Automated License Plate Recognition under Lighting Constraints at the Edge
The technological advancements have risen the growth in computer-intensive applications for smart cities and smart homes, where edge computing plays a significant role to provide effective and efficient solutions. The incorporation of deep-learning techniques in embedded systems enhances the capabilities of edge computing. Most of these solutions rely on high-end hardware and often require a high processing capacity, which cannot be achieved with resource-constrained edge computing. This study presents a novel approach and a proof-of concept for a hardware-efficient automated license plate recognition system for a constrained environment with limited resources. The proposed solution is purely implemented for low resource edge devices and performed well for extreme illumination changes such as day and nighttime. The generalisability of the proposed models have been achieved by using a novel set of neural networks for different hardware configurations based on the computational capabilities and low cost. The accuracy, energy efficiency, communication and computational latency of the proposed models are validated using different license plate datasets in the daytime and nighttime and in real-time. The obtained results have shown competitive performance to the state-of-the-art server-grade hardware solutions.
This paper proposes an Automated License Plate Recognition (ALPR) solution for edge computing with resource-constrained environments, which can lead to support smart city development and management processes. Although ALPR is a well-established area in the domain of image processing, research on ALPR is still challenging with the associated constraints in the environment such as varying weather conditions, plate variations across regions, vehicle motion, distorted characters, dirty plates, shadow and reflection. Moreover, most of the existing ALPR solutions are limited to execution in server-grade hardware with nearly unlimited resources and limited to daytime performance. Thus, currently, there has been less attention paid to build systems that work efficiently in constrained environments targeting low cost, energy efficiency, less computational power requirements, remote location deployments and work in night vision. The technological developments of deep learning techniques can be improved to use in edge devices to provide an efficient solution for ALPR in resource-constrained environments.
We present an approach and a proof-of-concept prototype for hardware-efficient ALPR at nighttime, while adhering to several constraints in terms of energy efficiency, resource utilization, low cost, low-latency communication and computation as the novel contributions. The proposed ALPR system can operate at nighttime without any visible additional illumination and require no internet connection for operation. Consequently, the system is fully implementable on low power edge devices like Raspberry Pi 3b+ and operated completely with a battery that lasts long due to the energy-saving strategies implemented in the solution. Therefore, the system recognizes license plates in real-time both day and night-time, and can be deployed in rural or forest areas, where there is no stable internet connectivity or a direct power grid, which is one of the main contributions of this study.
Our methodology uses deep learning based Neural Architecture Search (NAS) strategies to discover a novel set of hardware-efficient neural networks for autonomous management of license plate detection and recognition process for edge devices with low resources. However, the main limitation to train the discovered deep neural networks for the task of recognizing license plates is the lack of a large, annotated and diverse dataset. In order to circumvent this issue,we use a synthetic data generation process based on image-to-image translation techniques to convert daytime RGB (Red-Green-Blue) images into thermal infrared (TIR) images. Thus, the implementation of the NAS based data engineering techniques in IoT applications is one of the scientific contributions of this study.
The prototype of our solution simulates a case study of an animal poacher vehicle detection problem. At present, Wildlife has faced a capacious and prejudicial issue that has caused a countable number of wild animals to lose their lives. Most of the existing approaches to minimize illegal hunting of wild animals, rely on manual surveillance from the camera feeds. Poacher vehicle detection system uses modern image processing and deep learning techniques to detect poacher vehicles while tracking their license plate numbers and sending the detected vehicle details to authorized parties through SMS. It has been noticed that poachers arrive mostly at nighttime since the poacher vehicle detection system is designed to function at nighttime as well. The case study environment contains several constraints. This system relies on battery power only, thus the power consumption should be minimized. Since there is no internet connectivity in the wild, SMS is the only possible communication method, where images can be stored for later prosecution material. Also, the system should be deployed in an unnoticeable way to the poachers.