Monitoramento inteligente do perímetro operacional terrestre do Centro de Lançamentos de Alcântara utilizando processamento de imagens e inteligência artificial

• Ano de defesa: 2022
• Autor(a) principal: VALE, Juan Pablo do Nascimento
• Orientador(a): BARRADAS FILHO, Alex Oliveira
• Banca de defesa: BARRADAS FILHO, Alex Oliveira , BRAZ JUNIOR, . Geraldo , BARROS FILHO, Allan Kardec Duailibe , GOMES JÚNIOR, Daniel Lima
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA AEROESPACIAL/CCET
• Departamento: DEPARTAMENTO DE ENGENHARIA DA ELETRICIDADE/CCET
• País: Brasil
• Palavras-chave em Português: Centro de Lançamento de Alcântara; sistema de monitoramento; reconhecimento facial;
• Palavras-chave em Inglês: Alcântara Launch Center; smart security; face recognition.
• Área do conhecimento CNPq: Engenharia Aeroespacial
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/4055

The main space launch base in Brazil is the Alcântara Launch Center, which, due to its territorial extension, sensitive work, and security requirements, requires an automated form of perimeter control. Currently, new forms of perimeter monitoring are emerging based on artificial intelligence technologies, the so-called Smart Surveillance. In this work, we propose an embedded intelligent monitoring system based on facial detection and recognition, to control critical regions of terrestrial space at the Alcântara Launch Center. Embedded hardware is responsible for capturing and transmitting images, while processing is done on a central server. For facial detection, the Viola-Jones and Multi-Task Cascade Convolutional Neural Network (MTCNN) algorithms are used, with F1-Score of 61.28% and 87.29% respectively in the DroneFaces database. In the facial recognition process, FaceNet was used to extract features, and the kNN, SVM and Random Forest models were compared for classification. In facial recognition quality tests, an ROC-AUC of 0.999 was achieved using the SVM classifier in the FaceScrub database, and using the Random Forest classifier in the DroneFace database, an F1-Score of 56.6% was reached. In real-time tests, the system averages 5.8 fps using Viola-Jones, Facenet and Random Forest, and an average of 2.3 fps using MTCNN, Facenet and Random Forest.