How the counting works

A Ystara unit is a small solar-powered computer with a camera, mounted several metres above the sand and aimed at a defined patch of beach. Every five minutes it wakes, takes a short burst of frames, runs a person-detection model on the device, counts what it finds inside the zone it is responsible for, and discards the frames. Nothing but numbers leaves the beach.

Taking a burst rather than a single frame matters: people move, umbrellas block, and a gull crosses the lens. The burst is reduced to one figure per cycle, and the spread across the burst becomes part of the confidence score.

A raw count is not yet useful — two hundred people is empty on one beach and packed on another. Each zone therefore carries a practical capacity: its usable area divided by a comfort standard in square metres per person. The occupancy of a zone is its smoothed count against that capacity, and the occupancy of a beach is its zones combined, weighted by how much of the beach each one covers.

The comfort standard is a choice, not a law of nature. Ystara uses roughly ten square metres per person for a typical summer day, with looser and tighter presets available per beach, and the figure in use is recorded against the site rather than hidden in a constant.

Coverage is the share of a beach that was actually observed in a given period. A beach with three zones where one camera is offline reports lower coverage, and the estimate for the whole beach is extrapolated from what remained visible.

This is the number that keeps the platform honest, so it is published alongside every figure rather than buried. When coverage is 0.4, six tenths of that beach was not seen and the estimate is a projection. When coverage is zero, the platform reports unknown — never zero. An unobserved beach and an empty beach are different answers, and only one of them is a reason to go.

Each reading carries a confidence value between zero and one, reflecting how much the device trusts its own count: the agreement across the burst, the detector's own certainty, the light, and known conditions such as an obstructed view. Low confidence sustained over hours is treated as a fault rather than a curiosity — the platform raises an alert, because a camera that is powered, online and quietly wrong is the most expensive failure there is.

It will not fill a gap with a plausible number. If a camera is silent, the history shows a gap; if a beach has never reported, it says so rather than showing a comfortable zero. There is no back-fill, no smoothing across an outage, and no forecast presented as a measurement.

It will not claim an accuracy figure it has not measured. A published accuracy number requires an independent evaluation set — a body of annotated frames spanning crowd density, distance and light — scored against the model. That work is under way, and until it is finished the platform reports confidence and coverage instead of a precision claim it cannot defend.

Signed-in visitors can report how busy a beach actually looked, and the platform stores that beside what it was publishing at that moment. Those reports accumulate into a per-camera verdict — agrees, reads high, reads low — and no verdict is shown below twelve reports, because “this camera disagrees with visitors” is an accusation about hardware and a handful of reports from one afternoon cannot support it.

If you want to check the numbers yourself, the data service gives you the occupancy series with its confidence and coverage attached, and corrections are the most valuable mail we get.