Vision Based Object Identification and Detection in Maritime Environments

December 6, 2021

The work developed in this thesis will have as its main focus applying deep learning object detection algorithms to an autonomous vehicle performing unmanned missions. It aims at providing said vehicle with the ability to detect and track different kinds of obstacles in real-time using only a conventional RGB camera. We will explore the general features of object detection algorithms and more specifically the chosen neural network for the task, Scaled-YOLOv4, currently a state-of-the-art object detection network. The workflow consists of researching appropriate datasets representing the environment experienced by the vehicle, converting the trained network to an embedded system and finally applying visual tracking to the detection targets. The explored datasets include the SeaShips and ABOships datasets, both presenting thousands of images for proper neural network training. The trained neural network will run on a Jetson Nano 2GB module, a single board computer designed with artificial intelligence applications in mind, ideal for implementation in embedded systems like CEiiA's ASV. To complement object detection we add a tracking feature, allowing the algorithm to track objects resulting in a sense of object permanence that can be further extended upon in future works. Tracking occurs through the use of reliable solutions such as the SORT algorithm and a custom made particle filter. This thesis is done in collaboration with CEiiA and aims at combining state-of-the-art artificial intelligence with autonomous systems with hopes to provide an easy and scalable solution for object detection and tracking in maritime environments.

Authors in the community:

• 

  D

  Duarte Filipe Cotrim Nunes

  ist186624

Supervisors of this institution:

• 

  B

  Bruno Duarte Damas

  ist45387

mechanical-engineering - Mechanical engineering

eng - English

Instituto Superior Técnico