Please use this identifier to cite or link to this item: http://hdl.handle.net/1946/42024
Maritime surveillance and situational awareness are of topical interest but require massive amounts of human labour if done manually in a traditional setting. The human factor can make this process also highly expensive and unproductive.
Computer vision techniques, through maritime object detection and tracking, can transform maritime surveillance and improve maritime navigation. Classic object detection algorithms trained on general-purpose datasets do not yield satisfactory results for maritime vessel detection, one reason being that ships constitute only a small fraction of objects in these datasets. In this context, availability of domain-specific datasets is crucial, hence a few maritime detection datasets have been published, for example, ABOships, Seaships, Singapore Maritime Dataset, MCShips and MARVEL.
Various object detection algorithms (one- and two-stage detectors) have been employed for ship detection with varying performance. More recently, a new class of detectors, key-point detectors, emerged, yielding promising results on some of the maritime detection datasets mentioned above. This thesis investigates the use of transfer learning techniques on a CenterNet implementation (with various feature extractors), an evaluation of their performance is done on a locally collected maritime vessel dataset, ABOships. Results are evaluated under several augmentation techniques, using Average Precision (AP) and Intersection over Union (IoU) metrics.
|msc_jesper_winsten.pdf||6.89 MB||Open||Complete Text||View/Open|