RISC-V MCU-based system performs ALPR at 117 mW, 73x more efficient than Pi

A team led by Lorenzo Lamberti (ETH Zurich and University of Bologna) has demonstrated the first automatic license plate recognition (ALPR) system running entirely on a microcontroller-class device. The design is built around GAP8, a 9-core RISC-V parallel processor, paired with an ultra-low-power QVGA greyscale image sensor. The pipeline uses a multi-model approach: SSDlite-MobilenetV2 for localizing license plates and LPRNet for optical character recognition. On standard benchmarks, the system reaches a 38.9% mean average precision for detection and over 99.13% recognition accuracy. Remarkably, it can read license plates as small as 30×5 pixels in real-world images.

To fit such complex neural networks (totaling 687 million multiply-accumulate operations) into a low-power embedded device, the team applied aggressive model compression and optimization. The resulting system processes 1.09 frames per second while consuming only 117 mW of power. Compared to a prior ALPR implementation on a Raspberry Pi 3, this RISC-V solution is 73 times more energy-efficient. Crucially, the design does not rely on dedicated hardware accelerators, meaning it remains fully programmable for future algorithmic updates. The work was accepted at IEEE ISCAS 2021 and demonstrates that even complex vision pipelines can now run on ultra-low-power microcontrollers, opening the door to pervasive, battery-powered surveillance and smart city applications.