A few weeks back I got to sit in on a site safety review where the footage wasn’t shot by a person on the ground, it was shot by a drone hovering somewhere above a half-built tower. The idea sounded simple enough on paper: fly a camera over the site, let an AI model look at every worker in frame, and flag anyone who isn’t wearing a helmet, a vest, or proper shoes. In practice, watching it actually work was a different thing altogether, and it changed how I think about site safety audits.

Why Someone Even Bothered Building This

Anyone who has spent time around a construction site knows the drill. A supervisor walks the floor a couple of times a day, and if he happens to be looking the other way when someone takes their helmet off to scratch their head, that moment just never gets recorded. Multiply that across a site with a hundred workers spread over ten floors, and it’s obvious why so many near-misses go unnoticed until something actually happens. A drone doesn’t get tired, doesn’t have a blind spot on the fifth floor, and doesn’t need a coffee break. That’s really the whole pitch.

The First Site I Looked At

This is one of the frames from the review. Seven workers picked up in a single pass, and the system is already scoring each one for helmet, vest, and shoe compliance in real time.

Persons: 7 | Helmet: 4/7 | Vest: 3/7 | Shoes: 1/7 — rebar zone near the crane base

What struck me here wasn’t the technology; it was the number. Only one worker out of seven had shoes that the model was confident enough to flag as proper protective footwear. On a rebar-heavy floor like this one, that’s exactly the kind of gap a manual walkthrough would probably miss, because from ground level nobody is checking footwear from thirty feet up.

A Bigger Crew, A Bigger Blind Spot

The second clip was from a different angle, a wider stretch of the same site with ten people working across scaffolding and open rebar columns.

Persons: 10 | Helmet: 5/10 | Vest: 0/10 | Shoes: 0/10 — scaffolding and rebar section

Zero out of ten wearing a vest that the system could confirm. Now, in fairness, not every one of those readings means a worker is genuinely non-compliant, sometimes it’s just the angle, sometimes a vest is tucked under a jacket, sometimes the drone’s camera catches someone half in shadow. But even accounting for that margin of error, a number like this on a site with ten people is worth a second look, and that second look is exactly what the alert is for. It doesn’t replace a human decision; it just makes sure the human gets called in at the right moment instead of after the fact.

Where It Actually Gets Useful

The third example is smaller, only five people in frame, but it’s a good illustration of how confidence scoring works day to day.

Persons: 5 | Helmet: 5/5 | Vest: 3/5 | Shoes: 0/5 — with one worker tagged No_Vest

Here every single worker had a helmet on, which is a good sign, but one of them is directly flagged as No_Vest rather than just an unscored box. That distinction matters. A missing score usually means the model wasn’t sure. A direct tag like No_Vest means it’s reasonably confident that piece of gear simply isn’t there. That’s the difference between a soft warning and a real alert, and it’s the kind of nuance that makes this useful instead of just noisy.

It’s Not Just About Helmets and Vests

Once the drone is already up there filming, the same footage can be used for a lot more than PPE checks. If a worker is standing too close to an open edge, or a scaffold section looks like it’s leaning wrong, or there’s a pile of loose material sitting where a crane needs to swing, that same feed can be reviewed for it. Some teams are even using it to build a rough timeline of what happened right before an incident, since the drone footage becomes a record you can rewind, which a supervisor’s memory of the day usually can’t compete with.

The Part Nobody Tells You Upfront

None of this works if people treat the numbers as gospel. A confidence score of 0.34 on a helmet detection is not the same thing as certainty; it’s a starting point for someone to actually look at the frame and decide. The sites that get real value out of this aren’t the ones that automated the whole safety process away, they’re the ones that used the drone to catch what a person would have missed, and then still had a person make the call. I think that’s the honest way to describe what I saw. It’s a second pair of eyes, a very consistent and very tireless one, but still just a pair of eyes feeding information to the people actually responsible for the site.

Would I Recommend It

If you’re running a site with more than a handful of workers spread across multiple levels, yes, I’d say it’s worth trying. Not because it catches everything perfectly, the numbers above show it doesn’t, but because it catches things that would otherwise just slip by unnoticed until an inspection or, worse, an incident. Start small, run it alongside your existing safety walks for a few weeks, and see how often the alerts line up with things your team would have wanted to know about anyway. That’s usually the fastest way to tell if it’s pulling its weight on your site.