| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 24.51 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Os últimos 70 anos têm vindo a registar um aumento de detritos plásticos encontrados no oceano. Estes englobam uma quantidade considerável de lixo derivado de atividades piscatórias, sendo estas maioritariamente artes de pesca perdidas ou deliberadamente abandonadas, denominadas de Ghost Nets. Dada a necessidade da preservação do ambiente, um Autonomous Underwater Vehicle (AUV) poderá servir
como plataforma para a deteção e recuperação destas redes. Assim, a presente dissertação insere-se no contexto do projeto NetTag. Esta propõe o desenvolvimento de uma solução simulada para a navegação do AUV Inspection and Recovery Intelligent System (IRIS) baseada na deteção visual de Ghost Nets com recurso a Deep Learning, a fim de conduzir o veículo ao encontro das mesmas para a sua posterior recuperação.
Desta forma, foram desenvolvidos os diferentes componentes do sistema, tendo por base o Stonefish para o ambiente de simulação. Este é lançado e gerido por um orquestrador que controla o modelo do IRIS com base nas inferências executadas pela sua câmara frontal. Estes valores traduzem-se, posteriormente, em comandos destinados aos thrusters. Por sua vez, as inferências fundamentam-se no modelo YOLOv5 nano treinado com um dataset criado para esta dissertação. Após a validação isolada de cada componente, procedeu-se ao teste da simulação completa que incorporou todos os componentes desenvolvidos. Foram obtidos resultados satisfatórios, tendo-se verificado a aproximação do submarino à rede de forma autónoma.
The last 70 years have been registering an increase of plastic debris found in the ocean. These encompass a considerable amount of waste derived from fishing activities, being mostly lost fishing nets or deliberately abandoned, called Ghost Nets. Given the necessity to preserve the environment, an AUV may serve as a platform for the detection and recovery of these nets. Therefore, this dissertation should be seen in the context of the NetTag project. It proposes the development of a simulated solution for the navigation of the AUV IRIS based on the visual detection of Ghost Nets using Deep Learning, in order to lead the vehicle to meet them for later recovery. In this way, the different components of the system have been developed, having Stonefish as the basis of the simulation environment. This is launched and managed by an orchestrator that controls the IRIS model based on the executed inferences through its frontal camera. These values are later translated into commands destined to the thrusters. The inferences are fundamented on the YOLOv5 nano model, trained with a custom dataset created for this dissertation. After the isolated validation of each component, the full simulation that incorporated all the developed components was tested. Satisfactory results were achieved, by having verified the autonomous approach carried out by the submarine to the net.
The last 70 years have been registering an increase of plastic debris found in the ocean. These encompass a considerable amount of waste derived from fishing activities, being mostly lost fishing nets or deliberately abandoned, called Ghost Nets. Given the necessity to preserve the environment, an AUV may serve as a platform for the detection and recovery of these nets. Therefore, this dissertation should be seen in the context of the NetTag project. It proposes the development of a simulated solution for the navigation of the AUV IRIS based on the visual detection of Ghost Nets using Deep Learning, in order to lead the vehicle to meet them for later recovery. In this way, the different components of the system have been developed, having Stonefish as the basis of the simulation environment. This is launched and managed by an orchestrator that controls the IRIS model based on the executed inferences through its frontal camera. These values are later translated into commands destined to the thrusters. The inferences are fundamented on the YOLOv5 nano model, trained with a custom dataset created for this dissertation. After the isolated validation of each component, the full simulation that incorporated all the developed components was tested. Satisfactory results were achieved, by having verified the autonomous approach carried out by the submarine to the net.
Description
Keywords
Ghost Net Ghost Fishing Simulation Stonefish AUV IRIS Deep Learning YOLO ROS Simulação