Loading...
Publication
Identificação de embarcações usadas para contrabando humano com recurso a veículos não tripulados e sistemas multiagentes
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 15.37 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
O contrabando de migrantes por via marítima é um dos grandes desafios à segurança de fronteiras. O controlo das extensas áreas oceânicas exige soluções inovadoras que ultrapassem as limitações dos métodos convencionais, sobretudo no que diz respeito à cobertura e rapidez na resposta. A incorporação de veículos aéreos não tripulados (UAVs) com apoio a sistemas multiagentes apresenta uma hipótese de melhorar a vigilância marítima, possibilitando operações mais autónomas, escaláveis e resilientes. Esta dissertação tem como proposta uma solução que envolve o uso de UAVs num sistema de multiagentes heterogéneo, apoiados por algoritmos de visão computacional. A arquitetura inclui vários tipos de agentes (patrulha, acompanhamento, apoio e comunicações) organizados por uma estação terrestre, garantindo redundância e a robustez das operações. Modelos de deep learning da família YOLO (v8, v11 e v12, variantes nano e small) foram utilizados para fazer a identificação das embarcações suspeitas. Esses modelos foram treinados com recurso à framework Ray Tune e através de um dataset próprio, desenvolvido especificamente para esta dissertação e aperfeiçoado com técnicas de aumento de dados (data augmentation). Os resultados experimentais revelaram que os modelos, como o YOLOv11n, atingiram elevados níveis de desempenho (F1≈0,8569; mAP@50≈0,8832), ultrapassando outras versões testadas. A análise comparativa também demonstrou que as variantes small conseguem oferecer um
melhor equilíbrio entre precision e recall, enquanto as variantes nano são mais apropriadas para uso em ambientes com recursos computacionais limitados, pelo nível de performance em termos de processamento de frames por segundo. Embora os resultados tenham sido entusiasmantes, a pesquisa enfrentou desafios
consideráveis, tais como o tamanho reduzido do conjunto de dados, a falta de imagens reais de operações marítimas no contexto esperado e limitações de hardware. Estes fatores restringiram a capacidade de treino dos modelos e destacam a importância de reunir mais dados em circunstâncias operacionais reais, além de utilizar recursos com maior capacidade de processamento. Mesmo assim, a aplicação de técnicas de aumento de dados e a implementação de frameworks para auto fine-tuning de hiperparâmetros permitiram contornar essas limitações de forma parcial, confirmando a abordagem proposta. Em termos práticos, a solução aqui apresentada contribui para reforçar a eficácia da vigilância marítima, oferecendo às autoridades competentes uma ferramenta tecnológica de auxílio à tomada decisão que alia rapidez e precisão. Do ponto de vista científico, este trabalho
acrescenta evidência sobre a aplicabilidade de sistemas multiagentes em circunstâncias críticas, bem como a viabilidade de aplicar algoritmos de deteção de objetos em tempo real em dispositivos não tripulados.
Em suma, os resultados desta dissertação ressaltam o valor da aplicação de UAVs integrados a sistemas multiagentes para a vigilância marítima, ajudando a reforçar a sua efetividade no combate ao contrabando humano e a enfatizar a necessidade de implementação de soluções inovadoras que sejam simultaneamente éticas, seguras e ajustadas aos princípios de proteção de dados.
One of the biggest threats to border security is the smuggling of migrants by sea. Large-scale oceanic surveillance requires innovative solutions that get around the disadvantages of traditional techniques, especially regarding coverage and reaction time. A new strategy for improving maritime surveillance is the integration of unmanned aerial vehicles (UAVs) supported by multi-agent systems, which allows for more independent, scalable, and robust operations. This dissertation proposes a solution based on the incorporation of UAVs into a heterogeneous multi-agent system. Operational robustness is ensured by the architecture's multiple agent types (patrol, tracking, logistic support, and communications), which are managed by a ground station. To find suspicious vessels, deep learning models from the YOLO family (v8, v11, and v12, in both nano and small variants) were used. The Ray Tune framework and a specially created dataset for this study, enhanced by data augmentation methods, were employed to train these models. According to experimental results, models like YOLOv11n outperformed other tested versions and demonstrated a better performance (F1≈0.8569; mAP@50≈0.8832). The comparative analysis also revealed that the nano variants' higher frames-per-second (FPS) performance made them more appropriate for environments with constrained computational resources, while the small variants offered a better balance between precision and recall. Despite the promising results, the study had several obstacles, including a small-scale dataset, a lack of accurate maritime imagery in the intended context, and hardware constraints. These factors limited the model’s training capacity and emphasized the importance of using greater computational resources and gathering more data under actual operating conditions. However, these limitations were somewhat attenuated by the implementation of data augmentation and automated hyperparameter fine-tuning frameworks. Practically speaking, the solution helps to increase the efficacy of maritime surveillance by providing authorities with a technological tool that supports decision-making by combining speed and precision. Scientifically speaking, this work provides further evidence of the viability of implementing real-time object detection algorithms on unmanned platforms and the suitability of multi-agent systems in critical situations. In conclusion, this dissertation's findings demonstrate the importance of UAVs incorporated into multi-agent systems for maritime surveillance, highlighting their potential to prevent human smuggling. They also highlight the demand for innovative solutions that follows data protection laws while also being legal and security robust.
One of the biggest threats to border security is the smuggling of migrants by sea. Large-scale oceanic surveillance requires innovative solutions that get around the disadvantages of traditional techniques, especially regarding coverage and reaction time. A new strategy for improving maritime surveillance is the integration of unmanned aerial vehicles (UAVs) supported by multi-agent systems, which allows for more independent, scalable, and robust operations. This dissertation proposes a solution based on the incorporation of UAVs into a heterogeneous multi-agent system. Operational robustness is ensured by the architecture's multiple agent types (patrol, tracking, logistic support, and communications), which are managed by a ground station. To find suspicious vessels, deep learning models from the YOLO family (v8, v11, and v12, in both nano and small variants) were used. The Ray Tune framework and a specially created dataset for this study, enhanced by data augmentation methods, were employed to train these models. According to experimental results, models like YOLOv11n outperformed other tested versions and demonstrated a better performance (F1≈0.8569; mAP@50≈0.8832). The comparative analysis also revealed that the nano variants' higher frames-per-second (FPS) performance made them more appropriate for environments with constrained computational resources, while the small variants offered a better balance between precision and recall. Despite the promising results, the study had several obstacles, including a small-scale dataset, a lack of accurate maritime imagery in the intended context, and hardware constraints. These factors limited the model’s training capacity and emphasized the importance of using greater computational resources and gathering more data under actual operating conditions. However, these limitations were somewhat attenuated by the implementation of data augmentation and automated hyperparameter fine-tuning frameworks. Practically speaking, the solution helps to increase the efficacy of maritime surveillance by providing authorities with a technological tool that supports decision-making by combining speed and precision. Scientifically speaking, this work provides further evidence of the viability of implementing real-time object detection algorithms on unmanned platforms and the suitability of multi-agent systems in critical situations. In conclusion, this dissertation's findings demonstrate the importance of UAVs incorporated into multi-agent systems for maritime surveillance, highlighting their potential to prevent human smuggling. They also highlight the demand for innovative solutions that follows data protection laws while also being legal and security robust.
Description
Keywords
UAVs Multi-Agent Systems Maritime Surveillance Human Smuggling Computer Vision Deep Learning YOLO Ray Tune