The incident nobody has footage of: auditing coverage after the fact

The incident is reported, somebody goes to pull the clip, and there is nothing usable. That is a finding, not the end of the investigation, and it has 5 distinct causes with 5 unrelated remedies. The spot was never inside any field of view. It was, but something was standing in the way. It was, but the camera had been knocked out of aim months ago. The camera was fine and the recorder was not. Or the area is excluded on purpose and everyone had forgotten.

Which one it turns out to be decides whether you are ordering hardware, moving a pallet-stacking rule, tightening a maintenance procedure, or fixing storage. Getting that wrong is expensive in a specific way: sites buy cameras for coverage they already had and never fix the reason the coverage was not usable.

Fix the spot and the minute before you open any recorder

Almost every failed footage pull starts with a search on the wrong time or the wrong aisle. Reported times come from memory, from a clock on a wall, or from a form filled in after the shift, and they are routinely out by 10 to 20 minutes. Reported locations are usually a zone name rather than a position.

1. Walk to the spot with the person who reported it and mark it physically. A bay number, a rack upright, a floor marking: something a camera view can be aligned against later.
2. Record it on the site plan with a coordinate, not a description. A dot on the plan is what the next step needs, and an aisle name is not.
3. Widen the time window deliberately. Take the reported time and search plus or minus 20 minutes as a starting point; narrow it only once you have found the sequence.
4. Anchor the clock. Compare the recorder's timestamp against a known-good clock and note the drift, because a recorder that has been running unsynchronised for a year can be minutes off on its own.
5. Write down what you expect to see. Two people walking in from the north door and a truck reversing gives you something to search on when the exact moment is not visible.

Which cameras could have seen it, and which could have resolved it

From the dot on the plan, list every camera whose field of view could plausibly include it, including ones aimed at something else nearby. Then pull a still from each, from the incident window, and look at the spot rather than at the scene. Coverage on paper and usable coverage are different claims, and the gap between them is where most of these investigations actually land.

The arithmetic is worth doing once so the argument stops being subjective. If the camera at that distance frames a vertical span of about 12 metres and the sensor gives you 1,080 lines, a 1.7 metre person occupies roughly 150 pixels of image height. That is more than enough to say a person was present and moving, and nowhere near enough to say whether they were wearing gloves. Sorting an existing estate by that test, per camera, is set out in reusing the cameras you already own for safety analytics.

A camera that covers a spot and a camera that can answer a question about that spot are 2 different claims, and only 1 of them is on the coverage map.

Five findings, and the very different remedies they carry

Classify before you fix. The evidence column is what separates them, and each one can be settled in minutes once the frames are in front of you.

Only the first row justifies buying anything. If the answer to every coverage question is another camera, that is a procurement reflex rather than an engineering finding, and it is one of the things worth testing in whoever is advising you, along the lines set out in choosing an AI development partner. It is also worth being honest about the incidents cameras will never resolve: a forklift near-miss behind a full rack line is invisible from any practical mounting position, and a proximity system on the vehicles measures what no lens can, as compared in camera versus wearable tag for forklift proximity.

The stock moved, the season changed, and the camera did not

Obstruction and drift are the 2 findings that hide behind a coverage map claiming the site is fully covered, because both were true when the map was drawn. A warehouse at seasonal peak stacks 2 pallets higher than it does in a quiet month, and the view that was clear in March is a wall of shrink-wrap in November. Cameras get nudged by scissor lifts, by mounting-pole vibration, by anyone working on the light fitting next to them.

• Capture a reference frame per camera at commissioning. One still per camera, dated, stored with the coverage map. It costs an afternoon and it is the only thing that turns drift from an argument into a diff.
• Diff quarterly and after any work at height. A scheduled job that pulls a current still and compares it to the reference is small automation of the kind we build under AI agents and automation, and it finds obstruction before an investigation does.
• Mark the mount. A paint line across the bracket makes a nudge visible from the floor without a ladder or a monitor.
• Treat peak season as a separate coverage state. If stock height changes the view for 8 weeks a year, the coverage map has 2 versions and both need to be signed off.
• Record the reason when a view is knowingly compromised. An obstruction nobody logged becomes a surprise during the one week it matters.

A recording gap is not a blind spot, and confusing them costs months

If the camera was aimed correctly and unobstructed but there are no frames, the fault is downstream and no amount of camera work will fix it. Prove it from the recorder rather than from the absence: a stream that stopped, a disk that filled, a network segment that dropped during a shift, motion-triggered recording that never triggered, or a retention setting that rolled the window before anyone asked for the clip.

The last one is the most common and the most avoidable. How long each class of footage needs to survive, and why only 1 of the 3 retention clocks needs to be long, is worked out separately in clipping, retaining and handing over safety footage. The rule that matters during an investigation is simpler: place a hold on the relevant cameras and the relevant window the moment an incident is reported, before anyone has decided whether the footage is useful.

Deliberate exclusions belong in the same conversation, because they produce identical symptoms. Welfare areas, changing rooms and any area covered by an agreement with worker representatives may be masked or unmonitored by design, and local privacy and employment rules constrain what may be recorded at all. Do not reason about that from memory: confirm the current position with the applicable regulator's guidance and the site's own agreements before proposing new coverage anywhere near it.

The quarterly walk that keeps this from recurring

The audit above is reactive. Run it as a standing check and the same finding surfaces before it costs anybody an investigation. Two hours a quarter is enough on a site with 40 cameras, and the output is a signed coverage map rather than a feeling.

1. List the hazards, not the cameras. Take the top 10 highest-consequence hazards on the site and check each one against the estate; a camera-first walk always ends with the answer that all cameras are working.
2. Stand a person at each hazard and pull the frame. If nobody can identify what that person is holding, record the limitation on the map rather than assuming the view is fine.
3. Diff every current still against its reference frame and open a maintenance job for anything that moved.
4. Check recorder health for the quarter: gaps per camera, disk headroom, and the oldest frame still retained.
5. Sign the map. A named person accepting that these 3 areas are not covered is a far better position than an assumption that everywhere is.

One boundary worth keeping clear: this is a safety-coverage exercise, and the camera specification questions that dominate an inspection cell are a different discipline with different answers, which is why area-scan versus line-scan camera choice is argued separately. Coverage work belongs to the safety, PPE and site monitoring programme, alongside the rest of what we build for manufacturing and industrial vision.

• video analytics
• incident investigation
• coverage
• safety