Most camera systems are recording devices. You only interact with them after something has happened, and then you scrub through hours of footage hoping the angle was right.

Kissimmee Bay had sixteen cameras across two vehicle gates, a guard house and a marina. All of them recording. None of it searchable in any useful way, and no alerting anyone had confidence in.

What follows is what we built instead.

Part one: the gates read plates now

Four of the sixteen cameras are licence plate readers covering both entrances. They had never successfully read a plate — a story we have written up separately, because the fix turned out to be configuration rather than the licence everyone assumed was missing.

Getting them reading was the beginning. A camera that recognises a plate still leaves you with a recorder full of clips and no way to answer “was this vehicle here on Tuesday?”

So we built a plate log that runs continuously:

• Motion-gated polling. Rather than analysing every frame of four cameras around the clock, it watches for motion and only then pulls an image. That keeps both the processing cost and the cellular data cost sane — which matters on a site where bandwidth is genuinely finite.
• Reading from the full-resolution frame. Our first approach downscaled and cropped before reading, which is the obvious efficiency. It returned nothing. Plates are small, the detail is in the pixels you just threw away, and the fix was to stop being clever about it.
• A deduplicated registry. One vehicle passing a gate generates many frames. The log records the vehicle, not the frames, and only when the read is confident. An uncertain read is discarded rather than guessed at.

That last point is deliberate. A plate log that invents a character to fill a gap is worse than no log, because someone will eventually rely on it.

Part two: the marina watches itself overnight

The marina sits apart from the gates, on its own connection, with cameras covering the dock ramp, parking, boat storage, the entrance and the waterfront pad.

It runs local AI detection on the site itself — person, vehicle and boat classification on 4K streams, processed on the marina’s own hardware rather than shipped elsewhere. On a typical day that produces around a hundred detection events.

A hundred events a day is not an alerting system. It is a way to train people to ignore alerts.

So the overnight watch is narrow by design: between midnight and six, people and vehicles only, in the areas where neither should be. Everything else is logged and left alone.

Part three: the problem nobody expects

Here is where it got interesting, and it is the part we are most pleased with.

An alert with a bad photo is nearly useless. The system would faithfully detect a person, grab a frame, and send a picture of the back of someone’s head as they walked out of shot. Technically correct. Practically worthless.

The reason is timing. Detection fires when something enters the frame, but the identifying view — a face, the front of a vehicle — often happened moments earlier, on a different camera, as they arrived.

So the alerting does something less obvious:

• When a subject is detected, it gathers footage of that subject across every marina camera, not just the one that triggered
• It looks back ninety seconds, deliberately, to catch the arrival rather than the departure
• It extracts full-frame stills from those clips at real resolution
• Then an AI model reviews the candidates and picks the single most identifying one — favouring a front-facing face or the front of a vehicle, rejecting rear shots and figures walking away

The alert that goes out carries that frame. Not the first frame, not the triggering frame — the useful one.

If a subject only ever crosses one camera walking away, no identifying frame exists and the system says so rather than sending something misleading. That is a physical limit, not a software one, and pretending otherwise would be the same mistake as guessing at a plate.

Part four: it is searchable

All of it feeds a single portal. Plate history from the gates, detection events from the marina, and the ability to ask a question rather than scrub a timeline.

That is the difference between a camera system and a security system. The cameras were always recording. What changed is that someone can now get an answer out of them in under a minute.

What this actually is

It is worth being precise, because “AI-powered” attaches itself to a lot of things that are not.

There is no AI in the cameras beyond plate recognition. The intelligence sits in the layer we built on top: deciding when to look, what is worth reading, what is worth waking someone for at 3am, and which single image out of many is the one worth sending. Those are judgement calls, made continuously, that would otherwise need a person watching sixteen screens.

It also runs on a real site with real constraints — metered connectivity, hardware in outdoor enclosures, a network that predates the project. Anything that only works in ideal conditions does not work here.

Why we are telling this story

We build automation into our own operations before recommending it to anyone. This is the same idea pointed at a client’s problem: not replacing anything, just handling the part that no human was realistically going to do — watching sixteen cameras all night and deciding what matters.

If you have cameras that record but cannot answer a question, or alerts your team has learned to ignore, that gap is closable and it usually does not need new hardware.

Talk to us

iTech Plus supports communities, associations and businesses across Central Florida — Davenport, Haines City, Kissimmee, Lakeland, Orlando and the surrounding Polk and Osceola county areas. We work on camera and access systems, networks, and the automation layer that makes them useful.

Call (321) 221-7117 or email info@itechplus.co.