Low-Power License Plate Detection and Recognition on a RISC-V Multi-Core MCU-Based Vision System

Jul 16, 2026

Upload by Skyler

The paper demonstrates a low-power MCU-based edge device for automatic license plate recognition (ALPR), using a 9-core RISC-V processor (GAP8) with a QVGA ultra-low-power greyscale imager. A multi-model inference visual pipeline employs SSDlite-MobilenetV2 for license plate detection and LPRNet for optical character recognition, achieving 38.9% mAP and >99.13% recognition accuracy on public datasets. On real data, it recognizes small LP crops down to 30×5 pixels. Compression and optimization yield 1.09 FPS at 117 mW, embedding high network complexity without hardwired accelerators and improving energy efficiency by 73× versus a Raspberry Pi 3 system.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages5

Reading Time13 min

Introduction

Related Works

This paper has been accepted for publication in the IEEE International Symposium on Circuits and Systems (ISCAS) ©2021 IEEE. Low-Power License Plate Detection and Recognition on a RISC-V Multi-Core MCU-based Vision System Lorenzo Lamberti∗ , Manuele Rusci ∗‡, Marco Fariselli‡, Francesco Paci‡, Luca Benini∗† ∗ Department of Electrical, Electronic and Information Engineering -University of Bologna, Italy †Integrated System Laboratory -ETH Zurich, Switzerland ‡GreenWaves Technologies, Grenoble, France arXiv :2607 .09768v 1 [ cs .CV] 7 Jul 2026 Abstract—In this paper, we present the first (to the best of our knowledge) demonstration of a low-power MCU-based edge device for Automatic License Plate Recognition (ALPR). The design leverages on a 9-core RISC-V processor, GAP8, coupled with a QVGA ultra-low-power greyscale imager. The proposed visual processing pipeline uses a multi-model inference approach based on SSDlite-MobilenetV2 for license plate detection and LPRNet for optical character recognition, reaching a 38.9% mAP score for the first task and a recognition rate of >99.13% for the latter on public datasets. On real-world data, the pipeline recognizes registration numbers when the size of LP crops is as small as 30×5 pixels. Thanks to the applied compression and optimization strategies, the multi-model inference (687MMAC) achieves a throughput of 1.09FPS at a power cost of 117mW when running on GAP8. Our solution is the first MCU-class device embedding such a level of network complexity, resulting to be 73 × more energy-efficient w.r.t. precedent mobile-class ALPR system featuring a Raspberry Pi3. The proposed design does not resort to any hardwired acceleration engines, thus retaining full flexibility for future algorithmic improvements. I. INTRODUCTION Automatic License Plate Recognition (ALPR) is an optical character recognition (OCR) task consisting of detecting and transcribing the vehicle’s License Plates (LPs) . Currently, Deep Learning (DL) algorithms are the state-of-the-art (SoA) solution to this problem [1] . However, because of the high complexity (>10GMAC) and stringent latency requirements, the deployment of ALPR pipelines is limited to high-performance embedded platforms that feature a high power envelope (>5 Watts) [2]–[4]: such systems would operate for no longer than 30 minutes when powered with a single AA battery, hence not suitable as an always-on smart sensing solution. Conversely, MicroController Units (MCUs) are the typical data processing engines for battery-powered systems, because of the low-cost and low-power characteristics (as low as few mW) and their SW programmability. However, MCUs, e.g. ARM Cortex-M4-based devices, present severe limitations in terms of on-chip memory budget (few MB) and computing resources (a single-core running up to few hundreds of MHz) that prevent the implementation of complex vision DL pipelines [5] . Indeed, a typical SSD object detector is estimated to execute in 3s at 262mW on an STM32H7 MCU [6], hence not meeting the requirements of low-power embedded devices. To tackle this challenge, in this work we propose a vision pipeline for ALPR, which relies on a 2-step DL-based algorithm composed of a SSD object detector [7] and LPRNet [8], tailored for LP detection and recognition respectively. To gain an ultra- high energy-efficient smart sensing solution, the design of the DL model is optimized for the deployment on an ultra-low power smart camera platform, which includes GAP8 [9], a Parallel Ultra-Low-Power (PULP) [10] architecture featuring an 8-core cluster that has a peak computing capability ∼1GMAC/s within ∼100mW of power envelope, and an ultra-low-power Himax image sensor. To fit the memory, latency, and energy consumption constraints of the target platform, we apply a multi-objective optimization procedure aiming at minimizing both the computational complexity (MAC operations) and the model size of a baseline solution [11], which exceed

FAQs

What hardware and platform does the proposed ALPR edge device use?

It uses a GAP8 9-core RISC-V processor with a QVGA ultra-low-power greyscale image sensor, targeting an MCU-based vision edge node.

Which models are used for license plate detection and character recognition?

SSDlite-MobilenetV2 is used for license plate detection, and LPRNet is used for optical character recognition.

How efficient and accurate is the proposed pipeline?

It reaches 38.9% mAP for detection and >99.13% recognition accuracy on public datasets, and it runs at 1.09 FPS with about 117 mW power cost on GAP8.

Jul 16, 2026

Upload by Skyler