Digital twins are often presented as showpieces: a polished 3D model that mirrors a building, a production line or a city block in real time. In my experience, the real value of a twin is less glamorous. A good twin tells you the truth about how your facility actually runs, including the parts that are not ready for automation.
A study my research group recently published in the open-access journal Smart Cities shows this clearly. With colleagues at De La Salle University and partners in the Asia-Europe for Artificial Intelligence (AE4AI) Network, we built a smart parking management system for an existing 30-slot facility, without new construction.
What we built
Many parking facilities still log vehicles by hand and know little about how their spaces are used. Our system connects three parts:
- Machine vision. Cameras detect vehicles in each slot and read license plates at the entrance and exit.
- A building information model (BIM). We laser-scanned the facility and turned the scan into a 3D model. Occupancy updates appear on the model as vehicles arrive and leave.
- A data dashboard. Managers see occupancy, turnover, peak hours, dwell time, parking fees and revenue in one place.
The AI performed well. Vehicle detection reached a mean average precision of 94.86%, license plate detection 99.27%, and plate recognition 90.50% accuracy. Occupancy tracking based on detecting the vehicle itself reached 94.86% accuracy; tracking based on reading license plates reached about 89.9%.
What the twin revealed
The more useful findings were about everything around the AI.
- Infrastructure, not just intelligence. Without a barrier, cars entered and left without slowing down, and standard 30-frames-per-second cameras captured blurred plates. At one point the dashboard showed more than 30 vehicles in a lot with 30 spaces. A gate barrier or speed bump that makes vehicles pause would fix this more cheaply than buying faster cameras.
- Light and line of sight. Strong sunlight reflecting off plates made them unreadable, and low-mounted cameras were easily blocked. Camera placement turned out to be a design decision, not an installation detail.
- The right signal for the job. People walking past parked cars briefly hid the plates, so the system logged false departures and returns. Vehicle detection is the better signal for deciding whether a space is occupied; plate reading is better suited to identifying a vehicle once, when it arrives.
- Pilots hide complexity. The prototype was reliable in a controlled setting. A commercial facility with several entrances, multiple levels and unpredictable traffic needs a planned camera network and more computing capacity before it can scale.
- Platform lifespan matters. Design software versions lose support over time. Choosing a twin platform with long-term support and backward compatibility protects the investment.
The paper calls the principle behind these findings reciprocal flexibility. New technology should not disrupt operations that already work, but operations also need to adjust so the technology can do its job: better lighting, sensible camera positions, a controlled entry point.
Why this matters beyond parking
The same pattern shows up in factories, warehouses, campuses and hospitals. A vision system on a production line struggles with glare and occlusion just as a parking camera does. A twin of a plant floor will expose the gap between documented processes and what operators actually do. The study also points to natural next steps for the same approach: building energy management, airflow systems, manufacturing process optimization and predictive maintenance.
Four questions to ask before you build a twin
- What decision will the twin support? Occupancy, throughput, energy or maintenance. Pick one and design the data around it.
- Is the physical environment ready? Check lighting, camera or sensor placement, and whether objects pause long enough to be measured. Fixing these is often cheaper than upgrading the model.
- Which signal fits each decision? Combine sensors and models deliberately, based on what each one does reliably.
- Can it scale and last? Plan the sensor network, computing capacity and software support for the full facility, not only the pilot area.
Treat the twin as a diagnostic first. If it shows you where your real operations differ from what automation assumes, it has already paid for itself.