AtalaIA : Aperfeiçoamento de sistema para reconhecimento automático de placas de licença automotiva e construção de um produto para segurança de áreas restritas

Abstract

Given the recent advancements in computer vision technology, it is necessary for security products and services to utilize this knowledge. By using deep learning, it is possible to develop intelligent tools that can operate in scenarios such as the enforcement of traffic laws through vehicle license plate recognition systems, facial recognition for access authorization to restricted areas, and identification of criminal behavior through the detection of suspicious actions. Security issues at entrances can arise from human errors in the decision-making process regarding granting access to individuals or vehicles, thus technological assistance is useful for error prevention. With this in mind, this work aims to enhance a system of Automatic License Plate Recognition (ALPR) developed using Convolutional Neural Network (CNN) models to increase security at establishment and condominium entrances. To achieve this enhancement, an object tracking module was developed in the system, leveraging data from previous image processing to recognize a new image of the same vehicle. This allows for a reduction in processing latency and enables the development of additional functionalities, such as temporal redundancy treatment, which enhances the accuracy of the ALPR system’s results, as well as vehicle counting functionality. Furthermore, a prototype IP camera was constructed using a NodeMCU Esp32Cam, aiming to provide a low-cost IP camera option. Interfaces were developed in the system for MQTT and RTSP communication between cameras and processing units, allowing the ALPR system to be used with various cameras available on the market compatible with the RTSP protocol, as well as the IP camera prototype developed using the MQTT protocol. Additionally, compression methods were applied to a YOLOv5 Nano model, such as pruning and quantization. As a result, it was possible to enhance the ALPR system by increasing the accuracy of the results from 27.18% to 56.05% through temporal redundancy treatment, and there was a 33% reduction in processing latency with the implementation of the tracking module.