How Accurate Is Video People Counting? MAPE, Confidence Intervals and Ground Truth

How accurate is video people counting? It is the right question and the most abused one. Every vendor quotes a shiny figure — 98%, 99.5% — yet almost none tell you under what conditions, against what truth, or with what margin, so "how accurate are people counters" gets answered with marketing instead of measurement. This is the honest explainer: what accuracy really means, what MAPE and confidence intervals are, why a single headline percentage is usually a red flag, and where mobile panel estimates go blind on the exact sidewalk you are trying to measure. If you are about to trust a footfall number, read this first.

It sits alongside the foot traffic study guide and feeds directly into how to audit a footfall claim.

Key takeaways

• A single "99.5% accurate" claim with no conditions attached is marketing, not measurement.
• MAPE tells you average percentage error — but only means something when the ground truth is disclosed.
• A confidence interval is the honest way to state a count from a sample; a number with no interval overstates its precision.
• Mobile panel data models an area from phone samples; it cannot count your exact doorway. Video of the real street is ground truth for that address.

Why "99.5% accurate" tells you almost nothing

Accuracy is not a property of a product; it is a property of a measurement under conditions. The same counter that nails a bright, uncrowded pavement will struggle in heavy rain, dense crowds, low light or a bad camera angle. So a headline "99.5%" with no context is not wrong so much as meaningless — it does not tell you which conditions produced it, whether it was measured on the vendor's easiest test clip, or what to expect on your footage.

Worse, most such claims ship no evidence at all. You cannot see the test footage, the ground truth, or the error on hard scenes. A number you cannot check is a number you cannot trust — which is exactly why we ship a 60-second annotated overlay with every study, so you can watch the counter work and judge the hard frames yourself.

Ground truth: the thing accuracy is measured against

"Accuracy" only exists relative to a ground truth — the real answer. In counting, ground truth usually means a careful human count of the same footage, frame by frame, that the automated count is compared against. Every honest accuracy figure is really a comparison: automated count versus ground truth, on a specific clip.

This is why where the number comes from matters more than the percentage. A count taken directly from video of the actual street is ground truth for that address — you are counting the real people who really passed. An estimate modelled from somewhere else is, by definition, one step removed from the truth it claims.

MAPE, in plain language

When people quantify counting error, one common measure is MAPE — mean absolute percentage error. In plain terms: across a set of tests, how far did the count land from the truth on average, expressed as a percentage. If the true count was 100 and the estimate was 90, that is 10% error; MAPE averages that across many tests.

MAPE is useful because it is comparable across big and small locations — a percentage travels where a raw count does not. But it comes with a catch worth stating loudly: MAPE is only meaningful when the ground truth it is measured against is disclosed. A vendor can quote a flattering MAPE from easy footage and hide it from hard footage. Ask what conditions the figure covers, and treat any refusal as an answer in itself.

We deliberately do not headline a single global MAPE, because it would imply a precision that does not survive contact with real streets. Instead, every study carries a per-report confidence interval and a plain-language confidence tier — honesty at the level of your footage rather than a marketing average.

Confidence intervals: how to state a count honestly

Every count taken from a sample is an estimate, and estimates have uncertainty. A confidence interval is the honest way to express it: a range the true value is likely to fall within, given how much you observed. "About 1,200 a day, likely between 1,000 and 1,400" tells you far more than a lone "1,214" — because the second version pretends to a precision no sample can deliver.

Two rules follow:

1. More data, tighter interval. The more crossings we observe, the narrower the range. Short clips get wide intervals; representative multi-day samples get tight ones.
2. The interval is not the whole story. A confidence interval captures the randomness of the sample — the counting noise — not whether the hour you filmed represents the whole day. That second kind of uncertainty is why we add a confidence tier: a very short clip is labelled a spot reading and gets no monthly projection at all. The full logic is in reading your Location Score.

A report that gives you an interval and a plain-language tier is telling you the truth about how much to lean on it. A report that gives you one round number is asking you to trust it blindly.

Where mobile panel data goes blind

The other big source of footfall numbers is mobile-phone panel data — platforms that infer visits from a sample of app-tracked phones and model them across an area. This has real strengths: national chain benchmarks, trade-area comparisons, long-run trends. But for a single doorway it has structural blind spots:

• It models an area, not your door. A panel estimates movement across a block or catchment; it cannot count the specific people who pass one storefront.
• Coverage is uneven and thin where privacy rules bite. In much of Europe, phone-tracking panels are sparse, so a "visit" estimate for your street may rest on very few real devices.
• Estimates ≠ counts. A modelled visit figure is a projection with its own hidden error; a video count of the actual pavement is the ground truth that projection is trying to approximate.

None of this makes panels useless — they answer a different question, at a different altitude. But when the decision is "how many people pass this address," the honest tool is a ground-truth count of that address. That contrast is the heart of auditing a footfall claim.

How accurate is video people counting, then?

Accuracy you can trust looks like this: disclosed method, disclosed conditions, a ground truth you can inspect, a confidence interval instead of false precision, and evidence you can audit. That is the bar we hold ourselves to — and the bar you should hold anyone to before their number changes your offer.

See how it comes together in a real report, or start a $49 Spot Check to get a count of your own street with its uncertainty shown honestly. Full study options are on the pricing page.

Frequently asked questions

How accurate are people counters? It depends entirely on conditions and on how honestly accuracy is reported. A single headline figure like "99.5% accurate" with no conditions attached is a marketing number, not a measured one. A trustworthy answer states the method, the test conditions, and a margin of error for your specific footage.

What is MAPE in foot traffic counting? MAPE is mean absolute percentage error — on average, how far the counts land from the true number, as a percentage. It is useful because it is comparable across locations, but it is only meaningful when the ground truth it is measured against is disclosed.

What is a confidence interval and why does it matter? A confidence interval is the range a true value is likely to fall within, given a sample. It matters because every count from a sample is an estimate, and an interval tells you how much to trust it. A number with no interval is pretending to a precision it does not have.

Why can't mobile panel data see my exact sidewalk? Panels infer movement from a sample of phones and model it across an area. They are strong for trade-area trends but cannot count the specific people passing one doorway, and coverage is thin where privacy rules limit tracking. Counting from video of the actual street is ground truth for that address.