Camera audience analytics works when it counts faces and measures dwell or attention time; the 2013 field research validated the measurement. It rarely pays when it tries to target demographics or fire on attention, because viewers are not looking and operators will not fund the fees. If you add a camera, trigger on a face count crossing from zero to one or more.
Camera audience analytics on a sign means a webcam plus software that counts faces and times how long people look. Vendors label it AI. The working parts are a face detector, a frame counter and a set of timers, and the evidence splits cleanly down those parts. Counting and timing work: a 2013 peer-reviewed study measured dwell time, display in-view time and attention time with computer vision on a live signage installation. Making money from demographic targeting rarely works: the estimates are the weakest metric, the fees are real and few operators keep it past a trial. This page explains what the camera measures, the one trigger worth building, and what changes on a real sign. If the analytics needs a host, start with the guide to choosing a digital signage player.
What the camera measures
The software watches the frame for faces and reports five numbers. The names below are the ones the measurement literature uses, and the 2013 study published its timing metrics as measurements taken on a live signage installation, so the published numbers are field measurements.
| Metric | What the software counts | What it is good for |
|---|---|---|
| Face count | Faces in frame right now | Arrival triggers, footfall at the screen |
| Dwell time | Seconds a person stays inside the camera's zone | Compare locations, find where people stop |
| Display in-view time | Seconds a person faces toward the screen | Estimate exposure for a spot on the loop |
| Attention time | Seconds a full face points at the screen | The closest measure of actual reading |
| Demographic estimate | Age range and sex guessed from face geometry | Targeted content, and the least accurate of the five |
Only the first four are measurements. The fifth is a classification from face geometry, and it is the one that drags privacy rules into the project.
What the numbers are good for
Measurement is the use the evidence supports. Programmatic out-of-home advertising runs on standardized definitions for impressions and audience measurement, because advertisers buy those numbers, so a network that sells airtime needs measured audiences. A single store screen answers smaller questions with the same metrics: did anyone stop at the screen today, how long did they stay, did the new loop hold attention longer than the old one. Those answers set placement and loop length, and that is where most operators collect the value. Steering the content itself is the open case, and the next two sections take it.
The trigger that works: a face count from 0 to 1 or more
Build the simplest content trigger that can still change something: when the full-face count crosses from 0 to 1 or more viewers, restart the loop from its first frame or swap in the counter promo. Arrival is the one audience event a camera detects reliably, because the threshold is a single face pointed at the lens. Tools that offer this expose it as a condition on the live face count bound to a layout change; there is nothing smarter to configure and nothing smarter needed.
It earns its keep where someone stops close to the screen: a counter or endcap unit, the kind in our small retail displays. It also works as an arrival reset on a lobby or waiting-room screen. Give the triggered piece a fixed run, 90 seconds for example, then hand playback back to the loop, so one lingering face cannot park the screen on a single layout.
Why the clever triggers fail
An attention trigger fires when a full face points at the screen, and full faces are scarce. A standing viewer glances for a second or two. A seated audience looks at its phones. The screen collects eyes only when something on it has already pulled them in, which is why field reports of look-based triggers describe them firing during popular live events and doing nothing through ordinary programming. A trigger that needs the content to succeed first has the logic backwards.
Demographic targeting fails on cost and accuracy together. Age and sex estimates are the least accurate metric in the table, every screen then needs a camera, an analytics subscription and a consent review, and demand has stayed thin: the firms selling these systems report few operators who keep any of it once the trial bill arrives. Use the camera to measure, and let the schedule carry the message.
Example: will a 1080p webcam resolve faces at your counter?
Assumptions, stated so you can swap in your own: a 1920 x 1080 webcam with a 78° horizontal field of view, a typical consumer figure; an adult face about 15 cm across at the cheekbones; and the classical Haar face detector that ships with open-source vision libraries, which scans 24 x 24 pixel windows, so a face under about 24 px across falls below the detection floor.
Stand 1.5 m from the camera. The frame width is 2 × tan(39°) × 1.5 m = 2.43 m. That is 1920 px / 2.43 m = 790 px per metre, so a 15 cm face spans 0.15 m × 790 px/m ≈ 119 px, five times the 24 px floor. Detection is comfortable.
Now push the camera out. A face hits the 24 px floor when the frame is 1920 px / (24 px / 0.15 m) = 12 m wide, which at 78° is 12 m / 1.62 = 7.4 m away. At the floor, detections come and go, so halve it: plan a camera range up to about 3.7 m, which is a frame 6 m wide.
Decision: a counter, an endcap or a doorway zone works with any 1080p webcam on a short bracket. A lobby camera 8 m from the seats under-counts; move the camera closer to the stop zone or fit a narrower-angle lens so the frame stays within 6 m.
What it changes on a real sign
The display side of the project is unchanged. The camera only decides what plays, and the viewer still has to read it, so keep the letter-height plan: 1 inch of capital height per 30 ft of viewing distance (FHWA), or 1 inch per 20 ft for fast reading. The screen size and viewing distance guide covers the rest of that layout work.
The player side is where camera projects stall. The analytics needs an operating system that hands the camera to its SDK, and signage players are built to push pictures out. Taking a live signal in is the rare case, the same wall as why HDMI input rarely works on Android signage players. The usual host is a small x86 PC beside the screen.
Our panel-mount screens are offline players: images and video from microSD or USB, no subscription, and the HDMI port is an output only. There is no camera analytics software in that package and no video input to accept a feed from an external PC, so they are the wrong host for camera analytics. Use them for the scheduled loop they are built for, and put the camera stack on a separate PC feeding a screen that takes HDMI in.
Mounting is short work: camera at or just above eye height on the screen bracket, aimed at the stop zone, with a passive USB 2.0 cable good for 5 m and an active extension beyond that. Keep bright windows behind the audience out of the lens; a backlit face turns into a silhouette and the count collapses.
Privacy is one design rule: process faces in memory, store counts only. If nothing is kept, the project stays clear of the heaviest biometric rules in most jurisdictions; the FAQ below names the two laws that decide it. Add one line of notice on screen or at the counter: this camera counts visitors, nothing is stored.
The decision, case by case
- Counter, endcap or fitting-room screen with people stopping within about 4 m: a 1080p webcam, the 0 to 1+ face-count trigger, counts kept as footfall. Low cost, real placement data.
- Lobby or waiting room with a seated audience 5 m or more away: skip the attention trigger. Use the count trigger as an arrival reset or just run a schedule.
- Network selling airtime to advertisers: measured audiences are the currency of programmatic out-of-home, so measurement pays at that scale. Buy proper camera placement, not counter webcams.
- Demographic targeting on a small screen: build it only when a measured trial shows the targeted content change is worth the camera, the subscription and the consent work. Demand for it has stayed thin.
Recommendation: If you add a camera at all, trigger on the full-face count crossing from 0 to 1 or more viewers, run it on a player whose operating system supports the camera in software, keep the camera within about 4 m of the stop zone so faces sit five times above the 24 px detection floor, store counts and discard frames, and leave demographic targeting until a measured trial pays for the camera, the subscription and the consent review.
Questions
Do any laws regulate counting faces with a camera on a sign?
Counting anonymous faces is treated lightly in most jurisdictions, but identifying them is not. Illinois' Biometric Information Privacy Act requires informed written consent before a private entity collects biometric identifiers such as face geometry, and it lets people sue over violations. In the EU and UK, GDPR treats biometric data used to uniquely identify a person as special-category data that needs explicit consent. If your software keeps face images or templates, assume the consent rules apply and get advice for your jurisdiction.
Can a card-based offline signage player run camera analytics?
A player that only shows images and video from a microSD card or USB stick has nowhere to run the vision software and no way to read a camera. Camera analytics needs an operating system that exposes the camera to an analytics SDK, which usually means a small x86 PC next to the screen, feeding the display over HDMI. Plan the PC's mounting and power with the screen, and expect the analytics software itself to be a subscription.
What camera do I need for face counting at a counter?
A 1920 x 1080 webcam at 30 frames per second over USB covers counter work; the example above shows a 1080p sensor carries five times the pixels it needs across a 2.4 m frame. Spend the budget on placement: mount at or just above eye height, aim at the stop zone, and keep window backlight off the lens. Higher resolution only matters once the frame has to cover more than about 6 m.
Planning information, not legal or engineering advice. Rules differ by place: check codes and permits with your local authority and your licensed trades.