Crowd Awareness and Evidence From the Cameras You Already Run.
At Pitru Paksha Mahasangam 2026 in Gaya, across a Mela network of 500+ cameras, CamBrain AI put every camera on officers' phones, with live video and alerts. The same system serves any authority with cameras on roads, ghats, stations and gatherings.
Simulated illustration, not real footage: what each detection checks and the alert it raises.
The problem
What you are up against.
Crowds are judged by eye, after the fact
A control room cannot see which of hundreds of cameras is filling. A count against each camera's usual level can.
Footage is slow to find and hard to use as evidence
Reaching an old moment on an NVR takes minutes. Clips leave on pen drives with no record of where they came from.
Searches for people need legal discipline
Crowd monitoring counts people and identifies nobody. Any search for a person is separate, authorised, logged and confirmed by a human.
What CamBrain AI can alert on
Use cases, with their readiness.
3 validated · 5 need site setup. Labels come from the product's capability register.
What the labels mean.
Simulation Watch larger
Crowd buildup
Validated
“A ghat fills well beyond its usual level on the main day.”
What CamBrain AI detects
A camera's count stays far above that camera's own usual level for several minutes.
Who gets the alert
The control room and the officers assigned to that area.
What happens next
Officers manage barricades, routes and deployment.
Measured live at Gaya from timestamps: about one to two minutes from a camera's reading to an officer's phone, with each camera read every 90–106 s. Count-based, against each camera's own usual level.
Simulation Watch larger
Occupancy over a set limit
Site setup required
“A temple courtyard passes the capacity the administration set.”
What CamBrain AI detects
The count inside a zone passes its stated capacity.
Who gets the alert
The sector officer.
What happens next
Entry is held until the count falls.
Runs in production today; not yet measured on a held-out set, so it is measured on your cameras first Needs a zone drawn on your camera.
Simulation Watch larger
Counterflow at gates and lanes
Site setup required
“People start pushing back against the flow at a narrow gate.”
What CamBrain AI detects
A share of movement goes against the expected direction in a zone.
Who gets the alert
The control room and the gate officer.
What happens next
Officers adjust the route at that gate.
Runs in production today; not yet measured on a held-out set, so it is measured on your cameras first.
Simulation Watch larger
Evidence clips between two times
Validated
“An officer needs the 20 minutes around a reported theft.”
What CamBrain AI detects
Recorded footage is scrubbed like a video and exported between two times.
Who gets the alert
The officer who requested it.
What happens next
The clip is shared with its SHA-256 fingerprint and the export is logged.
Used throughout the Gaya deployment. Recorded playback seeks in 21–52 ms; every exported clip carries a SHA-256 fingerprint.
Simulation Watch larger
Search recordings by description (authorised)
Validated
“A family describes a missing relative: red saree, blue bag.”
What CamBrain AI detects
Indexed sightings across cameras are ranked against the description.
Who gets the alert
Only the authorised officer who ran the search.
What happens next
The officer reviews candidates; nothing happens without a person confirming.
Every search is logged to a named officer with a reason.
Measured at the Gaya deployment: a text query, in Hindi or English, answered in 2.4–2.8 s over 1.52 million indexed sightings, on site.
Simulation Watch larger
Missing-person search from a photograph
Site setup required
“A family hands over a photograph at the help desk.”
What CamBrain AI detects
Every camera is re-searched every three minutes while the case is open.
Who gets the alert
The officer who opened the case.
What happens next
The officer confirms or rejects each likely sighting before police act.
Requires a legal basis, authorisation and human confirmation.
Once a case is opened, every camera is re-searched every three minutes, by appearance and, where face memory is enabled, by face. Matches are candidates for an officer to confirm; accuracy on phone photographs, and face accuracy by skin tone and lighting, are not yet measured.
Simulation Watch larger
Wrong-way and stopped vehicles
Site setup required
“A motorbike comes against the traffic on a one-way approach road.”
What CamBrain AI detects
A vehicle moves against the direction set for that lane, or stops where it should not.
Who gets the alert
The traffic control room.
What happens next
Traffic police respond.
Runs in production today; not yet measured on a held-out set, so it is measured on your cameras first.
Simulation Watch larger
Camera health across the estate
Site setup required
“Forty cameras go dark after a power cut and nobody notices.”
What CamBrain AI detects
Cameras that are dark, frozen, covered, moved or unreachable.
Who gets the alert
The control room and the maintenance team.
What happens next
Each camera is restored; the list shows which are still down.
Runs in production today; not yet measured on a held-out set, so it is measured on your cameras first.
Deployed at scale
Pitru Paksha Mahasangam 2026, Gaya.
A Mela network of 500+ cameras, live and recorded video on officers' phones, crowd alerts,
search and evidence export — measured on the live deployment.
2.4–2.8 sto answer a text search over 1.52 M sightings, in Hindi or English
≈ 1–2 minfrom a camera's crowd reading to an officer's phone, measured from live timestamps; each camera is read every 90–106 s
1.89 lakhcrowd readings in 24 hours — about two every second across the estate, each camera re-read every 90–106 s
262people counted in one camera at the peak hour
When an alert fires
Who is told, and what they see.
Who receives it
The control room, and the officers you assign to each ward, ghat or sector — on the control-room wall and on their phones.
What the alert contains
Camera, location and sector
The time and the count against the usual level
A snapshot, with live view one tap away
Who acknowledged it
What they can do
Open the live view or the clip, acknowledge it, and
close it when it is dealt with. Alerts that a fix has been made are checked by the camera before
they close. The decision is always a person's.
Technical conditions
What your cameras need.
Cameras on the approaches, gates and gathering areas, a server sized to the camera count, and a network path from each camera to it. Person search works best where people appear at a reasonable size in the frame.
Untargeted tracking of individuals across the city
Emotion, gait or 'suspicious behaviour' prediction
Identifying organisers or participants of lawful assemblies
Any action on a match without a person confirming it
Questions
Frequently asked.
Is this facial recognition?
Crowd monitoring counts people and identifies nobody. Search by description uses clothing and appearance. Where an authority uses face matching — at Gaya, for missing-person photographs — it runs under that authority's legal basis, for a named officer with a stated reason, and every match is confirmed by a person. Accuracy by skin tone and lighting is not yet measured, so it is not offered for sale.
Where is the video processed?
On servers inside the administration's premises. Authorised officers view cameras remotely over an encrypted connection; video is not stored on the way.
Does crowd density mean people per square metre?
Alerts at Gaya were count-based, compared with each camera's usual level. Absolute people-per-square-metre needs a measured reference on each camera, and we do not claim it until that is done.
See it on your own cameras.
Tell us about your site and the problems you want to solve. We will check your cameras and recommend one focused use case to demonstrate first.