An application of object detection in ship navigation

Δημητρέλος, Ιωάννης; Dimitrelos, Ioannis

dc.contributor.author	Δημητρέλος, Ιωάννης	el
dc.contributor.author	Dimitrelos, Ioannis	en
dc.date.accessioned	2022-11-10T11:13:47Z
dc.date.available	2022-11-10T11:13:47Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/56119
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.23817
dc.rights	Default License
dc.subject	Safe ship navigation	en
dc.subject	Object Detection	en
dc.subject	Neural Networks	en
dc.subject	Autonomous ships	en
dc.subject	Yolo	en
dc.subject	Ασφαλής πλοήγηση πλοίων	el
dc.subject	Νευρωνικά Δίκτυα	el
dc.subject	Αυτόνομα πλοία	el
dc.subject	Αναγνώριση εμποδίων	el
dc.title	An application of object detection in ship navigation	en
heal.type	bachelorThesis
heal.classification	Ship navigation	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2022-07-13
heal.abstract	The present thesis is an attempt to light a spark in the field of object detection in ship navigation. This introduction is followed by an application based on an object detection model, named You Only Look Once (YOLO). Even though it is not the first time an object detection algorithm was applied to ship navigation problems, it is very fascinating to further research and experiment in this field by firstly testing an already existing algorithm and commenting on the results. Many object detection algorithms have been proposed with approaches ranging from traditional to deep learning. However, the majority of them have limited applications in real-time applications as they are computationally intensive and have accuracy problems. Another challenge when dealing with ship navigation is the wide range of background sizes of the objects. To overcome these problems the most recent object detection algorithm was selected. In this thesis, the You Only Look Once version 5 (YOLOv5) model was used, which is created in 2020 and is fine-tuned with the more recent and best practices in object detection and also is constantly modified and readjusted to achieve better results. From the different models of YOLOv5, the smaller one was picked, because of its size and the comparatively good accuracy it performs. After the model was chosen, a sufficient database for the model’s training was created using images that contained the most common obstacles that a ship can face. These are other ships, buoys, humans on the surface, containers, and rocks. The images were categorized into 3 teams, ship, floating object, and rock, which are the classes of the problem. The images were then labeled in a way that the YOLO accepts as input while some of them were kept and created the validation dataset. With these data some scenarios were created, namely, images that contained only one class in them, images of ship at night, and images of at least 2 classes coexisting in the same image and their combination. The model was trained with these different datasets and the results were collected, compared, and analyzed. The model achieved a Precision of 84%, Recall of 74%, and mAP of 79% on average when trained with 400 images of all the classes combined for 300 epochs and batch size 8. The model was also tested in a real-time application using a video and it detected all of the ships in most cases with 33 frames per second reload time.	en
heal.advisorName	Βεντίκος, Νικόλαος	el
heal.committeeMemberName	Ηλιοπούλου, Ελευθερία	el
heal.committeeMemberName	Γρηγορόπουλος, Γρηγόρης	el
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ναυπηγών Μηχανολόγων Μηχανικών	el
heal.academicPublisherID	ntua
heal.numberOfPages	93 σ.	el
heal.fullTextAvailability	false