Automated License Plate Recognition for Resource-Constrained Environments

Area of Science:

• Computer Science
• Artificial Intelligence
• Embedded Systems

Background:

• Deep learning in embedded systems boosts edge computing.
• Resource-constrained edge devices face limitations with high-end hardware requirements for complex computations.
• Existing solutions often lack efficiency for low-power, limited-resource edge applications.

Purpose of the Study:

• To present a novel, hardware-efficient automated license plate recognition (ALPR) system.
• To demonstrate a proof of concept for ALPR on resource-constrained edge devices.
• To address the challenge of implementing advanced AI on low-power edge hardware.

Main Methods:

• Developed a novel set of neural networks tailored for diverse hardware configurations.
• Implemented a purely edge-based ALPR solution designed for low-resource environments.
• Validated models using diverse license plate datasets under varying daytime and nighttime conditions.

Main Results:

• Achieved robust performance across extreme illumination changes (day/night).
• Demonstrated competitive accuracy and energy efficiency compared to state-of-the-art server-grade systems.
• Validated low computational latency and communication overhead for real-time edge deployment.

Conclusions:

• The proposed hardware-efficient ALPR system is suitable for resource-constrained edge environments.
• Novel neural network designs enable generalizability across different low-cost hardware.
• The solution offers a viable alternative to power-hungry, server-based ALPR systems.