AI video analytics use case
AI crowd detection and density monitoring
Short answer
Crowd detection is video analytics that estimates how many people are in a monitored area, or how densely they are gathered, and raises an alert when a build-up crosses a configured level. Operations and safety teams use it to spot congestion at entrances, queues, platforms and gathering points before it becomes a safety or service problem.
How crowd detection works
For moderate numbers, a person-detection model counts individuals in a zone. In dense scenes where people overlap heavily, density-estimation models predict how many people occupy each part of the image instead of detecting each person separately.
Rules compare the count or density in each zone against a level the site defines, and can require the condition to persist for a period before alerting, so brief surges do not cause alerts.
In OrionAlerts
In OrionAlerts, the AI runtime reads decoded frames from the same video engine that records the cameras, so detection runs on the streams you already record rather than on a parallel feed into a separate analytics appliance. Inference runs on the on-site Nebula OS node with GPU acceleration, and each event is written to the same record as the camera and site it came from. Alerts reach the dashboard, mobile app and email.
What affects reliability
Detection quality depends on the scene as much as the model. These are the factors to check during site survey and pilot.
- Camera angle: high, downward-looking views reduce people hiding behind one another.
- Occlusion: dense crowds overlap heavily, which is why density estimation is used instead of counting at high densities.
- Lighting: changing daylight, stage lighting and night conditions affect how clearly people can be seen.
- Zone design: thresholds should reflect the real capacity of each area, set with the operations team.
What you need
No camera replacement is assumed. A typical OrionAlerts deployment needs:
- Existing IP cameras that expose video over ONVIF, RTSP or a supported vendor SDK. Compatibility is confirmed per camera model during solution design.
- An on-site Nebula OS node with GPU acceleration, sized per deployment to the channel count and the detections you run.
- Camera views that actually cover the area to be monitored, with zones and schedules defined for each rule.
- Optional: Orion Cloud for a central view of alerts and health across multiple sites, without uploading raw video.
Where it is used
This use case commonly applies in the industries below. Each industry page covers the risks, detections and modules that matter for that kind of site.
Solution
Which solution it belongs to
Crowd Detection is part of AI Vision & Real-Time Alerts: turn live video into real-time AI alerts for safety, intrusion, PPE, and operational events.
Explore AI Vision & Real-Time AlertsFAQ
Crowd Detection: common questions
What is the difference between crowd detection and people counting?
People counting tallies individuals, usually as they pass an entrance or line. Crowd detection estimates how many people are gathered in an area, or how densely, and alerts when a build-up crosses the level set for that zone. Counting suits footfall reporting; crowd detection suits congestion and safety monitoring.
Can crowd detection use existing cameras?
Often, yes, especially elevated cameras that look down over the area, because high angles reduce people hiding behind one another. OrionAlerts connects existing IP cameras through ONVIF, RTSP and vendor SDKs, with compatibility confirmed per camera model during solution design. Low, eye-level cameras are less suited to dense crowds.
Who sets the crowd thresholds?
The site's operations or safety team sets them, based on the real capacity of each area and how it is normally used. Thresholds and schedules are configured per zone, and a rule can require the build-up to last for a set period, so brief surges at a doorway do not raise alerts.
Plan this use case for your sites
We will review your camera views for this use case, scope the deployment for your sites and show it in a live demo.