Publicação
Maritime Object Detection using Deep Learning for USV
| Resumo: | Deep Learning has been well known in the field of computer vision for its accuracy in detecting objects in images and videos streams; its has been being use for UAVs to detect objects in mid-air, or in self-autonomous cars, which are becoming a reality in the upcoming years. However, there have been a lack of research done with Unmanned Surface Vehicles (USV), used in maritime environments, in comparison with the other two fields mentioned before. These happens due to the harsh environment on the sea and the lack of specialized datasets, making it difficult to train deep learning algorithms that can detect and identify objects with high accuracy. Also, implementing an system with autonomous capabilities also comes with many difficulties regarding autonomy and performance, given the high computing power they need to operate. Nonetheless, the creation of an autonomous vehi- cle that is capable to recognize object in water bodies is of great importance in presents days, as it will greatly reduces cost for companies trying to help clean the ocean, or coastguard to save people lives, among many others use case scenarios. The study proposes the implementation of an perception pipeline for an USV based on a Deep Learning models, that will allow the system to identify obstacles and react accordingly, as well as keep track of the objects. First, an analysis on several object detec- tion models was made. After this, the deep learning models to be trained (if necessary) needed to be chosen. For this, tests were made in order to assert which framework yielded better results and, after choosing the best model, some modifications were made in order to guaranteed optimal results. The same was done for the tracker that will be used in the system. Then, once the model was chosen, the integration of the model and tracker into the robotic system of the USV needed to be implemented by using a perception pipeline, capable of communicating with other modules of the system. |
|---|---|
| Autores principais: | Genovese, Dani Andres Rodrigues |
| Assunto: | Deep Learning Computer Vision USV Object Detection CNN Tracking |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Deep Learning has been well known in the field of computer vision for its accuracy in detecting objects in images and videos streams; its has been being use for UAVs to detect objects in mid-air, or in self-autonomous cars, which are becoming a reality in the upcoming years. However, there have been a lack of research done with Unmanned Surface Vehicles (USV), used in maritime environments, in comparison with the other two fields mentioned before. These happens due to the harsh environment on the sea and the lack of specialized datasets, making it difficult to train deep learning algorithms that can detect and identify objects with high accuracy. Also, implementing an system with autonomous capabilities also comes with many difficulties regarding autonomy and performance, given the high computing power they need to operate. Nonetheless, the creation of an autonomous vehi- cle that is capable to recognize object in water bodies is of great importance in presents days, as it will greatly reduces cost for companies trying to help clean the ocean, or coastguard to save people lives, among many others use case scenarios. The study proposes the implementation of an perception pipeline for an USV based on a Deep Learning models, that will allow the system to identify obstacles and react accordingly, as well as keep track of the objects. First, an analysis on several object detec- tion models was made. After this, the deep learning models to be trained (if necessary) needed to be chosen. For this, tests were made in order to assert which framework yielded better results and, after choosing the best model, some modifications were made in order to guaranteed optimal results. The same was done for the tracker that will be used in the system. Then, once the model was chosen, the integration of the model and tracker into the robotic system of the USV needed to be implemented by using a perception pipeline, capable of communicating with other modules of the system. |
|---|