Camera placement for face recognition

How many pixels does a face need?

The useful measure is not the camera's megapixels but the pixels landing across the face at the point of capture — conventionally counted between the centres of the eyes. A 4K camera pointed down a long corridor can deliver fewer usable pixels on a face than a 1080p camera at a doorway, because resolution is spent on everything else in the frame.

• Below roughly 40 pixels between the eyes, detection is possible but recognition becomes unreliable regardless of algorithm.
• Around 60 to 80 pixels between the eyes is a workable operating range for cooperative capture at a door or gate.
• Above roughly 120 pixels the returns flatten: extra resolution stops improving the match and starts costing bandwidth and storage.
• Compression matters as much as resolution. Heavy H.264 or H.265 compression destroys the fine texture that matching depends on, and a highly compressed 4K stream can be worse than a lightly compressed 1080p one.

The practical consequence is that field of view is the setting to fight over. A camera covering a four-metre-wide corridor spreads its sensor across four metres; the same camera covering a one-metre doorway spends all of it on the person walking through. Narrowing the view is usually cheaper and more effective than buying a higher-resolution camera.

What mounting height and angle work?

Face recognition is trained predominantly on frontal and near-frontal images, so every degree away from frontal costs accuracy. The dominant error in real installations is a camera mounted too high, looking down at the tops of heads.

1. Aim for the lens to sit close to the eye level of the population being recognised — roughly 1.5 to 1.8 metres for standing adults.
2. Keep the downward tilt under about 15 degrees. Beyond that the face foreshortens, the eyes fall into shadow under the brow, and the geometry the matcher relies on distorts.
3. Keep horizontal yaw under about 15 degrees as well. A camera off to one side of a doorway sees a profile of anyone walking straight through it.
4. Where a high mount is unavoidable for tamper reasons, move the camera further back and use a longer lens rather than tilting it down more steeply. Distance flattens the angle; tilt does not.
5. For seated or wheelchair-using populations, the eye-level assumption changes and a single camera height will not serve both. Two cameras at different heights is the honest answer, not a compromise height that serves neither.

Why does lighting break more deployments than anything else?

A camera exposes for the whole scene. If there is a window or a glass door behind the subject, the camera exposes for the bright background and the face becomes a silhouette with no recoverable detail. This is the single most common cause of a system that worked in the demo and fails in the lobby, because demos are not usually staged in front of windows.

• Never point a face recognition camera at a window, a glass entrance or a bright exterior. Face the camera into the building wherever the layout allows.
• Where backlight is unavoidable, add front fill light rather than relying on the camera's wide dynamic range. WDR helps and does not solve it.
• Avoid harsh overhead downlighting directly above the capture point: it puts the eye sockets in shadow, which is where the most matching information sits.
• Watch for time-of-day variation. A west-facing entrance that works all morning can fail every afternoon in summer, and a site survey run once at 10am will not find it.
• Infrared illumination gives consistent capture in darkness, but IR images and visible-light enrolment images do not always match well. If IR is used for capture, enrol under IR too.

How does subject movement change the requirement?

Everything above assumes the person is roughly still or walking slowly. Motion adds two problems: blur, and the fact that a moving subject presents a usable frontal view for a shorter time.

• Shutter speed has to be fast enough to freeze the face. In low light this trades against noise, and the trade is usually resolved by adding light rather than by slowing the shutter.
• A capture zone the subject passes through in under a second gives very few usable frames. Widening the zone or slowing the flow — a door, a turnstile, a queue rail — buys frames.
• Corridors work better than open concourses because they constrain both the path and the direction people face.
• For gates, the geometry is easier than it looks: the person stops, looks forward at a screen and is at a known distance. That is why gate accuracy and corridor accuracy are not comparable numbers.

What should a site survey record?

A survey that records only mounting positions will not predict performance. The list below is what makes a survey useful enough to design against, and it is short deliberately — a survey nobody completes is worth less than a brief one everybody does.

1. For each capture point: mounting height, tilt angle, horizontal offset from the walking line, and distance to the capture zone.
2. Measured pixels between the eyes at the near and far edge of the capture zone, from the actual camera, not from a specification.
3. Light levels at the capture point at the three worst times of day, and a note of every window, glass door and light fitting in the field of view.
4. The expected direction of travel, and whether people stop.
5. The population: whether it includes children, wheelchair users, people wearing lawful head coverings, and whether enrolment images exist or will be captured on site.
6. What happens when recognition fails — the staffed fallback. If there is no answer to this, the design is not finished.

The last two items are the ones most often skipped and most often decisive. A system that recognises 95 percent of a population still needs a plan for the other 5 percent, and that plan sizes the staffing, which sizes the business case.

What does good placement not fix?

Placement raises the ceiling on what an algorithm can do. It does not address enrolment quality, threshold policy, or what an operator does with a match — and a deployment can fail on any of those with perfect cameras.

• Poor enrolment images cap performance permanently. A gallery built from low-quality scans will underperform whatever the capture side does.
• Threshold choice is a policy decision, not a technical one: it trades missed matches against false alarms, and the right setting depends on which costs more at that site.
• Operator response under time pressure is where automation bias appears. A high score shown to someone with a queue behind them is a request to disagree with a machine, and people do that less often than governance documents assume.