Near miss detection with AI video: catching what goes unreported

What a near miss is, and what near miss detection means

A near miss is an unplanned event that did not cause injury, illness or damage, but could have. The difference between a near miss and an accident is often a second, a step or luck. That is why safety management treats near misses as warnings: the same conditions that produced one near miss can, if nothing changes, produce an accident.

Near miss detection is the practice of finding these events so they can be investigated. Traditionally it depends on people reporting what they saw. With AI video, part of that work can be done on the cameras a site already has: the near misses that are visible on camera can be recognized as they happen and recorded, whether or not anyone reports them.

Near misses a camera can see

• Vehicle–pedestrian near misses. A forklift and a person on foot whose paths converge at a blind corner, a dock door or a vehicle lane; a forklift reversing toward someone behind it. See forklift and pedestrian safety.
• Dangerous maneuvers and speed. A forklift taking a corner without slowing, or a route violation through a pedestrian area.
• Reaching into an energized machine. An operator about to intervene on a machine whose guard is open and whose energy has not been isolated. See lockout/tagout and machine guarding.
• Unauthorized entry into a danger zone. A person inside a robot cell or another restricted area who is not part of permitted maintenance.
• Work at height without a harness. A person on a platform, roof edge or high racking without a harness attached. See work at height and restricted areas.
• Missing PPE where the task needs it. Equipment missing for the zone and the task, read in context. See PPE compliance.

Near misses a camera cannot see

A camera can only read what it shows. A near miss outside camera view, or one that is not visible at all — a brief exposure to a chemical vapour, an electrical shock that didn't land, a tool that slipped inside a machine — still depends on the people involved reporting it. Video adds to a reporting culture; it does not replace it.

Why traditional methods miss near misses

Many sites have a near miss reporting form, a safety team and regular inspections. Near misses still slip through, because they are short, they end without harm, and they happen across a large site around the clock.

• Under-reporting. Reporting takes time, nobody was hurt, and people may worry about blame — for themselves or a colleague. A near miss that ended well is easy to shrug off. Good safety management treats these events as warnings to investigate, but that investigation can only start from events someone records.
• Periodic audits see a moment, not the site. A safety walk observes one area for a short time, and people behave differently when they know they are being watched. The near miss at the dock on the night shift is not on the audit route.
• Manual CCTV review doesn't scale. The cameras are often already recording. But finding a near miss in the footage means knowing which camera and which minute to watch, so footage is usually reviewed after an accident, not to find the near misses that came before it.
• Rule-based detectors fire without context. A fixed zone or a distance rule triggers when a person enters an area or comes close to a vehicle. It can't tell a person working beside a parked forklift from one stepping into the path of a moving one, or a technician on permitted maintenance from an unauthorized entry. The result is many alarms for normal work, silence for the situation nobody drew a zone for, and a team that learns to ignore the alarms.

What these methods lack is the reading a safety professional makes on the spot: what is happening, what was about to happen, and why.

How vision-language models make it solvable

Primarch builds on and enhances vision-language models (VLMs): models that read video and language together. Instead of drawing boxes around objects frame by frame, the model reads the scene the way an experienced safety professional would — and does it on every camera, on every shift.

• It reads movement over time. A VLM reads a window of time, not a single frame, so the before and after of a movement are read together. That is what makes a near miss visible: two paths converging, a hand moving toward an open guard, a step onto a platform edge.
• It reads the situation, not just the objects. A person and a forklift in the same picture is normal work. A person stepping into a reversing forklift's path is a near miss. Context is what separates the two, and reading the scene in context is designed to remove the biggest source of false alarms in rule-based systems: rules firing without context.
• It writes the finding in plain language, with a clip. Each near miss becomes a sentence a supervisor can read — what happened, where, and why it matters — with a time-stamped evidence clip, so everyone discussing it sees the same thing.
• It chains events by shift, zone and time. Single events are recorded and alerted for review; chained together, they show where and when the same situation repeats, and help diagnose why: an attention lapse, or a process gap.

A near miss report starts with someone deciding to write it down. A near miss seen on camera can become a record whether or not anyone does — with a time, a place, a plain description and a clip.

What to look for in a solution

Whichever vendor you talk to, these are fair questions to ask about a system meant to detect near misses on video:

• Works on your existing cameras. Can it connect to the IP cameras you already have, or does it need new hardware?
• Reads situations, not just proximity. Does it consider what people and vehicles are doing and where they are heading, or does it only measure distance and zone entry?
• Is clear about what it covers. Which kinds of near misses does it recognize, and which still depend on people reporting them?
• Turns near misses into records. Each event should come with a time, a place, a plain description and an evidence clip, so it can be investigated like a reported one.
• Handles false alarms openly. Ask how findings are reviewed by a person, and how single events are separated from repeating patterns.
• Explains repeats. Can events be grouped by shift, zone and time, and does it help separate an attention lapse from a process or layout gap?
• Respects privacy. Is footage anonymized? Is there an on-premise option where footage never leaves the site? Does it meet GDPR and other data-protection requirements, and does it report situations rather than individuals?
• Can be piloted on your own site. One zone, a few of your cameras, your own work — not a demo video.

Example use cases

The walk-throughs below are illustrative examples of how a finding is built. They are not customer cases and contain no performance figures.

Example 1: a forklift near miss at a dock door

1. 01

   Moment

   Dock door 2 at a distribution center, 07:50, during morning receiving. A camera covers the dock floor.

2. 02

   What the camera shows

   A forklift reverses out of a trailer with a pallet. A checker with a clipboard walks behind it along the dock edge. The forklift stops sharply; the checker steps aside.

3. 03

   The finding + evidence clip

   "Near miss at dock door 2: a forklift reversed out of a trailer toward a pedestrian on the dock floor behind it; it stopped short and the pedestrian stepped aside." The clip covers the seconds before and after.

4. 04

   Who is alerted

   The dock supervisor, instantly; the event goes into the safety team's weekly digest.

5. 05

   What the shift and zone pattern shows

   Similar near misses at doors 1–3 fall in the first hour of receiving, when checkers work on the dock floor between forklift runs.

Example 2: a hand toward an open guard

1. 01

   Moment

   A packaging line, 18:40, on the evening shift. A camera covers the machine and its guard.

2. 02

   What the camera shows

   The guard of a filling machine is open and the machine has not been isolated. An operator reaches toward the jammed area, then pulls back and stops the machine at the panel.

3. 03

   The finding + evidence clip

   "Near miss at the filling machine: an operator reached toward the jammed area with the guard open while the machine was still energized, then stopped it at the panel." Clip attached.

4. 04

   Who is alerted

   The shift supervisor, instantly; the maintenance lead in the weekly digest.

5. 05

   What the shift and zone pattern shows

   Similar events at this machine cluster around jams late in the evening shift, which points at the jam-clearing procedure rather than one operator.

Example 3: a step onto high racking without a harness

1. 01

   Moment

   A warehouse, 13:15. A camera covers a section of high racking.

2. 02

   What the camera shows

   A person climbs onto the racking from a pallet on raised forks to free a stuck carton, with no harness attached, and climbs back down a minute later.

3. 03

   The finding + evidence clip

   "Work at height without a harness: a person climbed onto the racking from raised forks to free a carton and returned without fall protection." Clip attached.

4. 04

   Who is alerted

   The warehouse shift supervisor, instantly.

5. 05

   What the shift and zone pattern shows

   Asked in plain language — "show work at height flagged in this aisle last month" — the event memory returns a short list, all linked to stuck cartons on the top level.

Example 4: an entry into a robot cell during production

1. 01

   Moment

   A welding cell, 10:05. A camera covers the cell and its gate.

2. 02

   What the camera shows

   A person enters the cell through the gate while the robot is moving, with no maintenance work under way, and leaves after picking up a part from the floor.

3. 03

   The finding + evidence clip

   "Unauthorized entry into the welding cell while the robot was operating; no permitted maintenance was in progress." Clip attached.

4. 04

   Who is alerted

   The line supervisor, instantly; the event appears in the zone's weekly safety score.

5. 05

   Corrective action

   A recommendation written for the production manager: give dropped parts a retrieval routine that does not require entering a running cell, and review the gate interlock.

How Primarch's OHS Expert helps

This section is about our product. The OHS Expert is the Primarch expert module that reads the floor like a safety professional, on the cameras a site already has. For near miss detection it:

• Reads the forklift's route and the pedestrian's direction together, and recognizes the near miss and the dangerous maneuver before the intersection happens.
• Understands an operator is about to intervene on a machine that is still energized, and flags that moment.
• Tells unauthorized entry apart from permitted maintenance, and catches harness-free work at height at the level of the action.
• Judges missing PPE against the zone and the task: a visitor and a welder are not the same thing.
• Narrates each finding in natural language, with a time-stamped evidence clip, and recalls past repeats of the same scene from memory.
• Chains events by shift, zone and time and diagnoses why they repeat — an attention lapse or a process gap.
• Writes corrective actions in the language an inspector would use, assigns owners, and prepares regulation-referenced records, weekly zone safety scores and executive summaries.

It needs no rule set: it comes pre-configured with OHS domain knowledge and regulation, site-specific thresholds are tuned together at onboarding, and situations nobody wrote a rule for are still caught through scene understanding. Multi-stage validation and temporal noise suppression are added on top.

Putting it to work

1. 01

   Connect

   Connect your existing RTSP cameras; no new cameras or special hardware.

2. 02

   Pilot on your own site

   Pilots typically start with a single zone and a few cameras — a busy dock or a line with frequent jams, say. Thresholds are tuned to your site together, and you review the findings with your safety team.

3. 03

   Roll out

   Scale facility-wide as the system proves out, with dozens of streams watched in parallel.

• Alerts, digests and dashboards. Instant notifications, periodic digests and role-based dashboards, so each finding reaches the person who acts on it.
• Ask the past in plain language. Findings are written into a queryable event memory: "Which near misses were recorded at the docks last week?"
• Records you can defend in an audit. Every finding carries its time-stamped clip; reports carry the regulatory reference and the corrective action with its owner.
• Privacy by design. Footage is anonymized in a GDPR-compliant way, and the system is designed to report risky situations and process gaps, not individuals. With on-premise deployment, footage never leaves the facility.