Computer VisionShipped
License-Plate Recognition for Yard Gates
Nhận diện Biển số Xe Cổng Bãi
Truck ALPR at container yards — reads day & night white/yellow plates, auto-opens the barrier.
PaddleOCRPyTorchOpenCVFastAPITCP relay
Problem
A container yard processes 180 trucks/day; guards manually checked plates in 30-45 s per truck. Gate jams at peak. RFID was tried but external drivers don't carry tags.
Architecture
Hikvision 4K camera + IR fill light → Jetson Nano → 2-stage pipeline: PaddleOCR det (detect plate region) + custom CTC recognizer (trained on 28k VN plates: white/yellow/red/blue) → DB whitelist match → TCP relay opens barrier. Every truck snapshot stored 90 days for audit.
Stack & rationale
- PaddleOCR-det: detection robust to skew + dirty plates.
- Custom CTC over PaddleOCR-rec: VN domain has unique chars (Đ, Ô; no F), custom train is 8 points more accurate.
- IR fill light over thermal: cheap, enough for night ALPR.
Results (6-month run)
| Metric | Value |
|---|---|
| Avg gate time | 45 s → 6 s |
| Daytime read accuracy | 99.2% |
| Nighttime read accuracy | 98.1% |
| Peak-hour gate jams | gone |
| False positives (wrong open) | 2 (added manual confirm for new plates) |
Lessons
Training a VN-specific recognizer is worth the effort (generic PaddleOCR-rec scored only 91% on VN test set). IR light at 30° is better than straight-on (less driver-headlight glare).