There is a number every safety professional quotes and almost none can verify for their own site: the ratio of near misses to serious injuries, the broad base of the incident pyramid that supposedly sits beneath every recordable. The exact figures attached to that pyramid have been argued over for ninety years, and the argument misses the operational point, which is simpler and harder. Whatever the true ratio is at your plant, your reporting system is capturing a fraction of it, you do not know what fraction, and the fraction changes with every reorganization, every staffing squeeze, and every rumor about what happened to the last person who filed one. A safety program steering by near-miss reports is steering by an instrument whose error is unknown and unstable, and this article is about what that costs and what continuous detection does to it. It is part of our full guide to AI for workplace safety in manufacturing.
Why near misses go unreported
The reporting gap is usually discussed as a culture problem, which is half true and manages to insult everyone involved. The fuller truth is that not reporting is frequently a rational act by a reasonable person, and until a program grapples with that, exhortation changes nothing.
Consider the picker who steps back as a forklift swings past closer than it should have. Reporting means stopping, finding the form or the kiosk, reconstructing the event in writing, and possibly triggering a review in which the picker, the driver, or both explain themselves to people with authority over their employment. The event took two seconds and ended fine. The report takes twenty minutes and might not. Every incentive in that arithmetic points one way, and it points the same way harder when the near miss involved a shortcut the worker chose, when the shift is behind, when the driver is a friend, when the form asks for names, and when last quarter's reports produced no visible fix. None of this requires a bad culture, only an ordinary one.
Programs then compound the problem with the metric itself. A site that sets near-miss reporting targets gets reports, of the safest possible kind: paper cuts and stubbed toes, filed to make the number, while the genuinely instructive events, the ones involving judgment, pressure, and machinery, stay unreported because those are precisely the ones with social cost attached. Meanwhile a falling report count gets read in two opposite ways depending on the meeting: as improving safety or as collapsing trust. Both readings are available because the instrument cannot distinguish them, which is the whole indictment.
The consequence lands where a safety leader can least afford it. Resources follow reported risk, so the loading dock that generates paperwork gets the barrier while the quiet intersection that generates nothing gets discovered by the incident investigation. The Occupational Safety and Health Administration's (OSHA) own turn toward leading indicators, measures that reveal problems before injury rather than after, presumes the leading data is trustworthy, and self-reported near misses are the least trustworthy data most programs hold.
Weight near misses by severity potential
The intellectual scaffolding under most near-miss programs is the incident pyramid, descended from Herbert Heinrich's work in the 1930s, which proposed a fixed ratio between minor incidents, serious injuries, and fatalities, and drew the conclusion that driving down the base of the pyramid drives down its peak. The model earned its longevity by being half right in a useful way: frequency of exposure does relate to probability of harm, and attention to small events genuinely surfaces fixable conditions. Modern safety science has spent two decades documenting where it is wrong, and the correction matters for how detection data should be read.
The core critique is that serious injuries and fatalities are not, in general, scaled-up versions of minor ones. They have their own precursors, concentrated in specific high-energy situations, suspended loads, mobile equipment, stored energy, work at height, and a site can drive its total near-miss count down for years while the handful of events with fatality potential continue at their old rate, invisible inside an improving aggregate. A reporting program that treats all near misses as equal weight is therefore optimizing the wrong curve, and a detection program that simply automates the same counting would inherit the same blindness.
This is where detection data offers something the pyramid era never had: the events come with attributes that allow severity-potential weighting. A dwell violation in a marked walkway and a person inside a crane's swing radius during a lift are both "near misses" to a tally; they are different universes of potential, and because each detected event carries its zone, detection type, and context, the record can be read through a severity-potential lens, tracking the high-energy subset as its own population. The safety review that matters then has two lines rather than one: total exposure events, which measure general conditions, and high-potential events, which measure proximity to the outcomes that end careers and lives. A site whose total is falling while its high-potential line is flat knows something no pyramid could have told it, and knows exactly where to spend.
What cameras count that forms miss
Continuous detection changes one thing, and it turns out to be the thing that matters: it removes the decision to report. A camera running zone and proximity rules does not weigh social cost, schedule pressure, or friendship before logging a vehicle-pedestrian near pass. It logs all of them, identically, on every shift, including the night shift no observation program has ever covered properly.
Sites that run detection alongside an existing reporting program get to perform an experiment almost none have ever run: comparing the two records for the same events in the same zones. The pattern is consistent and clarifying. Detected near passes outnumber reported ones not by percentages but by multiples, the gap is widest exactly where reporting friction is highest, night shifts and high-pressure areas, and the events that do get reported skew heavily toward the ones with witnesses. None of this shames the reporting program. It calibrates it, for the first time, and the ratio of detected to reported events becomes a measurable index of reporting culture itself, trackable by zone and shift the way any other indicator is tracked. That ratio, and the exposure rates detection makes possible, are the subject of our companion article on turning detections into safety Key Performance Indicators (KPIs).
What detection cannot do is equally important to state, because the two records answer different questions and a program that discards one for the other has misunderstood both. A camera counts the event: the proximity, the zone entry, the person down. It does not know why, whether the driver was undertrained or the route badly designed or the picker distracted by a defective pallet jack, and it has no access to the events outside camera coverage or the ones with no visual signature at all, the almost-dropped chemical, the wrong-valve-nearly-opened. The narrative half of the near-miss record, the why, still arrives only through people, which means the reporting program does not retire. It gets a different job.
Redesigning the reporting program
Once the counting is automatic, the human program can stop pretending to be a census and become what it should have been all along, an investigation resource aimed by data.
The redesign that works in practice has a consistent shape across the sites that have run it. Detection carries the burden of frequency: the counts, the trends, the rankings of which intersections and zones carry exposure. The human program is then pointed at the top of that ranking, and instead of asking the whole workforce to report everything, supervisors ask specific crews about specific patterns: the record shows forty near passes at dock three on nights this month, walk us through what is happening there. That question is easier to answer than a blank form, carries no accusation because the count is about a place rather than a person, and produces exactly the narrative detail the cameras lack. Reporting friction falls because the ask is smaller; report quality rises because the subject is real.
Two design rules keep the whole structure trustworthy. The detection record must stay no-discipline, in writing, for all the reasons covered in our article on works councils and worker privacy, because the moment a near-pass clip appears in a write-up, the workforce will treat the cameras as the new form, to be avoided rather than filed. And the fixes must be visible: every top-of-ranking pattern that gets a barrier, a reroute, or a schedule change should be announced against the data that motivated it, because nothing rebuilds reporting culture like evidence that the system produces concrete change. Sites report the human program strengthening after detection arrives, which surprises people who expected replacement, and should not: the program finally has proof that somebody reads the reports.
Using the data in toolbox talks
Detection data changes the monthly safety meeting, but its cheapest wins happen at the weekly and daily level, inside routines the site already runs. The toolbox talk is the clearest example. Instead of the generic slip-hazard reminder recycled from the corporate deck, the supervisor opens with this week's pattern from the crew's own area: dock three logged eleven near passes on nights, here is what the corridor looks like at shift change, anonymized and framed as a condition of the place rather than conduct of a person. Talks built on the crew's own footage get a different quality of attention, and they routinely surface the missing context on the spot, the staged pallets that narrow the aisle on Thursdays, the second forklift borrowed from receiving, the detail that turns a count into a cause.
The same record shortens the argument that follows every proposed fix. When the reroute or the barrier is trialed, the event rate for that zone before and after is a two-line chart, and the trial either worked or it did not, in numbers the crew watched being generated. Sites consistently underestimate how much goodwill this produces: workers have spent careers reporting hazards into a void, and a system that visibly measures, fixes, and confirms is the opposite of a void.
How VIDIZMO fits
VIDIZMO AI Live Insight produces the near-miss record this article describes, from zone, line-crossing, dwell, and proximity rules running on the cameras a plant already owns, processed on the customer's own hardware on site, which is where a continuous surveillance workload practically belongs. Every event carries its timestamp, camera, detection type, and severity, the surrounding clip is preserved automatically with footage from before the trigger, and the events aggregate into the exposure reports the safety committee reviews, exportable to whatever the site already uses. The clips live in the Nexus portal the deployment works alongside, under its access control and retention rules, so a disputed count can always be traced to the moments behind it, and where the ambiguous event needs a second look before it enters the record, a short clip can be put to a multimodal model with a written question, the escalation pattern covered in our article on unsafe behavior beyond the preset list.
The honest first step costs a site nothing but nerve: run detection silently for a month in the two or three zones the safety team believes it understands, then set the detected count beside the reported count for the same period. That one comparison, done once, reframes every near-miss slide the site has ever presented, and it is the beginning of a program that measures exposure instead of paperwork.
One closing caution about what to do with the number the comparison produces, because the first instinct is usually wrong. A detected-to-reported ratio of twenty to one is not an indictment to be announced at a town hall, and handling it that way teaches the workforce that the cameras exist to prove them liars. The sites that convert the finding into progress present it the other way around: the reporting system was asking people to do an unreasonable amount of clerical work under production pressure, the measurement now carries that burden, and what the program asks of people going forward is smaller and more valuable, context about the patterns the record surfaces. Same number, opposite meaning, and the difference decides whether the months that follow produce a stronger safety culture or a quieter one.
