One camera. Every space, counted.
We wanted to find out whether a single camera, watching a car park watches a car park the way a person would — and knows, at any moment, exactly how many spaces are free.
Anyone who has circled a full car park looking for a space knows the frustration. The usual fix is a sensor buried under every single bay — reliable when it works, expensive to install, and quietly unreliable once a few units start failing underground. We wanted to know whether one camera, watching the way a person already would, could do the job instead.
What it sees
Point it at a live feed, and it learns the shape of every bay in view. From there it watches continuously – no new hardware in the ground, nothing bolted to a bumper. Just the footage the site already has, or a camera you'd install once.
| Ordinary camera feed | What the algorithm understands |
|---|---|
 |  |
Same footage, two views: the raw feed on the left, the system's read of every bay on the right.
What it does
- Counts free spaces in real time – the one number a driver actually needs before deciding whether to pull in.
- Tracks how long each vehicle has stayed – so long-stay and overstaying vehicles can be flagged automatically, without anyone walking the rows with a chalk.
- Ignores the noise — a passing pedestrian, a shadow, a reflection off a windscreen — so it doesn't cry wolf over nothing.
Results, from real footage
We tested against camera footage from three different car parks — different layouts, different camera angles, different levels of activity — rather than one clean, controlled example.
| 3 | 56 | 12 min |
| Real car parks tested | Spaces watched by one camera | Continuous busy-lot footage |
The 56-space lot pictured above under "What It Sees" was that busiest test: one camera, one vantage point, the kind of coverage that would otherwise mean 56 separate sensors. Across that stretch of footage, the system kept an accurate, continuously updating count as availability shifted between roughly a quarter and a third of the lot being free. Just as importantly, when we checked its calls against the original footage by hand — a car pulling into an empty bay, another backing out and leaving — the timing held up.
Built for real time
We ran the same system on a dedicated Linux server to see what real-time monitoring actually costs in hardware. The answer: not much. Nothing exotic or purpose-built — the kind of graphics card you'd find in a mid-range gaming PC handled it with room to spare.
Test rig — real-time, headroom to spare
| Processor | 9th-gen Intel Core i7 |
| Graphics | NVIDIA GeForce RTX 3060 |
| Video memory | 12 GB (~4% in use while running) |
| System memory | 64 GB (13 GB in use while running) |
That headroom matters more than the raw speed: it means the same single, modest machine has plenty of capacity left to watch several car parks at once, rather than needing dedicated hardware bolted on for every additional site.
Why it matters
A sensor above and under every bay means a costly install and an ongoing maintenance burden — and a single failed unit quietly reports a phantom full space until someone notices. A camera-based system covers dozens of spaces from one vantage point, extends to a new car park without new infrastructure, and — as this testing showed — holds up across genuinely different real-world conditions, not just a single tidy demo.