Cameras that speak up: AI video analytics on jobsites

There is a scene I have seen on almost every jobsite I have visited. Monday morning, cable is missing, tools are missing, and someone says the usual line: “don’t worry, we have cameras”. Then the sad part begins: sitting down to go through sixty hours of footage to find the ten minutes that matter.

The cameras were there. They recorded everything. The problem is that nobody was watching when it happened.

That was the problem on the table when we started working with BGS Vision, the BGS Group company focused on jobsite monitoring. This post is what we learned along the way.

The problem is not having more cameras#

The first instinct is to add more cameras, with higher resolution and night vision. It helps, but it solves nothing if the camera only records. An active jobsite produces hours and hours of video nobody will ever watch in full, and a site that is closed at night is exactly when you most need someone paying attention.

So the question changed from “how many cameras do we need?” to a much more useful one: how do we get the camera to speak up on its own when something that matters happens?

Build or integrate#

At PROOQ we build a lot of our own software: Suite HUB, the new version of E-BOP, portals, integrations. So the natural temptation was to build our own analytics.

We thought about it seriously and decided against it. AI video analytics at scale means vision models you have to train and maintain, cloud recording, mobile apps, encryption, support for dozens of camera brands. That is not something you do well in a few months, and there are mature platforms that have already solved it.

So BGS Vision became an integrator of an AI cloud video platform, and our job became something else: fitting it to the reality of a jobsite and explaining it in the client’s language, not the vendor’s. It is the same idea that guides me with other tools: not everything has to be built from scratch.

The eight analytics that actually make sense on a jobsite#

The platform offers a lot. We picked the ones that solve problems a contractor recognizes in five seconds:

• PPE detection. It flags anyone on site without a hard hat or a safety vest. Fun detail: a person without a vest is also a typical sign of an intruder.
• After-hours intrusion. If someone gets in while the site is closed, the system can play a voice warning over the speakers, or trigger lights or sirens. The idea is to deter before anything goes missing.
• Loitering. It spots someone who stays too long near materials or an entrance, from 1 to 2 minutes depending on the setup.
• People and vehicles. It tells people, vehicles and animals apart. It sounds minor, but it is what removes the false alarms from wind, shadows or a dog walking by, which are the ones that make people ignore everything.
• License plate recognition. It logs every truck coming in and out. For deliveries and subcontractors, that is gold.
• Fire and smoke. Visible flames and dense smoke, indoors or outdoors.
• People counting. How many people are in each zone throughout the day.
• Natural-language video search. My favorite. Instead of going through sixty hours, you type something like “white truck coming through the gate” and it brings back the matching moments. That Monday morning scene gets solved in minutes.

On top of the analytics come things that matter just as much: cloud recording (if someone steals the camera, the footage is no longer on site), camera trailers on cellular for sites without internet, and a time-lapse of the whole project, which is also a great way to show progress to the client.

What we do not turn on by default#

How you offer something matters as much as what you offer. There are three things that are not turned on in any project:

• Facial recognition. It carries serious legal implications and requires explicit consent. We only turn it on by request, when the case justifies it and after a legal review. Not lightly, just because the technology allows it.
• Weapon detection. It is not a typical jobsite problem, so it is not included. It is available on request for the sites where it does make sense.
• Crew attendance and productivity. BGS wants it and it makes a lot of sense, but the platform does not cover it the way we need yet. I would rather keep it as a draft than half promise it.

AI filters, a person decides#

One rule we set from day one: the AI reviews every event, describes it and suggests what to do, but a person confirms before anyone acts. AI is great at separating what matters from the noise. Deciding whether to call the police or send someone to the site is a different thing.

What we are still working through#

Not everything is settled, and I would rather say so:

• Connectivity. On sites without wired internet everything depends on the cellular signal. The upside is that the equipment stores footage locally if the connection drops and uploads what is missing once it is back.
• Privacy and legal questions. How long footage is kept, which notices need to be posted on site, and what the law says wherever the work happens. All of that goes through legal review before signing with each client.
• Real footage. We want to show real monitored jobsites, not stock images. That will come with the first projects.

What I take away#

The difference between a camera that records and a camera that speaks up is huge, and it is rarely about the hardware. It is about deciding what is worth detecting, filtering the noise well, and always having a person with good judgment at the end of the chain.

If you have a jobsite, a warehouse or a yard with cameras nobody is watching, write to me. There is probably a way to get them to speak up.