We built and deployed a detection and OCR pipeline on NVIDIA Jetson hardware at each gate. Reads are validated against plate-format rules, with uncertain results routed for review.
- 99.4% plate-read accuracy across the deployment’s tested conditions and formats.
- 45 ms from camera frame to verified plate string, with no cloud dependency in the critical path.
- 10 weeks from kickoff to live toll-gate deployment.
Technical approach: plate recognition
Camera frames pass through plate detection, perspective correction, denoising and contrast normalisation, then character recognition and confidence-gated format validation.
We collected and annotated site footage across regional plate formats, night conditions, rain and peak traffic. The recogniser was trained on regional fonts and degraded plates; TensorRT runtimes support inference on Jetson hardware.
Local VMS integration, alerting and failover logging keep the gate operating during connectivity outages. Low-confidence reads remain available for human review.