Step 01 — Connect
Connect any camera.
Point an existing CCTV, IP camera, or frame pipeline at our endpoint. No new hardware, no firmware, no vendor lock. If it produces a frame, Crowdcount can count it.
RTSP · HTTP · WebRTC→
§ 01 / Operating principle
One dot per person. One count per frame.
Crowdcount predicts a single point on every detected person. Counts follow from those points — not from a coarse density estimate, not from a guess. Every number is traceable back to what the camera actually saw.
Step 02 — Process
Process every few seconds.
Frames are sampled at whatever cadence your use case needs — once a minute for occupancy, once every few seconds for live flow — and scored by a point-prediction model tuned for dense scenes.
Cadence · 1s — 5 min→
Step 03 — Count
Count in near real time.
Results feed a live and historical dashboard with alerts on thresholds you define. Export to your stack, or pull counts back through the API. No black box — every detection is inspectable.
Alerts · Webhooks · Exports→
§ 02 / Use cases
Operational visibility, wherever people gather.
Built for teams who already have cameras and need the count those cameras can give them.
Featured · 001
Stadiums & arenas.
Matchday operations without staffing every corridor. Watch concourses, exit gates, and seating bowls for pressure points before they become incidents — and replay density against the event timeline after the final whistle.
Camera load30 – 300 feedsCadenceEvery 3 – 10 sLatency< 1 s
Use case · 002
Concerts & festivals.
Stage-front density, entry-gate queues, and camp-ground occupancy across a multi-day footprint — one live number per zone, with history replayed against the line-up.
Cadence5 sZonesStage · Gate · Camp
Use case · 003
Airports & transit hubs.
Security-checkpoint queue times, platform occupancy, and gate-line pressure — so controllers can hold, release, or re-route before things back up.
Cadence5 – 10 sAlertsThreshold
Use case · 004
Theme parks & attractions.
Live ride-queue times, zone-level density, and park-wide capacity — so operators can open lanes, stagger shows, and balance the park before guests notice the strain.
Cadence30 sZonesRide · Land · Gate
Use case · 005
Demonstrations & public gatherings.
Verified crowd-size estimates for organizers, public safety, and the press — grounded in the footage, not in guesswork from a podium.
Cadence30 sArchivePer-event
§ 03 / Evidence
Every count is traceable.
Frames in, points out. Mask the regions that don't matter, keep the ones that do. What you see in the viewfinder is exactly what the count was built from.
Overlay · key
Point prediction (1 = 1 person)
Masked region (excluded)
Frame 00418 · 3840 × 2160 ● 2,847 DETECTED · 312 MASKED
§ 04 / By design
Accurate, inspectable, EU-compliant — by construction.
Six things that are true of every count we ship. Not bolted on for the deck, built into the product.
● Privacy by design
We count people. We don't identify them.
No faces, no biometrics, no tracking IDs. Raw frames are discarded after inference — only predicted points and aggregate counts persist. GDPR-compliant by construction, not by consent dialog.
GDPR · Art. 5 · Data minimisation
● EU AI Act aligned
Point prediction, not identification.
Counting humans as dots is not biometric identification. We maintain full model documentation, evaluation records, and operational logs to support minimal-risk classification and any compliance review.
EU AI Act · Minimal-risk category
● EU hosting
Data stays in the EU.
Frankfurt-hosted by default. On-prem deployment available for sensitive sites where no frame should leave the building. Zero outbound data if your policy requires it.
eu-central-1 · On-prem ready
● Built for density
Thousands of people per frame.
The model is trained and evaluated on the hardest crowd scenes — marathons, matchdays, festivals. Where most counters fail past a few hundred, ours is just beginning to work.
Max 3,000 / frame · MAE 4.1%
● Inspectable outputs
Every count ships with its points.
No black box. The predicted points that produced every number are available alongside the number itself. Draw masks, audit the dots, trust the total — every count is traceable back to the frame.
Points · Masks · Audit trail
● Web app & API
Two ways in, one model underneath.
Browser-based workspace for ad-hoc frames, per-camera configuration, and mask drawing. REST plus webhooks for live feeds and monitoring. Same model, same numbers, same evidence.
REST · Webhooks · Browser