Pasqual et al.'s license plate system hits 11.5 FPS on FPGA

A team of researchers led by Anuki Pasqual from Sri Lanka has developed an embedded real-time license plate recognition system designed specifically for the chaotic, unstructured traffic scenes common in developing countries. The work, accepted at IEEE ITSC 2026, tackles two core challenges: detecting license plates in multi-vehicle images and recognizing characters on those plates. Both steps rely on lightweight convolutional neural networks (CNNs) to balance accuracy with the strict compute and power budgets of embedded devices. To handle the lack of representative training data, the team created the SL-LPR dataset, capturing a wide variety of vehicle types (rickshaws, buses, trucks) and lighting conditions typical of roads in Sri Lanka and similar regions.

On the SL-LPR dataset, the detection model reached 93.6% mean average precision (mAP), and the character recognition model achieved 87.88% accuracy—competitive with larger models on public benchmarks. To enable real-time performance on a resource-constrained device, the researchers employed low-bitwidth quantization using the Brevitas library and synthesized the models onto a Xilinx Kria KV260 FPGA via the FINN framework. The end-to-end system operates at 11.5 frames per second (FPS) on that platform, sufficient for real-time applications like automated toll collection, traffic monitoring, and law enforcement. The SL-LPR dataset is publicly available to foster further research in low-cost intelligent transportation for emerging economies.