Recognition systems that work in the field, with your cameras and your constraints.
Automatic number plate recognition, object detection, OCR, and integration of models into existing products. We train models on your real data rather than deploying a generic model that fails as soon as conditions change.
The difference between an AI demonstration and a system you can operate is measured in the field. A plate recognition model trained on generic data collapses when it meets Tunisian plates, a badly angled camera, rain, or low sun behind the vehicle.
Our approach is empirical. We start from your actual video streams, build an annotated dataset from your operating conditions, and train a dedicated model. On the Bizerte economic activity park project, this took recognition accuracy from roughly 50-65% with a generic model to approximately 95% with a model trained specifically on Tunisian plates, from 3,200 images.
We also handle the least visible and most decisive part: ingesting several simultaneous RTSP streams, latency, GPU hardware, monitoring, and how the system behaves when a camera drops.
Projects where this work was carried out.
Image to replaceEconomic activity park
PAEB — Parc d'Activités Économiques de Bizerte
A plate recognition system driving three barriers from six video streams, with a model trained specifically on Tunisian plates.
Logistics
It depends on how variable the conditions are. For the Tunisian plate model on the Bizerte project, 3,200 annotated images were enough to reach approximately 95% accuracy. A site with more camera angles or harder lighting will need more.
For real-time video, almost always on-premise. Pushing several RTSP streams to the cloud is expensive in bandwidth and adds latency that is incompatible with opening a barrier. We size a dedicated GPU workstation on site and send only the results upstream.
A scope to challenge, an architecture to validate, a system to take over — start by writing to us.
