A camera missed a vehicle: what the supervisor should check first

By Gate & Shift Editorial Desk · Planned week of 2026-11-23 · Documentation reviewed 2026-11-15

Someone says a van arrived at 02:17, but the alert log is empty. Before changing anything, preserve the permitted footage and the operator's time note. You need to know which part of the chain failed — turning the sensitivity up right away can bury that answer.

This is a proposed investigation procedure. It does not diagnose a particular vendor or establish a product error rate.

Start with the camera view

Write down the camera, location, clock and reported event interval, then ask the witness what they actually saw. Was the van in the camera's view, behind an obstruction, or on another approach? Keep “a vehicle was on the property” apart from “the defined vehicle event was visible on this camera.”

For scoring, use a reference event with the target activity, camera and start/end times. NIST's activity-evaluation material illustrates why the event boundary matters M01. A delivery arriving at the site and a vehicle stopped on a marked shoulder are not interchangeable reference events.

Check whether the system was observing

Look for a feed interruption, disabled detection, an excluded schedule or an unexplained gap, and preserve the available health evidence. A reference event that falls inside an observation gap should be reported as such: silently dropping it can make a system look better, while calling every gap a model error hides the operational problem.

Keep two views of the result: end-to-end scheduled coverage and performance during eligible monitored time. The house definitions treat coverage as monitored time divided by scheduled time, with gaps recorded explicitly M02. That proposal normalizes differently from ActEV's time-based false-alarm metric.

Reconstruct the saved rule

Use the version active at 02:17 and record the class, schedule, zone and duration condition. If the rule needs a vehicle to remain for thirty seconds and the visible visit lasts twelve, a missing alert does not by itself show the rule failed. Those durations are hypothetical examples, not a feature claim.

Check whether a camera move changed the physical area inside the drawn zone, and take a screenshot of the tested view under your evidence policy. The next shift should be able to read the geometry without rebuilding it from memory.

Follow the event beyond the detector

If the event exists in the application but no recipient saw it, move to delivery troubleshooting. Keep application creation time, send time, receiving endpoint time and acknowledgment time separate. Our proposed alert-delay measure begins when the annotated rule first qualifies and ends at the chosen receiving endpoint M02.

That separation matters during an incident review. “The camera was late” could otherwise cover a slow rule, a slow network, a rejected webhook or a person away from the desk.

Turn the finding into a small repeatable test

Use permitted labeled recordings or an approved on-site exercise. Keep the camera and rule stable, repeat the relevant activity and list every reference event. Match system events using a rule declared before scoring M03.

Say eight of ten eligible reference events matched. Recall is 8/10, or 80%, in that small hypothetical sample — it is not “80% accurate” because false alerts need their own count. Undefined denominators stay undefined M04. The measurement buying guide develops the calculation.

FAQ

Should I delete the failed trial and start again?

Keep its evidence and label the change. A failed observation can tell the installer exactly which condition needs attention.

When can we call it a missed detection?

After establishing an eligible reference event, observation coverage, a qualifying rule and no matching system event under the declared matching procedure. Record unresolved cases separately rather than forcing a verdict.