Know a crowd is dangerous before it is. Crowd disasters are rarely a surprise to the cameras; they are a surprise to the people who were not told in time, and crowd analytics for public safety turns density into an alert while it is still an operational problem.
The Module
At the core of CrowdCount is a convolutional network trained to detect human figures in open spaces such as railway platforms and airport terminals. It annotates figures on sampled frames and tracks them in between, then checks those figures against three things: whether they are inside the region of interest you drew, how many of them there are, and how long they have been there. When all three hold, a Crowd event fires.
That gives the module two jobs at once. The first is real time: an alert on potentially dangerous crowding, early enough for stewards to open a gate, hold a flow or redirect an entry. The second is counting, within any region you define and up to 3,000 people, which is what turns a crowd-management plan into something you can review and defend afterwards.
One behavior is worth knowing before a pilot rather than during one. A crowd is a durable event: it is raised when the crowd builds up, and a crowd that later dissipates raises no second notification. Whether it is still there is read off the event duration indicator on playback, and the recount interval is also why a playback timeline can sit slightly behind the timestamp on the card.
The module is documented as able to run in parallel with any other module, so density, unattended items, weapon detection and watchlist alerts arrive in one operational picture rather than out of four systems.
The Settings
Every value below is drawn or typed per camera. The two thresholds and the recount interval are what a commissioning visit exists to settle, because a stadium queue and a platform edge want very different answers from the same module.
The part of the frame the count applies to, drawn as a polygon and editable vertex by vertex. A concourse usually needs several regions rather than one, because a number that averages a full doorway with an empty hall tells nobody anything.
Silhouette where the expected number is moderate, Top of head for big crowds, where whole bodies stop being separable. Choosing top of head also enables Save people’s position, which is what makes the on-the-fly count below possible.
The number of people that makes it a crowd, an integer from 3 to 35. This is the alarm threshold and not the counting ceiling: the module estimates up to 3,000 people in a region, and this setting only decides when it says something.
How long that many people have to stay inside the region before the event fires, from 15 to 600 seconds. It is the setting that separates a crowd from a train unloading, and it has to be at least three times the recount interval.
The check interval, from 5 to 60 seconds, and at least three times shorter than the crowd duration so a detection is confident rather than lucky. It is also why an event card can sit slightly ahead of the playback timeline.
Where positions are saved, an operator can redraw the region inside the media player while the video is playing and get a count for that new area. An incident review is often a question about a doorway nobody instrumented in advance.
Under Load
Density analytics earns its place at the three places a venue actually loses control: the queue outside, the concourse behind the turnstile line, and the entry point between them. That is where a delay of two minutes changes the outcome, and where a number arriving while the flow can still be held is worth more than a precise one arriving afterwards. IREX runs crowd management and banned-fan watchlists across every Premier League national stadium in Bosnia and Herzegovina, and crowd density monitoring at Dublin Airport during peak travel periods.
Requirements
Where It Runs
Density becomes a safety question wherever a large number of strangers moves through a space with a fixed capacity and a fixed number of exits.
Turnstile lines, concourses and stands on match day, with banned-fan watchlists running on the same cameras, across every Premier League national stadium in Bosnia and Herzegovina.
Read More →The platform edge, where density is a safety problem rather than a service one, on a platform that fills and empties on a ninety-second cycle.
Read More →Security queues, terminals and gate areas at continuous high throughput, monitored at Dublin Airport to prevent dangerous congestion during peak travel periods.
Read More →Squares, interchanges and event routes as one layer of a city operating picture, so the services that would have to act on a crush are looking at the same number.
Read More →Up to 3,000 within a region of interest you define on the camera. Larger areas are covered by defining regions across several cameras. Note that this counting ceiling is a different number from the "people in crowd" threshold that decides when an alert fires, which is an integer between 3 and 35.
Both, and they are different jobs done by the same module. Real-time alerting flags potentially dangerous crowding while there is still time to act; counting gives numbers per region for review and reporting. The settings that separate them are the crowd threshold and the crowd duration, and both are set per camera.
None, and deliberately. The module documentation is candid that performance efficiency can only be estimated in each customer’s own conditions, and IREX policy is not to publish per-module accuracy figures. Crowd counting is unusually sensitive to camera height, angle and pixels per meter, so the honest answer is the pilot: measured on your own cameras, against criteria agreed in writing beforehand, with the measured numbers going into the contract.
Not with a notification, and this catches people out. A crowd is a durable event: it is raised when the crowd builds up, and no second notification is released when it dissipates. What the platform gives you instead is the event duration indicator on playback, which shows the moment the crowd built up and the moment it cleared.
Yes. The guide states directly that CrowdCount can run in parallel with any other module, which is the normal configuration at a venue or a station: density, unattended items and watchlist matching on one channel, arriving on one alarm monitor.
Yes, unlike the intrusion and abandoned-object detectors. CrowdCount documents fixed or PTZ cameras, with the caveat that an en-face view of the crowd should be ensured as far as the site allows, and that image quality inside the region has to stay above 100 pixels per meter.
Any camera that supports ONVIF or streams RTSP with H.264 or H.265 on a static IP can be connected, but connection is not the bar. Each module sets its own resolution, mount position, height, and angle-of-view requirements in the camera requirements, the flagship modules need higher-resolution imagery of the scene than a bare stream, and the recommended primary stream is 1920 × 1080 or 1920 × 1440. A site survey confirms every camera against the module it will run.
It runs on its own and alerts a person. Detections are signals for a human to verify: no response runs autonomously, and the verification and decision are logged against the same Case ID as the detection.
Keep Reading
Crowd analytics can only really be judged under load. Pick your busiest fixture.