Near misses: how to count them without a survey

A forklift turns into an aisle, a picker steps out from behind a rack, and both stop half a metre apart. Nobody is hurt. By the end of the shift, nobody remembers it either. For a warehouse safety lead this is the most frustrating kind of event: it says more about the next accident than any injury report, and it almost never reaches you.

This article is about closing that gap without another form, another survey or another campaign asking people to report more.

What counts as a near miss in a warehouse?

A near miss is an event that could have injured someone but did not. In a warehouse the typical one is a vehicle and a pedestrian in the same place at the same time: a forklift crossing a walkway, a pedestrian cutting through a loading zone, someone standing in a marked area while a truck reverses. Other near misses matter too (falling loads, blocked exits), but the vehicle–pedestrian encounter is the one that repeats every day and the one with the worst outcome when luck runs out.

What makes a near miss useful is not the single event but the pattern. One close call in aisle 3 is an anecdote. The same close call several times a week in the same aisle, at the same shift change, is a design problem you can fix.

Why do near misses go unreported?

Near misses go unreported because reporting them costs the person more than staying silent. The reasons are human and understandable:

• Nobody was hurt. The event feels finished the moment it ends. There is nothing to show, nothing to treat, nothing to explain.
• The form takes time. On a busy shift, ten minutes at a terminal is ten minutes the line is waiting for you.
• A report feels like reporting a colleague. If the driver was a friend, writing it down feels like blame, even when the system is designed not to blame anyone.

None of these reasons go away with a poster or a reminder. A reporting culture can be built, and it is worth building, but it takes years and it always depends on someone choosing to speak up. In the meantime you make safety decisions on what does get reported: injuries. That means deciding after the fact, not before.

Can the cameras already on the wall count near misses?

Yes, if each camera is given a clear rule about what to watch for. Most warehouses already have IP cameras covering docks, main aisles and crossings. They are recorded, and the recordings are watched only after something has gone wrong. The same streams can be analysed as they happen.

The approach is simple to describe:

1. On each camera you draw the areas that matter: a walkway, a crossing, a loading zone, a line nobody should pass while a truck is moving.
2. A small computer at your site looks at the camera images and finds people and vehicles in them.
3. When a rule is broken (a person inside a vehicle zone, a vehicle crossing a pedestrian line, someone lingering where they should not), it creates an event with the time, the place and a short clip.

Two things to know before you rely on it. First, a standard detection model knows people and vehicles; it does not have a dedicated "forklift" class. Whether it reliably picks up your trucks, at your camera angles and lighting, is something to verify on your own footage before you rely on it, not something to assume. Second, a camera only sees what it sees: blind corners stay blind until you add or move a camera.

Who decides what is real?

A person does. Every event goes to someone on your team who looks at the clip and marks it as confirmed or as a false alarm. That decision is part of the record. Software that "decides" what happened on its own would be faster, but it would also be a number nobody can defend in front of management, the works council or an inspector.

The confirmation step does a second job: it is how the rules get better. If one zone produces mostly false alarms, the zone is drawn wrong or the camera angle is wrong, and you see it within days rather than months.

Will employees and the works council accept it?

They are far more likely to accept it when the system cannot be turned against individuals, and when that is a property of the system, not a promise. In practice that means:

• no facial recognition, no identification of who it was, no score for any worker. The data says that a person was in the zone, not which person;
• the video is analysed at your site; only events and their short evidence clips are sent on, and in an on-site installation nothing leaves your premises at all, except during a remote-support session you open yourself;
• the output is a count per zone and per time of day, which points at layouts and processes, not at people.

In many European countries the works council or employee representatives must be consulted before a system like this goes live. Plan that conversation from the start and bring the facts above to it. It is easier to agree on what a system cannot do than to argue later about what it might do.

What changes in the conversation with management?

The conversation moves from luck to numbers. Today, a safety lead arguing for a new barrier or a separate walkway usually brings anecdotes and the memory of the last accident. With counted near misses the argument looks different. An illustrative example: "In aisle 3, pedestrians and forklifts met nine times a week in September, almost all of them at shift change. We moved the picking start by ten minutes and painted a crossing. In October it was twice a week."

That is a decision management can take, a cost it can weigh and a result it can check a month later. It also changes what you report upwards: not only injuries, which arrive late and rarely, but leading indicators that move week by week.

How do you start without a big project?

Start small, on cameras you already have, and measure before you commit. A sensible first step looks like this:

1. Check your cameras. List the cameras that cover the aisles and crossings you worry about, and check that they can provide a video stream the analysis can read.
2. Pick one site and a handful of cameras. Three or four well-placed cameras in the busiest area tell you more than thirty spread thinly.
3. Run it for a few weeks with a person confirming every event. At the end you should have a count per zone, a list of false alarms and what caused them, and a clear answer to the question that matters: does this show you something you did not know?

With VAP this first step is a fixed-scope trial: three weeks on up to six of your own cameras, followed by three months of live operation, for €2,900 excluding VAT. At the end you get a written report with the numbers from your site, and you decide whether to continue.

You can check whether your cameras are suitable in a couple of minutes, and you can see what the report at the end of the trial looks like before you talk to anyone:

• Check your cameras
• Sample PoC report (PDF)