Every camera, on a phone
All cameras, live, on officers' phones.
Officers opened any camera from the map or a grid, from anywhere on the ground, day and night.
Live video started in about 5–7 s on a phone
Deployment case study · Government & public safety
We gave Pitru Paksha Mela's CCTV cameras a super brain.
For sixteen days at Pitru Paksha Mahasangam 2026 in Gaya ji, across a network of 500+ cameras on ghats, gates and lanes: live and recorded video on every officer's phone, evidence clips, search, and crowd alerts. This page separates what we measured from what was reported to us.
Measured
Measured on the live deployment.
From our evidence pack of 5 October 2026: our own footage and hardware, not a lab benchmark.
500+cameras in the Mela's CCTV network
15.36 Mpeople-sightings indexed for search
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
| What | Result | Note |
|---|---|---|
| Cameras | A Mela CCTV network of 500+ cameras | Measurements below: live deployment, 5 Oct 2026 |
| Hardware | Analysis servers with NVIDIA RTX A4500 (20 GB) GPUs, on site | Spare GPU capacity at peak load |
| Search by description | Text query in 2.4–2.8 s over 1.52 M sightings; Hindi text precision@10 ≈ 0.97 | Ranked candidates, not automatic matches |
| Find the same person again | 3.8 s over 1.52 M sightings (72 h) | From an indexed sighting, not a phone photo |
| Crowd buildup alert | Reading → flag p50 66 s / p90 96 s; flag → phone ≈ 1 s | Count-based; plus up to 90–106 s for each camera's turn in the cycle |
| Crowd counting | Mean absolute error ≈ 20–24 people on held-out frames | Weak in very dense crowds; reported as uncertain there |
| Coverage cycle | Every camera revisited every 90–106 s | Round-robin across the estate |
| Continuity | Server up 2 days continuously; alerting 48 h unbroken at measurement | Observed, not a formal MTBF figure |
Reported
What officers used it for.
These outcomes are as reported by the deployment team and officers on the ground. They are not measured figures, and arrests and recoveries were the police's own work.
- Officers used photo and description search to find where separated pilgrims had last been seen, so police could search the area around it.
- From a reported theft, officers searched by appearance across cameras and identified suspects for the police's own investigation; every candidate was confirmed by a person.
- Crowd-buildup flags reached the control room and officers' phones; officers report adjusting barricades, routes and deployment in response.
- Recorded footage was found and exported in minutes instead of being searched on each recorder.
The officer app
What officers had on their phones.
Illustration of the app's screens — drawn, not screenshots
Playback like a video
Scrub back through hours of recording.
CamBrain AI kept its own recording, so officers could drag the timeline and the picture followed.
Recorded playback seeks measured at 21–52 ms
Evidence between two times
Pick a start and an end; export the clip.
Any window from any recorded camera became an evidence clip with a SHA-256 fingerprint.
Every exported clip fingerprinted with SHA-256
Shared and logged
Download it, or send it on.
Clips went to the officer who needed them, and every view and export was logged.
Every view and export logged to a named user
Alerts on the phone
The officer responsible is told.
Crowd buildup, a possible sighting for an open case, or a camera gone dark reached the officer for that area.
Camera reading → officer's phone ≈ 1–2 min, measured
Search by description
“Red saree, blue bag” — across every indexed camera.
Authorised officers searched sightings by clothing, in Hindi or English, and opened the moment in playback.
Text search 2.4–2.8 s over 1.52 M sightings
Missing-person cases
A family's photograph becomes an open case.
While a case was open, every camera was re-searched every three minutes; likely sightings went to the officer who opened it, to confirm.
Re-searched every 3 minutes while open
Theft investigations
From a reported theft to candidate footage for the police.
Officers searched by appearance across cameras; every candidate was reviewed by a person before police acted.
Targeted, logged, human-confirmed
Crowd counts and alerts
Each camera counted against its own usual level.
Cameras whose counts stayed well above normal for minutes were flagged on the control-room wall and phones — once, not every frame.
1.89 lakh crowd readings in 24 hours; peak 262 in one camera
Limits
What did not work, or was not measured.
Published next to the results, because a case study without its limits is an advertisement.
- Following one person automatically across cameras: measured to fail at this crowd density, so it was not used. Officers searched and confirmed instead.
- Absolute people-per-square-metre: camera geometry was recovered for 239 cameras, but absolute scale needs a measured reference per location before it drives an alarm.
- Number-plate reading: 0 plates read on the existing overview cameras, which do not have enough pixels on plates. It needs a camera aimed at the lane.
- Face-matching accuracy by skin tone and lighting: not measured, so face matching is not claimed.
What this means for other customers. Gaya was a government deployment under
the administration's authority. Search for a person and missing-person cases ran only for
authorised officers, with every search logged. These features are not offered to every customer,
and readiness for your site is stated on each industry page.
Running a large gathering or a city estate?
Tell us about your cameras and the event. We will assess what your existing CCTV can support.