define a rapid regrouping or dispersal. Test movement alerts on real footage from your site before staff rely on them.
- Undercounting in dense, high angle views
- The same committee reported that the software sometimes had difficulty counting people packed tightly together because of the height at which cameras were placed. Check counts at the densest points and add a camera angle or a second source of counts where the error is largest.
- Figures that look precise without a published test
- The Saudi platforms report camera numbers and the number of algorithms, but no count accuracy or incident result has been published for them. Ask a vendor for accuracy on crowded footage like yours, measured by someone other than the vendor.
What are the risks and rules?
EU AI Act
Depends on design
Counting people and estimating flow without identifying or categorizing individuals is not listed in Annex III. The tier changes with the design: remote biometric identification of people in a crowd falls under Annex III point 1(a), biometric categorization under point 1(b), and use by or on behalf of law enforcement to assess or profile individuals under point 6. Real time remote biometric identification in publicly accessible spaces for law enforcement is prohibited under Article 5(1)(h), with narrow listed exceptions. France ran its Paris 2024 trial under a national law that authorized eight listed situations, after an opinion from the data protection authority CNIL, according to the French press.
Controls to put in place
- A data protection impact assessment before cameras are linked to analytics, with a stated purpose limited to safety and flow
- Counting and flow models that output aggregate figures and alerts, not identities
- Retention limits for footage and analytics logs, and signage that tells visitors about camera analytics
- Independent testing of count accuracy on the venue's own footage before an alert is used to act
Frequently asked questions
- Does AI crowd monitoring identify the people in a crowd?
Counting and flow models measure density and direction and do not need to know who anyone is. Some deployments add separate functions, for example the Baseer platform at the Grand Mosque is also reported to help locate missing persons, and how it does so is not described. Identification changes the legal tier and needs its own basis.
- How accurate are AI crowd counts?
No public deployment on this page publishes a count accuracy. The Paris evaluation committee, as quoted in the press, called density analysis satisfactory according to transport operators but with reservations: the software sometimes had difficulty counting people packed tightly together because of camera height. Measure accuracy on your own cameras against manual counts.
- Can AI predict a crowd surge before it happens?
According to SDAIA, as reported by the Bahrain News Agency, the Baseer and Sawaher platforms identify congestion points and predict crowd surges at Hajj. No lead time, hit rate or false alert rate has been published, so treat prediction as a claim to test. Movement detection was the weaker part of the Paris trial: the committee found real movements hard to tell from groups walking together.
- Who decides what to do when the system raises an alert?
An operator in the control room. In the Saudi and French deployments the output goes to operations rooms or command rooms, and staff choose the response. The system supplies counts and alerts; it does not open gates or move people.
How to cite this page
Blits.ai AI Use Case Library, "AI crowd density and flow monitoring for venues, transport hubs and pilgrimage sites", last verified 14 October 2026, https://www.blits.ai/ai-use-cases/crowd-density-and-flow-monitoring. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 14 October 2026: First published