Person Detection: The Foundation Layer

Before you can count a crowd, draw a line, or alarm a restricted zone, one primitive has to work: person detection – is there a person in the frame, and where. It is the foundation layer that every higher rule is built on, and getting it right is what makes the rest reliable.

What follows covers what person detection returns, how it differs from face detection and from the counting or zone rules above it, and how to deploy it on cameras you own.

What Person Detection Returns

The model scans each frame and returns a bounding box around each person – their position, their size in frame, and a count-ready signal. From that single output a system can confirm a zone is occupied, track how many people are in view, and hand the positions to higher rules. It does not say who anyone is; it says someone is there.

Person Detection vs Face Detection

They are different models with different jobs. Face detection locates faces; person detection locates whole bodies. Person detection is the broader and more robust of the two, because a body is visible from more angles and distances than a face. Where a site does need to confirm identity, that is a separate project; face recognition and verification covers what that involves. Stay at the person layer and most privacy reviews stay simple; for the framework these deployments map onto, see the GDPR resources and the NIST Privacy Framework.

Person Detection vs Counting and Zone Rules

The rules above person detection are just configurations on top of it:

• Counting aggregates the detections into a headcount – the job of people counting on camera and vehicle and person count-limit detection.
• Intrusion fires when a detected person enters a zone that should be empty; intrusion detection is that rule.
• Line crossing fires when a detected person crosses a virtual line; line crossing detection is that rule.
• Dwell and departure track how long a person stays or when a post goes empty, handled by person dwell detection and person leaving detection.

None of those rules is meaningful if the base detection is weak. Person detection is where the tuning budget should start.

Accuracy Factors, Honestly

The same conditions govern any video model, and a site survey settles most before go-live:

• View and distance. A person needs to be large enough in frame. A camera that sees a forecourt from across a lot reads presence well and fine detail poorly – which is fine, because detection needs only presence.
• Angle. A body-on or three-quarter view reads reliably; a full top-down is harder. Positioning the camera on the approach path solves most cases.
• Occlusion. Crowds, racks and forklifts can hide a person. Design the zone around the open approach, not the cluttered middle.
• Lighting. A dark entrance or night-time yard degrades any model; that is a camera and lighting question, not an algorithm one.

Running It on the Cameras You Already Have

Detection does not need new cameras – it needs a reasonable view of the approach. An edge AI box ingests the streams your existing units already produce, runs the models on-device, and holds the rules as configuration rather than hardware. Systems scale from 2 to 128 channels with 1 to 256 TOPS matched to load, carrying a library of 198+ pre-built algorithms in software. Practically, the entrance camera can also be the one that detects a person, and adding a zone rule later is a settings change rather than a purchase order.

What It Costs

Adding analytics to existing cameras starts at USD 399 for Standard 2, with Standard 4 at USD 999 and Standard 6 at USD 1,599; larger channel counts are quoted per site. Each Starter Kit ships with 10 metres of cabling as standard across the whole kit, with additional cable at USD 25 per 10 metres. The full breakdown is on our pricing page.

For a configuration-led project, the first working configuration on your own footage typically lands in about seven days. Bespoke algorithm development is different: until we have seen your video – the angles, the lighting, the specific zone you need watched – nobody can honestly commit to a date, and we do not.

For a related deployment angle, see our guide to a Warehouse Security Camera System: No Monthly Fee.

For a related deployment angle, see our guide to a Ring Camera Alternative: No Monthly Fee, On Your Cameras.

Frequently Asked Questions

Does person detection identify who someone is?

No. It reports that a person is present and where, not whose name is on them. Identity is a separate model kept in a separate project; staying at the person layer keeps deployments simple.

How is it different from face detection?

Face detection locates faces; person detection locates whole bodies. Person detection is broader and more robust, visible from more angles, and it is the base that counting, intrusion and line rules are built on.

Is person detection enough on its own?

For presence and basic zone alerts, yes. For headcounts, restricted-zone alarms or line rules, those are configurations layered on top of detection – the same base output, different thresholds and zones.

Where does accuracy break down?

Partial occlusion in crowds or racking, a fully top-down angle, and poor lighting are the usual causes. A short commissioning review against your real footage, with the zone drawn on the open approach, resolves most of them.

Next Step

Send us a photo of the entrance or zone you want watched and we will tell you whether your existing camera can carry the detection. For a wider view of turning cameras into event systems, see what AI video analytics does. Start with a Starter Kit to validate person detection on your busiest door, or tell us your scenario for a configuration quoted per site.