Pilot guide
Agree a workflow, representative footage, evaluation measures and review responsibilities. Expand only after the pilot establishes what works and where the limits are.
- 1Main gate · 22:41 · plate read
- 2Campus road · 22:43 · candidate, no plate
- 3Loading bay · 22:46 · yard closed
- 4Waiting for a reviewer
How a pilot runs
Six steps, agreed before anything is installed.
A pilot is a test you can fail. It is scoped so that the result tells you something.
Choose one workflow
One problem, stated in a sentence, with the people who own it. The pilot is built on what is in the current release; a capability not yet confirmed on live footage in our own testing is evaluated separately, with its own measure.
Choose representative footage
A small set of cameras, or exported clips, that include your hard cases: night, rain, distance, crowds. A pilot on easy footage tells you little.
Confirm the sources
We check that each camera or file connects, note the ones that do not, and agree the machine it will run on.
Agree the measures
Pick the measures below that fit the workflow, and write down what result would make you expand, change or stop.
Agree who reviews
Named people who will mark alerts and label a sample. Without review, the false-alarm rate and the read rate cannot be measured.
Run, review, decide
Run for an agreed period, review the measures together, and decide on the evidence.
Evaluation methodology
Measures, each with a definition.
Defined so that two people would compute the same number. None of these is a result: the values come from your footage.
- False-alarm rate
- Alerts marked False positive divided by alerts reviewed, per rule and per camera.
- How it is measured: From the alert queue, once your reviewers have marked the alerts. Unreviewed alerts are excluded, so a neglected queue shows no rate rather than a good one.
- Plate read rate
- Vehicle sightings with a readable plate divided by all vehicle sightings, per camera, split by day and night.
- How it is measured: From plate statistics, checked against a hand-labelled sample. Relevant only where plate reading is part of the workflow.
- Candidate precision
- Of the sightings grouped or matched as candidates, the share a reviewer accepts as the same person or vehicle.
- How it is measured: A reviewer labels a sample of groups against their stills. Measured on a sample. Review decisions are not yet stored in the product.
- Missed events
- Events your team knows happened that raised no alert.
- How it is measured: Compare a hand-kept list for the pilot period with the alerts raised. The only way to find what was missed is an independent list.
- Time to acknowledge
- The median time from an alert being raised to its being marked by a person.
- How it is measured: From alert statistics. Measures your process as much as the product.
- Time to first useful evidence
- From the start of a search to the first sighting an investigator notes as relevant.
- How it is measured: Timed by hand in the pilot; the audit log gives the time of a plate, scene or route search. Compare with how long the same task takes today.
- Coverage health
- The share of cameras online, the cameras offline for more than an hour, and each camera's delay between capture and analysis.
- How it is measured: From camera health. A camera that is behind or offline cannot raise alerts.
What our current evidence establishes, and what it does not
Established
- On one outdoor camera recorded over a single day, 53 exported clips were analysed, placed on one timeline at their recorded times, and played back with the gaps skipped.
- In that footage, 116 sightings were grouped into 27 candidate people and vehicles for review.
- Zone-entry, item-left, count and scene-reading rules, their alerts and triage have been exercised on live streams in our own testing, on one workstation.
Not established
- Accuracy on your cameras, in your lighting and at your distances.
- Plate read rates. On night and dawn street footage only a small share of vehicles produced a usable read, and daytime rates have not been measured.
- That a plate watchlist alert, a line-crossing or dwell rule, or a route with more than one stop works on a live feed: these are verified by automated tests only.
- Operation on a server or in a cloud account, and connection to a camera directly rather than through a gateway.
- Any reduction in investigator hours. Skipping the gaps between clips shortens playback; it is not a measured saving in labour.
Sharing footage safely
This website never asks for footage, and neither does the demo form.
- For a pilot in your environment, footage does not leave it: the platform runs on your machine.
- If you want us to look at sample footage, we first agree in writing how it is transferred, who may see it, what it is used for and when it is deleted.
- Do not send footage, images of people, number plates or camera passwords by email or through the form.
- Choose sample footage you have the right to share, and remove anything you would not want reviewed.
Questions we are asked first
How many cameras should a pilot include?
Few enough to review properly. A single workstation analyses a handful of cameras at once; the right number depends on resolution, frame rate and which modules are on, and we size it with you.
What do we need to provide?
Stream addresses or exported clips, a machine to run on, named reviewers, and an hour a week to look at the measures together.
What happens if a measure comes out badly?
Then you have learned something before buying. We would rather change the scope, the camera positions or the rules, or stop, than expand on a weak result.
Can a pilot evaluate a capability you have not yet confirmed on live footage?
Yes, as a separate evaluation with its own measure, and not as something your operation depends on.
Plan a pilot on your own conditions.
Bring one workflow and a few representative cameras. We will agree the measures and the reviewers before anything is installed.