What a camera may call a near miss, and what still needs a human report

A camera can measure how close two things came, how fast they were closing, how long someone stood somewhere, and whether they crossed a line. A near miss is something else: a judgement that an unplanned event could have caused injury or damage and happened not to. That judgement rests on what the load weighed, whether the driver had seen the pedestrian, whether the machine was energised and what would have happened next — none of which is in the picture.

The practical consequence is a rule about naming. Camera output enters the near-miss process as candidates, and a person turns some of them into near misses. Skip that step and two numbers go wrong at once: the leading indicator inflates by an unknown factor, and the human reports carrying the actual context stop arriving, because the system appears to be catching things already.

One word, one safety definition, one sensor definition

The safety function defines a near miss by its potential consequence — an unplanned event that did not cause harm but could have. That definition is about the counterfactual, which is why it needs a person who knows the work. Your own incident procedure will have wording, and so will your regulator; the Health and Safety Executive in the UK and OSHA in the United States both publish guidance on near-miss reporting, and the wording that governs you is theirs, not a vendor's datasheet.

The sensor defines nothing. It reports geometry over time, and every safety meaning attached to that geometry was put there by whoever wrote the rule. That is an argument for writing the rule down, not against using cameras.

The four quantities, and the sentence each one cannot finish

All 4 inherit the geometry problem: they are computed in image space and mean something on the floor only through an assumption about the ground plane, the subject of what a restricted zone polygon really means. A separation of 1.4 m from a badly calibrated view is a number with no unit anyone should trust.

Three buckets, and only the middle one is contested

• Seen and reportable. A pedestrian steps out behind a reversing counterbalance truck and the driver brakes. The camera measured it, a person would report it, and the two agree.
• Seen and routine. Two operators pass within 1.2 m in an aisle designed for exactly that. The measurement is correct and the event is nothing, and this bucket is where an unclassified count does its damage.
• Reportable and invisible. A sling with a broken strand noticed before a lift, a mislabelled drum, a colleague working while unwell. No camera in the plant produces a candidate for any of these, and they are the reports worth protecting.

Candidates in, classified events out

1. The system raises a proximity or intrusion candidate with its measured values, an evidence clip and the rule version, in the shape set out in a PPE violation is a record with an owner.
2. It pre-ranks by severity — closing speed and minimum separation — so a reviewer opens the worst 20 rather than the oldest 20.
3. A named person classifies each candidate within a working day, from the clip, into 1 of 4 outcomes: near miss, observation worth coaching, normal operation, or false detection.
4. Near misses join the same register as human-reported ones, tagged by source, so the two streams stay comparable instead of merging into one total.
5. False detections route to engineering. Sorted by hour and camera they usually collapse into 1 or 2 causes — a dusk cluster is the day-night switch, a single-camera daylight cluster is closer to wearing a hard hat and still flagged.

That review queue is an operational tool with a throughput limit, and it is where these programmes fail. A queue where 4 in 5 candidates are noise stops being worked within a fortnight — the response that makes a rejection rate intolerable on a line, in what a false reject rate feels like. Sizing and instrumenting that queue is internal tools and ops work, not model work.

A camera can tell you two objects came within a metre of each other. Whether that was a near miss depends on facts that were never in the frame.

Reading the two counts against each other

Report the streams separately and permanently: proximity events per 1,000 forklift hours from the sensor, near misses per month from the register, with the camera-sourced share visible inside the second. Two patterns then become legible. Sensor events rising while human reports hold steady is a real exposure signal worth investigating. Sensor events rising while human reports fall is the failure this whole distinction exists to catch, and no dashboard that merges the two will ever show it.

The rest of this cluster — zones, PPE rules, alert routing and coverage — sits under safety, PPE and site monitoring, part of our manufacturing and industrial vision practice.

• near miss
• leading indicators
• video analytics
• safety reporting