An Embedded Real-Time License Plate Recognition System for Complex Traffic Scenes

Conference

IEEE Intelligent Transportation Systems Conference (ITSC 2026)

Naples, Italy

Authors

Anuki Pasqual, Dulan Lokugeegana, Manimohan Thiriloganathan, Nuthya Rathnayake, Udaya S. K. P. Miriya Thanthrige, Kithsiri Samarasinghe

Keywords

Abstract

This work presents an embedded real-time license plate recognition (LPR) system customised for developing countries with complex, unstructured traffic scenes. The method uses a two-stage pipeline: license plate detection on multi-vehicle images followed by character recognition on the detected plates. Both steps use lightweight convolutional neural networks to balance accuracy and efficiency. The SL-LPR dataset of Sri Lankan road images is introduced, covering diverse vehicle types and traffic conditions typical of developing countries. The LPD model achieves 93.6% mAP and the LPCR model achieves 87.88% plate-level accuracy. Low-bitwidth quantization (Brevitas) and FPGA acceleration (FINN framework) are applied for real-time edge deployment, reaching 11.5 FPS at 87 ms max latency per frame on the Xilinx Kria KV260.

Key Specifications & Results

PlatformXilinx Kria KV260 (Zynq Ultrascale+ MPSoC)

LPD ModelAdapted LPYOLO — 4-bit quantized, 576×576 input, 0.363B ops

LPCR ModelAdapted fast-plate-ocr CNN — 1.02M parameters, 132M multiplications

Total Parameters~1.4M combined (268 KB after quantization)

LPD Accuracy (SL-LPR)93.6% mAP, 92.2% recall

LPCR Accuracy (SL-LPR)87.88% plate-level, 97.42% character-level

System Throughput11.5 FPS (max 87 ms/frame with 10 license plates)

Power Consumption4.2 W

FPGA Resource Utilization63% LUTs, 38% FFs, 99% BRAM, 29% DSPs

Dataset (SL-LPR)2970 multi-lane images, 3412 license plates, 5+ vehicle types, up to 8 LPs per scene

Quantization ToolsBrevitas (training) + FINN framework (FPGA compilation)

Dataset

The SL-LPR dataset was created by capturing images from overhead bridges on urban roads and expressways in Sri Lanka. It contains 2970 images with diverse vehicle types (cars, motorbikes, buses, trucks, three-wheelers), up to 8 license plates per scene, and covers both daytime and low-light conditions. Images were split 80:18:2 for training, validation, and testing. The character recognition subset contains 3412 license plates. SL-LPR outperforms public datasets like CCPD, PKU, and UFPR-ALPR in terms of scene complexity and vehicle variety.

Key Novelty

The system is the first to combine complex multi-vehicle Sri Lankan road scenarios with FPGA acceleration using the FINN framework, achieving real-time performance (11.5 FPS) on a low-power embedded platform. The adapted LPYOLO detection model uses a novel 8-bit quantized sigmoid (replacing HardTanh) reducing max pixel error from 15 to 3, maintaining 93.6% AP vs 52% with the original approach.

My Contributions

Conducted literature review on license plate recognition, vision-based speed estimation, and vehicle detection; verified existing methods by running reference code
Fine-tuned the LPD model (LPYOLO) and trained it on the custom Sri Lankan SL-LPR dataset
Implemented input image resizing for the LPD model in C (optimised for efficiency) and Halide language
Implemented Non-Maximum Suppression (NMS) in C
Developed input image resizing/cropping from LPD detections for the LPCR model
Integrated the full end-to-end pipeline and displayed outputs on monitor
Designed enclosures for field testing and the final prototype
Contributed to field testing, real-time camera integration, and documentation

Follow Me