Methodology

Rendered from the engine's METHODOLOGY.md at commit 6b351b02ca7b

Draft. This describes the method as designed. Measured accuracy figures will be added once the engine has run in production. Changes are listed at the end.

What we measure

For each sweep of London’s public traffic cameras we count:

• people detected
• people with an outer layer (coat or jacket), or without one
• people with bare legs (shorts or a short skirt), or without
• open umbrellas

Each sweep is tagged with the weather at that time: temperature, feels-like temperature and precipitation. Aggregating many sweeps gives the share of people wearing each item at each temperature, for example “at 12 °C, 71% wore a coat”.

We do not measure anything finer than these coarse attributes (sleeve length, colours, brands). At 352×288 pixels they are not reliable, and the product does not need them.

Pipeline

Privacy

1. Camera frames and person crops exist only in memory. They are never written to disk, logs, caches, CI artifacts, git or analytics services.
2. No face detection, face recognition, re-identification, or tracking of anyone across frames.
3. Only aggregate counts are published. The site never shows camera images or crops. To see a camera, follow the link to TfL’s own camera page.
4. If a vision model is used for attribute labelling, it receives person crops only (never full frames), from a provider with data-processing terms, within a fixed monthly budget, and only the resulting labels are kept.
5. Accuracy checks are done by hand, live, on a sample of current frames. A person views the crops in a window on their own computer; nothing is saved, and only the tallies and each box’s height in pixels are kept. Before launch at least 300 boxes are checked, across daylight, dusk and rain; after that, 100 a month and after every change to the detector or its thresholds. For one-off tests of automatic clothing labels, the same crops of single people may be sent, in memory, to an open-weights vision model at a hosting provider whose terms exclude storing, logging or training on them. The spot-check tool refuses to run in CI.

CI enforces rule 1 in two ways. A static check blocks image-writing calls in engine/. An end-to-end test runs a full sweep against a local fake camera server and checks that it creates no files except the aggregate output and leaves no image bytes in its working, home or temporary directories. Deliberate leaks in the test suite prove that these checks catch them. The detector’s runtime library (ONNX Runtime) has its built-in telemetry switched off, and a test checks that loading it writes nothing.

Sampling and known biases

• Cameras watch roads, not pavements. In early tests about 40% of cameras showed at least one person and 5–7% showed five or more. People close to the camera are over-represented.
• Distant people are missed. In a manual check of one frame every detection was a real person, but about half of the visible (mostly distant) people were not detected. That check used the earlier YOLOX-s model; YOLOX-m, used since September 2026, detects 30–40% more people on the same frames, and its precision is measured by the spot-checks. This lowers the count but should not bias the share wearing a coat, because detection depends on distance, not on clothing. We will test this assumption with spot-checks.
• One city. London’s population is not everyone’s. Clothing norms at a given temperature differ between cultures and climates. Pages say where the data comes from.
• Near-field people only. We count and classify only people whose box is at least 31 pixels tall in a 352×288 image: the smallest size at which most boxes can be checked by eye and at least 90% of them are people. Smaller boxes are ignored. The threshold is recalculated when the hand checks grow.
• Daytime only. Sweeps run in London daytime, and only sweeps that start while the sun is no more than 6° below the horizon are counted.
• Weather is a forecast for one point in central London (the Met Office hourly time step nearest the sweep), not a measurement at each camera.

Uncertainty and minimum sample

• Every percentage is shown with its sample size n, time range and source.
• Every percentage has a 95% confidence interval (Wilson score interval).
• Below n = 200, pages show “collecting data (n observed so far)” instead of a percentage. Nothing is imputed, padded or estimated.

Validation

• Detection precision: at least 300 boxes checked by hand before launch, then 100 a month and after every change to the detector. Target ≥ 90% for near-field boxes; below 85% raises an alert.
• Attribute accuracy: at least 200 near-field people compared with human labels. Target ≥ 85% per attribute before an attribute is published.
• Umbrellas: validated during rain (at least 30 umbrellas).

Results will be published here with dates and sample sizes.

Sources and attribution

• Camera images: Transport for London JamCam feed. Powered by TfL Open Data.
• Weather (UK): Met Office Weather DataHub. Powered by Met Office data.
• Weather (US visitors): National Weather Service (public domain).
• Detector: YOLOX by Megvii, Apache-2.0.

Changelog

• 2026-09 — Method drafted; feasibility tested with two full sweeps (822 and 931 people).
• 2026-09 — Fetcher and weather adapter built. Facts updated: one attempt per camera per sweep (no retries), feels-like temperature recorded, weather point described exactly.
• 2026-09 — Detector switched from YOLOX-s to YOLOX-m after a same-frame comparison on 200 live frames (30–40% more people detected, about 190 ms per frame).
• 2026-09 — Accuracy checks redesigned: an open-weights vision model on the same machine judges detection crops in memory, with known-answer crops mixed in as controls. The manual mode remains.
• 2026-09 — Accuracy checks are done by hand: the automatic judge did not reach the required accuracy (it never recognised a box that was not a person). Counts are limited to near-field people.