TaskObject Detection, OCR
IndustryIntelligent Transportation
DeviceARM CPU, AMD Zynq UltraScale+
Context and Challenges
Our client had an effective but traditional solution based on image processing and classical machine learning. To stay competitive and aligned with the state of the art, they sought to enhance their system’s detection and reading capabilities.
The main challenges were twofold:
- Data scarcity: limited datasets from countries where the client had no current presence;
- Latency requirements: ensuring the system could accurately detect and read license plates from high-speed vehicles.
Solution
- We provided end-to-end AI integration, including:
- Data collection and annotation oversight to ensure high-quality training sets;
- Metric definition to track model performance effectively;
- Synthetic data generation to overcome geographic data gaps;
- Custom AI model design for both license plate detection and OCR;
- Optimized deployment: compiling the solution into a library tailored to the client’s ARM-based hardware.
Results
Our AI-powered solution introduced deep learning into the client’s product, significantly boosting automatic license plate detection and reading accuracy, even in regions without existing camera installations, such as parts of Africa and Asia. Crucially, these advancements were achieved without requiring any hardware upgrades, leveraging the client’s existing infrastructure.
