Illustration of a car seen from behind inside a detection frame, titled "ALPR - Lettura Targhe Intelligente".
ML & Recognition

ALPR System

License Plate Recognition

<100ms
Latency
<€200
Camera Cost
IT + abroad
Plates
Edge or server
Architecture

Overview

The Problem

Commercial license plate recognition systems cost between €5,000 and €10,000 per camera and are closed systems: proprietary hardware, proprietary software, no integration possible. For a 50-space parking lot or a condo with an automatic gate, the cost is prohibitive. The alternative was manual management or standard cameras with no automation.

The technical problem is more subtle than it seems: a plate needs to be read in real conditions. Night with blinding headlights, rain, mud-covered plates, side angles, fast-moving vehicles. Generic models don't hold up under these conditions.

The Solution

I developed a custom system that works with normal IP cameras, even cheap ones under 200 euros. The core is YOLO11, trained on photos of Italian and foreign plates collected in every light, weather, and angle condition. The model locates the plate in the frame, crops it, and passes it to the OCR layer for text conversion.

The architecture is flexible: a central server for processing streams from multiple cameras, or edge computing directly on the camera for sub-100ms latency without network dependency. A web dashboard manages transit logs, subscriber whitelists, and hourly traffic statistics.

The Result

The system is designed for parking lots with automatic barriers, condos with resident access, restricted traffic zones and warehouses tracking vans in and out. The cost per camera drops from thousands of euros to a few hundred.

Key Features

  • Recognition of Italian and foreign plates with YOLO11
  • Cheap IP cameras (even under €200) instead of proprietary systems (€5k-10k)
  • Flexible architecture: central server or edge computing <100ms on the camera
  • Web dashboard with transit logs, whitelists and hourly statistics

Technologies Used

YOLO11OCREdge ComputingReal-time