// INSIGHTS
Crowd Density Monitoring at a Large Festival
6 min read
A three-day music festival on the outskirts of Jaipur expects 60,000 people across two arenas, one food court and four gates. The safety officer does not need a dashboard. She needs to know which of the eleven pinch points on her site plan is filling up faster than people can leave it, and she needs to know it before it is visible from the ground. That is a measurement problem. Cameras and computer vision can solve part of it. Not all of it.
What density actually measures
Crowd risk is not a headcount. It is people per square metre in a defined area, combined with how fast that area is draining. A field at 1.5 persons per square metre is a comfortable festival crowd. The same field at 4 persons per square metre is a crowd where individuals lose the ability to choose their own direction. Above roughly 5 to 6 persons per square metre, pressure waves travel through the crowd and people can be injured while standing still. These are not our thresholds. They come from decades of pedestrian engineering literature and they are the numbers your safety plan should already reference.
Computer vision estimates density by counting heads inside a calibrated ground region. The camera sees a trapezoid of real-world floor. We map pixels to square metres during setup, run a crowd-counting model on each frame, and divide. The output is a number per zone, refreshed every 60 to 90 seconds, plotted as a colour grid on the site plan.
Flow is the second measurement and often the more useful one. At gates, bridges and corridors we count directional crossings per minute. A gate that ingests 900 people per minute and a corridor downstream that clears 400 per minute is a queue that will exist in ten minutes whether or not anyone has noticed yet. Density tells you the current state. Flow tells you where the state is heading.
Working numbers for a festival deployment
What the system provides
Zone density grid
Every monitored area is drawn on the site plan and coloured against agreed thresholds. Colour changes are logged with a timestamp so the post-event review has a record.
Directional flow counting
Gates and corridors report entries and exits per minute in each direction. Net accumulation per zone is derived from these counts rather than guessed.
Short-horizon trend
A 15-minute projection based on current flow rates, shown as a simple arrow and a number. It is a trend line, not a prediction of behaviour.
Control room and radio integration
Threshold breaches raise a card in the control room with the zone name in the language the operators use, and can push to the incident escalation chain.
How a deployment runs
Site survey before anything is mounted
We walk the ground with the safety plan and mark the zones that matter. Camera positions are chosen for counting geometry, which is usually higher and steeper than a surveillance mount.
Calibration against manual counts
During load-in and soft opening we compare model output against clicker counts at each zone. Calibration factors are recorded per camera, not applied globally.
Thresholds agreed with the safety officer
Every colour has a written action attached. Amber means one specific thing that a named role does. Without that, the dashboard is decoration.
Live operation and debrief
Operators run the grid through the event. Afterwards we hand over a time-series export of density and flow per zone for the incident review and the next year's plan.
What usually goes wrong
The most common failure is inherited cameras. Surveillance cameras are mounted low and oblique to capture faces. Counting cameras need height and a steep angle so that bodies do not fully occlude one another. Reuse a face-oriented camera for density work and error at 4 persons per square metre can exceed 30 percent, always in the direction of undercounting, which is the dangerous direction. Night is the second problem. Stage lighting produces strobes, colour washes and deep shadow within a single frame. Dust and winter fog in North India degrade accuracy further. We state a per-camera confidence and drop zones out of service rather than publish a number we do not believe.
The deeper failure is treating density as the whole of crowd risk. Crushes are usually flow failures at a constriction, not field-wide density events, and they can develop in under 90 seconds in a location that read amber a minute earlier. A dashboard cannot substitute for stewards, barrier design and a plan for closing a gate. On privacy, crowd counting does not require identity and should not collect it. We run these models on low-resolution feeds where individual faces are not recoverable, we do not retain frames beyond the processing buffer by default, and we keep density analytics on a separate purpose and retention basis from any surveillance recording under the DPDP Act. If a proposal bundles counting with face recognition without a separate justification, question it.
Common questions
Do you need to identify people to count them?
No. Crowd counting models detect head-like shapes and produce a number. There is no identity, no matching and no template. We deliberately run these pipelines at resolutions where individual identification is not possible, which also keeps the DPDP obligations narrow and easy to document.
How accurate is it really?
In a well-lit scene with correct camera geometry, expect roughly plus or minus 10 percent at moderate density. Accuracy degrades as density rises and occlusion increases. We publish the error band per camera and calibrate against manual counts on site rather than quoting a single laboratory figure.
Can it use the CCTV we already have?
Sometimes, for flow counting at gates where the geometry happens to be usable. For area density it usually cannot, because surveillance mounts are too low. We audit the existing estate first and tell you which cameras are usable, which need repositioning and which zones need new units.
Plan the measurement before the event, not during it
Send us your site plan and expected footfall and we will mark which zones can be measured reliably and which cannot.
