Foot Traffic Counting Methodology: How the Numbers Are Made

A number you cannot inspect is a number you should not trust. This is the StreetProof foot traffic counting methodology handbook — the full, plain-language account of how raw video becomes a verified pedestrian count, from the first frame to the final Location Score. It is written for entrepreneurs and brokers who want to understand what they are relying on, and for technical reviewers auditing the method. Nothing here is a black box: detection, tracking, line-crossing, deduplication, aggregation, sampling and extrapolation, plus the honesty rules that govern every figure we publish.

This is the reference the rest of our blog points back to, and the shared engine-trust document our sister brands link to rather than duplicate. Start with the foot traffic study guide if you want the decision-level view first.

Key takeaways

• Counting runs in four stages: detect, track, cross a line, aggregate — then sample and extrapolate.
• Each person gets a stable identity, so crossings are counted once per direction, with joins stitched across video segments.
• Categories beyond "person" exist, but beta ones are labelled, never overstated.
• Every projection carries a confidence interval and a plain-language confidence tier — and we never publish a number we did not measure.

Stage 1: Detection — finding the people

Every video is a stack of frames. The first job is to find, in each frame, where the pedestrians are. A detector scans the frame and returns a box around each person it is confident about, along with a confidence value. It also recognises a few broad categories — for example bicycles and dogs — so the mix of who is passing can be reported.

Two honesty points matter here. First, detection is probabilistic: the detector has a confidence threshold, and very faint or heavily obscured figures may be missed, which is one reason we report uncertainty rather than false precision. Second, detection is about shapes, not identities — the system is looking for the outline of a person to count a crossing, not for a face. That is the foundation of the privacy design explained in is video people counting GDPR compliant.

Stage 2: Tracking — linking frames into paths

A box in one frame is not yet a person walking; it is a single snapshot. Tracking links a person's boxes across consecutive frames into one continuous path, and gives that path a stable identity for as long as it is visible.

Good tracking is what makes counting trustworthy. It uses motion prediction to follow someone smoothly, and a two-stage matching approach so that when a pedestrian is briefly hidden — behind a lamppost, another person, a passing van — their track can be recovered on the other side rather than being lost and restarted. A track has to persist for a minimum number of frames before it is eligible to be counted, which filters out flickers and false detections. The result is a clean set of real trajectories, each belonging to one person.

Stage 3: The counting line — turning paths into counts

A count happens when a tracked path crosses a counting line you have drawn. The line is a virtual tripwire; the moment a person's trajectory passes from one side to the other, that is one count, tagged with the direction of travel. You can place up to three lines in a scene, and you can mark exclude zones — areas such as bus stops or terraces whose occupants are ignored so loiterers do not pad the total. How to place these well is covered in how to draw a counting line.

This line-crossing method is deliberately simple and auditable. Anyone can watch the overlay, see a box cross the line, and see the counter tick up. There is no hidden model deciding who "counts" — it is geometry you can check with your own eyes.

Stage 4: Deduplication — counting each person once

The subtle risk in any counter is double-counting. We guard against it in two ways:

• Once per direction, per short window. A given tracked identity is only counted once per direction within a brief time window, so someone hovering on the line does not rack up phantom crossings.
• Stitching across segments. Long videos are processed in segments for scale. Where a person walks across the join between two segments, the tracks on either side are matched and merged so they are recognised as one person, and any duplicate crossing at the boundary is removed.

The effect is a count of distinct crossings — the honest definition of foot traffic.

Stage 5: Aggregation — building the totals

With clean, deduplicated crossings in hand, aggregation rolls them up: totals by hour, by day, by direction and by category, on fixed time boundaries. This is where the raw events become the hourly profile, the weekday/weekend shape and the direction split you read in the report. Everything downstream — benchmarks, the Location Score — is computed from these aggregates, from the database, not re-derived by hand.

Stage 6: Sampling and extrapolation — days into an estimate

You rarely film every hour of every day, so the final stage projects a representative sample to a full picture. This is extrapolation, and it is where honesty is easiest to lose and most important to keep.

We scale observed counts to daily and longer figures using a stated statistical rule, and we attach a confidence interval whose width grows as the sample shrinks — the less you observed, the wider and more cautious the range. The projection is scaled by how much of the period was actually covered, so a partial sample is never quietly treated as complete.

Crucially, that interval models counting noise, not whether the hours you filmed represent the whole day. So on top of it we apply a plain-language confidence tier:

• Low — a spot reading. Too little footage to project a day. We report what was counted and offer no monthly projection.
• Medium. A projection is offered with a clearly wider interval.
• High. Enough footage, or enough distinct days, to make a daily and monthly estimate dependable.

The full reasoning, and how it shapes the Location Score, is in how accurate is video people counting and reading your Location Score.

The honesty rules behind our foot traffic counting methodology

Method is only as good as the rules that constrain it. Ours are simple and non-negotiable:

• Never publish a number we did not measure. If a component of the score cannot be computed honestly from the data, it is left blank and the rest renormalised — never filled with a flattering default.
• Label uncertainty, do not hide it. Confidence intervals and tiers are shown, not buried. Beta categories are marked as beta.
• Show the work. Every study ships a 60-second annotated overlay and a public QR-verified page listing the method and model version, so a counterparty can check the count instead of trusting it.
• Count silhouettes, not people. No faces, no identities — measurement without surveillance.

That is the whole pipeline, end to end. If you want to see it produce a real number for your street, start a $49 Spot Check, or compare the full study tiers on the pricing page.

Frequently asked questions

How does video people counting work? In four stages: a detector finds each pedestrian in every frame, a tracker links those detections into a continuous path, a virtual counting line records each path that crosses it by direction, and an aggregation step rolls the crossings into totals by hour, day and category — then sampling projects them with a confidence interval.

How do you avoid counting the same person twice? The tracker gives each person a stable identity across frames, and a crossing is only counted once per direction within a short time window. Tracks are also stitched across video segment boundaries so someone walking through a join is not double-counted.

What categories can you count? People are the core count, split by direction. We also detect broad categories such as children, bicycles and dogs. Some categories — for example strollers and wheelchairs — are marked beta and may not appear in every study; we label them rather than overstate them.

How do you turn a sample into a daily or monthly number? By extrapolation with a stated confidence interval that widens as the sample shrinks, scaled by how much of the period you actually observed. A clip too short to represent a day is labelled a spot reading and given no monthly projection.