Near-Miss Detection: Catching the Accident That Didn't Happen
Near-miss detection uses AI cameras to automatically spot and log factory close calls — a forklift stopping short, a hand pulled from a press — saving a timestamped clip with no worker paperwork. It turns close calls into a measurable leading indicator that points to the fix before the injury happens.
The important word above is log, not alert. Most vendors sell the live moment: a light, a buzzer, an SMS the instant someone breaches a zone. On an Indian floor that is the wrong product. The retrospective, de-identified log is the thing worth paying for.
Key takeaways
- A near-miss is a free warning. Cameras capture the ones your paperwork never will.
- The value is the pattern over weeks, not the alarm in the second. Live near-miss alerts get muted and quietly turn into a punishment tool.
- A working programme makes logged near-misses rise first, then injuries fall. An early drop is usually a blind camera, not success.
- Detection is only as good as per-camera ground-plane calibration. Accuracy is decided on your floor, not on a spec sheet.
Start with the scene, not the theory
Ask an Indian floor supervisor how many close calls happened last month. You will get a blank look, or a confident "none." That silence is not safety. It is the warning system switched off.
Near-misses happen every shift. Almost nobody reports them. A plant can run for months with "zero incidents" on paper while the same blind corner produces a near-strike every week. Management usually hears about that corner exactly once: when it finally produces a real injury. By then the warning has been paid for in blood.
The stakes are not abstract:
Around 48,000 workers die every year in India from occupational accidents, according to a study by the British Safety Council (British Safety Council India).
Only about 20% of India's workforce is covered by the existing occupational health-and-safety legal framework (same source) — so most of the floor where near-misses happen is barely measured at all.
Why near-misses go unreported in India
The reasons are structural, not personal. Report a forklift near-miss and you risk being blamed for standing in the wrong place. Production pressure means nobody stops a line to write a note. The report is paperwork in a register no one reads. Admitting a close call feels like confessing a mistake to the supervisor who signs your card. So the base of the pyramid goes uncounted, which is the whole game.
The safety pyramid: a heuristic, not a law
Beneath every serious injury sits a wider base of minor injuries and near-misses. H.W. Heinrich first drew it as a triangle in 1931; Frank Bird restated it on a larger sample in 1969. The directional idea is sound: the hazards that eventually cause a serious injury throw off many harmless warnings first.
Treat the ratios as a heuristic, not a fact. The fixed numbers have been challenged for decades. Safety researchers such as Fred Manuele, and reviews by bodies like the UK HSE, argue the neat ratio is not empirically robust and varies hugely by site. What survives is the useful part: a floor generating lots of near-misses is telling you where the serious injury will come from. And you can only act on the ones you know about.
What cameras can actually detect — and how reliably
A camera does not need a worker to notice, remember and report a close call. It watches every shift and flags defined patterns itself, saving a timestamped clip. But no vendor grid tells you which of these events actually survives dust, glare, crowding and occlusion on a real Indian floor. Here is that column.
| Near-miss type | What the camera detects | Reliability on a real floor | The fix it points to |
|---|---|---|---|
| Forklift–pedestrian close call | A person and a moving truck closing below a distance threshold | Reliable — clear, high-value signal when the aisle is framed and calibrated | Re-route the walkway, add a gate at the crossing, mark a keep-clear zone |
| Person in a machine danger zone | A worker crossing into a defined guard/press/robot zone while it is live | Reliable — a fixed zone is easy to define and hold | Fix or add guarding, add an interlock, relocate the control |
| Over-speeding vehicle in an aisle | A forklift moving faster than a set pace in a shared lane | Moderate — depends on good calibration; speed from 2D is easy to get wrong | Enforce a speed cap, add bumps or signage, review operator training |
| Worker crossing behind a reversing vehicle | A pedestrian entering the path of a backing truck | Moderate — occlusion by the vehicle itself hurts | Dedicated pedestrian route, spotter rule, wider turning bay |
| Crowding at a pinch point | Too many people at a door, conveyor or narrow gap | Noisy — high false-positive rate at shift change and breaks | Widen or stagger the flow, change break timing, add a second exit |
| Loitering in a no-go zone | A person standing under a load or near an edge | Noisy — "standing still" is hard to separate from legitimate work | Physical barrier, floor marking, supervised-access rule |
The contrarian part: don't buy the live alert
The category's headline feature — the instant ping — is the part you should be most suspicious of, for two reasons.
First, alarm fatigue. A buzzer that fires every time someone walks near a lane gets muted inside a week, and a muted alarm protects no one.
Second, and worse, the live alert rebuilds the blame culture that killed reporting in the first place. A supervisor pinged "worker X entered zone" now holds a punishment tool, and the floor learns to dodge the camera, not the hazard. That is Goodhart's law on a shop floor: once a near-miss count becomes a disciplinary trigger, it stops measuring safety and starts measuring camera-avoidance.
The defensible design is the opposite of the vendor pitch. Near-miss detection should be deliberately non-real-time, de-identified, and statistical: no individual callout in the moment, just a count by zone that a plant head reads on Monday. Keep threshold ownership with safety/EHS, never production, and never tie an individual event to worker discipline.
The KPI nobody tells you about
The honest success metric is counterintuitive. A working near-miss programme makes the logged count go up first, because visibility finally exists, before injuries fall. An early drop is usually a warning sign: a blinded camera, a shifted angle, or workers who learned to route around the lens. It is rarely a win.
This is also why you should discount the round numbers in vendor case studies. Claims like "35% fewer recordable incidents" circulate without published methodology. We will not invent our own. The trustworthy KPI is your before-and-after count at one zone, after one fix.
From a log to a real fix
A near-miss dataset lets a plant head do things a monthly injury count never allows:
- Rank hotspots by frequency. The crossing throwing the most near-misses gets the first barrier. Evidence-led, not gut-led.
- Justify capital with proof. "This lane produced N pedestrian close calls last month" beats a supervisor's hunch. (Ground N in your own log, not a vendor's.)
- Test whether a fix worked. Re-route the lane, then watch that zone's count drop. Or not. That feedback loop is what manual reporting can never give.
- Spot patterns in time. If near-misses cluster at shift change or the end of a long night, the fix is a roster change, not a wall.
The same footage answers a wider question in how to measure the cost of factory downtime, and the two deepest hazard classes get their own treatment: forklift and pedestrian safety with cameras and machine guarding and danger-zone detection.
Honest limits
A camera does not make a floor safe, and anyone selling it that way gets people hurt.
- It only sees what it frames. An unwatched aisle, or a hazard behind racking, generates no log. Occlusion, dust, glare and poor light all degrade detection.
- Distance and speed come from calibration, not magic. A 2D camera turns pixels into metres and km/h only through a per-camera map to the ground plane (a homography). Mis-calibration — or a camera someone nudged last week — is the dominant source of false and missed alerts. Per-floor tuning, not a datasheet, decides accuracy.
- A "near-miss" is a definition, not a fact. The system flags the thresholds you set. Set them badly and you drown in noise or miss real events.
- A log is not a control. Guarding, barriers, lanes, speed limits and training separate people from trucks; the camera only measures whether they are working.
Where this sits legally
The legal ground shifted recently, and an out-of-date article will mislead you. India's four Labour Codes came into force on 21 November 2025, and the Occupational Safety, Health and Working Conditions (OSH) Code, 2020 is now the primary statute — it consolidates and subsumes the old Factories Act, 1948 among 13 central labour laws (DLA Piper). Many rules under the Code are still being finalised, so the Factories Act lineage still matters as context.
The core duty carries straight through. The occupier/employer must provide safe systems of work "so far as is reasonably practicable" — the wording that lived in Section 7A of the Factories Act, 1948 (general duties of the occupier, inserted by the 1987 amendment) and survives in the OSH Code's employer-duty provisions. Indian OSH law has no near-miss-specific reporting rule; its statutory duties cover accidents and notified dangerous occurrences. But the general duty clearly applies, and a documented near-miss log is direct evidence you are finding and closing hazards rather than waiting for injuries.
Three more anchors a serious buyer will check:
- ISO 45001:2018, the modern OHS management-system standard, requires incident investigation and corrective action (clause 10.2), and near-misses are in scope. A camera log feeds that loop directly — evidence for a standard you may already be audited against.
- The technical body is now the Directorate General Occupational Safety and Health (DGOSH), renamed from DGFASLI effective 15 May 2026 under section 87(1) of the OSH Code. The old dgfasli.gov.in domain still resolves.
- Worker CCTV is personal-data processing under the Digital Personal Data Protection Act, 2023 (DPDP Act): notice and purpose limitation apply when you record identifiable workers. Another reason to run detection de-identified and for safety, not surveillance.
Analytics does not discharge the duty; guarding and training do. But a near-miss record is exactly the proactive management an inspector wants to see, and a timestamped near-miss-to-fix trail is a genuine legal-defensibility asset. (Whether it also becomes a workers'-comp or insurance input is plausible but unproven — treat that as forward-looking until an expert confirms it.)
Manual reporting vs camera-based logging
| Manual reporting | Camera-based logging | |
|---|---|---|
| Who initiates | The worker, voluntarily | The system, automatically |
| Blame risk | High — reporting feels like confessing | Low if kept de-identified |
| Coverage | Whatever gets written down | Every framed event, every shift |
| Over-time trend | Effectively none | A count you can rank and re-measure |
| Feedback loop | Absent | Before/after at each zone |
Where a plant head should start
Don't wire the whole plant on day one. Start where the pyramid is widest:
- Pick your two worst conflict points — usually a forklift crossing and a machine danger zone — and log near-misses there first.
- Reuse existing cameras where sightlines already cover those zones. Add cameras only where the hazard is unframed.
- Run for a few weeks, then read the log. Let the data name the priority fix. Expect the count to rise before it means anything.
- Fix, then re-measure. The whole value is the before-and-after. If a zone's rate doesn't fall after your fix, the fix was wrong.
Deciding which zones to watch, and from where, so a camera can actually see the close call it is meant to catch, is the survey problem Mama removes. Record a short phone walkthrough and it reads the space — aisles, crossings, machine zones, sightlines — then returns a floor plan and a camera-placement plan showing which near-miss zones each camera can genuinely cover, before you buy a single unit. Mama runs on a live pilot floor, so the placement logic is tuned against real aisles, not a catalogue.
FAQ
What counts as a near-miss on a factory floor? A close call that caused no injury or damage: a forklift stopping short of a pedestrian, a worker stepping out of a machine zone just before it cycled, a trolley over-speeding past someone. It is a free warning that the hazard exists, delivered before it hurts anyone.
Should near-miss alerts be real-time? No. A live "worker entered zone" ping gets muted within a week and quietly becomes a tool to punish individuals, rebuilding the blame culture that kills reporting. Run detection de-identified and statistical: a count by zone that safety reads over time, not a callout in the moment.
Does the safety pyramid prove a fixed ratio of near-misses to injuries? No. The Heinrich (1931) and Bird (1969) pyramid is a heuristic — many warnings precede a serious injury. The specific ratios are widely disputed (Manuele, HSE reviews) and vary by site. Treat it as a model, not a law.
Will this catch every close call? No. Cameras only see what they frame, and detection degrades with occlusion, dust, glare and mis-calibration. Near-miss logging sits on top of physical controls — barriers, lanes, guarding, training — never replacing them.