Urban and commercial car parks face a simple but expensive problem: nobody knows which bay is free at any given moment. Drivers circle the floors, a queue builds at the entrance, the ticketing system has no connection to the bays, and management learns the day's entries and exits only after closing. Ground sensors for every bay are costly to install and maintain and struggle in open-air facilities. What is needed is a way to measure the state of every bay continuously, cheaply and without human involvement.
The importance is not only the driver's convenience. For the city, every minute a car spends hunting for a space adds traffic and emissions. For the operator, not knowing real occupancy means being unable to price dynamically, offer reservations or analyse usage patterns. For security, knowing which vehicle is in which bay since when is the basis for following up any incident. Smart parking is therefore not a convenience product; it is infrastructure that turns into data, and data that turns into decisions.
A computer-vision architecture turns existing cameras, or a small number of new ones, into sensors. At the edge, a compact processing unit analyses the frames, detects occupancy per bay and reads number plates, and sends only meaningful events, such as bay twelve occupied or plate X entered, to the server. Those events are collected on a central platform, linked to ticketing and payment, and exposed through APIs to guidance signs, the mobile app and the management dashboard. Raw video never leaves the site; only data does.
The first practical consideration is real-world conditions. Models that perform well in good light with a clean lens may fail at night, in rain, under tree shadows or when a camera has been knocked out of alignment. The remedy is evaluating the model on data from the actual site and retraining periodically, rather than trusting a generic model. For example, an open-air car park in winter has a completely different lighting pattern from summer, and accuracy should be expected to drop without retraining.
The second consideration is privacy and data design. A number plate is personal data and must be handled with the same care as customer data in a bank. That means processing at the edge, discarding the image once the event is extracted, retaining plates only as long as operations and payment require, and role-based access for the dashboards. A good architecture builds these constraints into the design rather than bolting them on afterwards, which is both safer and cheaper.
The common pitfalls are clear. First, starting with the best model instead of the best camera angle; correct camera placement affects accuracy more than any algorithm. Second, streaming raw video to a central server, which consumes bandwidth and creates privacy risk. Third, building the detection system without integrating it with ticketing and payment, which produces a dashboard nobody acts on. Fourth, ignoring the failure state; when a camera goes down, the system should report bays as unknown rather than free, or drivers will be guided to spaces that do not exist.
Niadad's vision AI platform Bina («بینا») implements this architecture: edge processing, occupancy and plate detection, event generation and integration with operational systems through APIs. In Niadad's smart-parking project, Bina serves as the perception layer, and its events feed Niadad's data and event platforms so that the state of the car park is available to the driver, the operator and city-level analysis alike.
Smart parking is a small instance of the larger smart-city pattern: inexpensive sensing, intelligence at the edge, events instead of video, and integration with operational systems. An organisation that gets this pattern right in one car park can repeat it at a building entrance, on a production line or in a warehouse. The technology is now mature enough that the decisive factor is no longer the model but the discipline with which the surrounding system is built.