Near-Miss Detection Explained: What AI Video Counts as a Close Call

A near miss is an event that could have hurt someone and did not. AI video safety systems detect a narrow slice of near misses, mostly people coming too close to moving vehicles, loads or machines, by applying rules for distance, speed, zones and timing to what the cameras see. Those rules are a proxy for risk, so the number on a dashboard depends as much on how the site defines a near miss as on what actually happened. This article explains the standard definitions, how camera systems turn them into rules, where the counts mislead, and how to write a definition you can defend.

What counts as a near miss?

Safety regulators define a near miss by what could have happened. HSE's guidance on investigating accidents and incidents (HSG245) defines a near miss as "an event that, while not causing harm, has the potential to cause injury or ill health," and for the purposes of that guidance includes dangerous occurrences in the term ^[2]. HSG245 separates this from an "undesired circumstance," a set of conditions with the potential to cause harm, such as untrained staff handling heavy loads ^[2]. The same guidance points out that it is often luck that decides whether an undesired circumstance becomes a near miss or an accident ^[2].

OSHA's template near-miss reporting policy takes a broader view. It describes a near miss as an opportunity to improve safety based on a condition or incident with potential for more serious consequences, and lists unsafe conditions, unsafe behavior and minor incidents that could have been worse ^[1]. Many US employers use a policy like this, but OSHA does not generally require near misses to be reported to the agency.

In Great Britain, some events in this family are legally reportable. RIDDOR lists dangerous occurrences, which HSE describes as certain incidents with a high potential to cause death or serious injury, such as the collapse or overturning of lifting equipment or contact with overhead power lines ^[3]. These must be reported whether or not anyone was hurt, and whether or not a camera recorded them.

All three definitions share one feature that matters for technology: they rest on judgment about potential. A person deciding "that could have killed someone" is weighing speed, mass, distance, the people involved and what might have happened next. A camera system has to approximate that judgment with measurements.

How does AI video decide something was a near miss?

Camera-based systems detect objects (people, forklifts, trucks, loads), track them across frames, and apply rules. A near miss is whatever the rules say it is. Intenseye, for example, lists detections such as "person in an active forklift path," "converging paths with no sightline" and a worker beneath a lifted load ^[6]. Protex AI says it tracks forklifts, pallet jacks, automated vehicles and pedestrians together and flags proximity events and near misses ^[7].

The common rule types are shown below.

Each rule has parameters: a distance, a speed floor, a time window, a minimum duration, an active-hours schedule. A site that sets a 3 meter proximity threshold will log far more events than one that sets 1 meter, from the same footage, so any count is only comparable with counts made under the same settings.

What road safety research says about using conflicts as a proxy

Traffic engineers faced the same problem decades ago. Crashes are rare, so waiting for them is a slow way to judge whether an intersection is dangerous. The traffic conflict technique, in which observers count strong braking and evasive maneuvers, formally began with studies at General Motors Research Laboratories in the late 1960s ^[5].

The US Federal Highway Administration's 2008 report on its Surrogate Safety Assessment Model (SSAM) shows how the idea became automated. SSAM classifies vehicle interactions as conflicts using two measures: time to collision (TTC), the time remaining before two vehicles would collide if they kept their current speed and path, and post-encroachment time (PET), the gap between one vehicle leaving a point and another arriving at it ^[5]. SSAM uses a default TTC threshold of 1.5 seconds, based on earlier research, and lets analysts change it ^[5].

The same report is candid about the limits. It says the conflict method has shown some correlation with crashes, but that there is still debate about the connection between conflict measures and crash predictions ^[5]. Workplace systems that borrow TTC-style logic for forklifts and pedestrians inherit that debate, with less validation data behind them than road safety has.

Is a camera near miss the same as a reported near miss?

No, and treating them as one series will distort trends. The two capture different events and fail in different ways.

Keep the two as separate series. A site that adds camera events to its manual near-miss total will see an apparent jump in near misses the month the system goes live, which says nothing about risk.

Why near-miss counts can mislead

Thresholds and definitions drift

Vendors and sites re-tune rules during pilots to cut nuisance alerts. Every change resets the baseline. If the proximity distance moves from 3 meters to 2 meters in month two, a drop in events from month two onward is a change in measurement. Log every rule change with a date so trends can be read correctly.

Exposure changes

More forklift movements, more shifts or a seasonal peak will produce more conflicts at the same level of risk per movement. Report near misses per hour of operation, per thousand vehicle movements or per camera-hour, alongside the raw count.

Falling counts are not proof of safety

Vendor case studies often report large drops in detected near misses. Protex AI's website cites an "80% reduction in near-miss events within 54 days of deployment" at a North American automotive manufacturer, and a 97% decrease at a food distribution company ^[7]. These are vendor-reported results and have not been independently verified. A fall in detected events can come from safer behavior, which is the goal. It can also come from re-tuned thresholds, a camera knocked out of alignment, workers avoiding the monitored area, or reduced traffic. Before accepting a reduction, ask what changed in the rules and the coverage over the same period.

The pyramid does not hold as well as the poster suggests

The appeal of near-miss counts rests on the safety pyramid. Frank Bird's 1969 work, based on more than 1.7 million accident reports from 297 companies, popularized a ratio usually quoted as 1-10-30-600 ^[4]. Later studies have questioned whether such ratios are stable or predictive. One study of US mines found no significant difference in the probability of a later fatal event between mines with few incidents and those with none, and research in Chile found the ratios vary with the base incident rate ^[4]. A camera system that drives down minor proximity events may not touch the precursors of a fatality, which often involve different tasks, such as maintenance, lifting or work on unguarded machinery.

What camera near-miss detection cannot see

Cameras can only flag what is in view and visible. That rules out most electrical near misses, chemical releases below a visible threshold, pressure system faults and many ergonomic close calls. It also rules out events behind racking, inside vehicles or in areas without coverage. RIDDOR dangerous occurrences, such as a lifting equipment collapse, still need reporting through your normal process even if a camera captured the clip ^[3].

Proximity rules also struggle to judge intent and control. A banksman standing close to a reversing truck is doing their job; a visitor in the same spot is a near miss. Some systems handle this with zones and schedules, others by excluding people wearing particular clothing. Each workaround has its own error rate.

How to write a near-miss definition your site can use

Start from the hazards most likely to cause serious harm on your site, then write rules that a reviewer can apply to a clip and get the same answer as a colleague.

1. List the interactions that matter: forklift and pedestrian, truck and pedestrian at the dock, person under a suspended load, person inside a machine envelope. OSHA's pedestrian traffic guidance for powered industrial trucks starts from separating forklift traffic from pedestrians where possible, which tells you where to look first ^[9].
2. For each interaction, set measurable conditions: distance, vehicle moving or stationary, speed floor, minimum duration, zone and active hours.
3. Assign severity tiers. A pedestrian 0.5 meters from a forklift moving at speed is a different event from one at 2 meters beside a slow vehicle. Tiering lets you prioritize review and keeps the headline number from being dominated by low-risk events.
4. Define what the reviewer confirms: the objects were correctly detected, the conditions were met, and the event meets the site definition. Record false positives by cause.
5. Fix the definition for a set period, at least a quarter, and change it only through a logged decision.
6. Normalize the count by exposure and publish the definition next to the chart.

A short definition table might look like this.

The numbers above are illustrations of format. Your thresholds should come from your risk assessment, your vehicle speeds and the measurement accuracy your vendor demonstrates on your own cameras.

Watch the alert load

A system that produces hundreds of low-value alerts teaches people to ignore all of them. A 2006 NIOSH evaluation of a radar proximity warning system on haul trucks found it detected people and vehicles reliably, but frequent alarms from objects that posed no immediate danger were a problem, and the author recommended pairing sensors with cameras so operators could check the cause ^[10]. Send only high-tier events as live alerts, and route the rest to trend reports.

Using near-miss data without punishing workers

Camera near misses often show a person in the wrong place. The tempting response is discipline. HSG245 makes the case for investigating near misses because they reveal failures in risk controls without anyone being hurt, and because people talk more openly when no one is injured ^[2]. If camera events lead mainly to individual warnings, workers will learn to avoid cameras, and the useful signal about layout, scheduling and traffic design is lost. Use clusters of events to ask why people cross where they do, whether walkways match real routes, and whether vehicle and pedestrian flows can be separated by time or space.

Summary

A near miss is defined by its potential for harm, and that judgment has to be turned into rules before a camera can count one. Most AI-detected near misses are proximity, path-conflict or zone events involving people and vehicles or loads, and the count depends heavily on thresholds, camera coverage and traffic volume. Road safety has used conflict counts as a proxy for crashes since the 1960s and still debates how well they predict harm. Treat camera near misses as a separate series from manual reports, write and freeze a site definition with severity tiers, normalize by exposure, audit a sample, and read vendor claims of large reductions as vendor claims until you have checked what changed in the rules.