License Plate Recognition
License Plate Recognition (LPR) is an AI algorithm that detects the license plate region from vehicle images and outputs the plate number as text. It is an essential technology for advanced traffic infrastructure including parking management, vehicle tracking, traffic violation enforcement, and ETC gates.
This algorithm is based on a Chinese license plate recognition model. Support for Japanese license plates requires additional training and customization based on Japanese plate datasets and recognition rules.
Algorithm Overview
The system consists of the following three-stage pipeline:
- Plate Detection (LPR Det): Identifies the license plate region in the image using a lightweight CNN
- Plate Classification (LPR Cls): Tilt correction and plate type determination (single-row blue, single-row yellow, new energy, training vehicle, etc.)
- Number Recognition (LPR Rec): Directly outputs the plate number string via end-to-end OCR (no character segmentation required)
Supported Plate Types
- Single-row blue plates
- Single-row yellow plates
- New energy vehicle plates
- Training vehicle plates
- White police, embassy/Hong Kong-Macau, double-row yellow, armed police plates (limited support)
Performance Metrics
| Algorithm | Recognition Rate |
|---|---|
| EAI-LPR | 98% (entry/exit scenarios) |
Edge AI Board (RV1126B) Execution Efficiency
| Processing Stage | Model Size | Processing Time |
|---|---|---|
| Plate Detection (lpr_det) | 2.64MB | 14.3ms |
| Plate Classification (lpr_cls) | 1.02MB | 3ms |
| Number Recognition (lpr_rec) | 5.19MB | 10.2ms |
| Total | 8.85MB | Approx. 28ms |
Key Features
- End-to-end recognition: Direct plate number output without character-level segmentation
- High accuracy: 98% recognition rate in entry/exit scenarios
- Lightweight model: Compact total size of 8.85MB
- High-speed processing: Total inference time of approximately 28ms across three stages
Use Cases
- Parking lot entry/exit management (plate authentication)
- Vehicle access control (gate operation)
- Automatic traffic violation enforcement
- Vehicle trajectory tracking and anomaly detection
- Traffic volume statistics and analysis
- Smartphone app vehicle registration
Edge AI Board Implementation
The three-stage LPR pipeline runs on the RV1126B NPU in approximately 28ms. It is a compact implementation that can be embedded into parking lot gates and roadside units.