Abstract/Summary

This thesis proposes new computational methods for detecting maritime objects in video data and analyses their performance in the context of a counter-piracy surveillance system. A key challenge in the maritime anti-piracy context is the wide range of possible environments and objects which may be encountered. The focus of this work is therefore on methods which do not make strong assumptions about the visual appearance of the scene or targets. This has the additional benefit that they can be used in other applications and domains. Two novel approaches are investigated. The first is a saliency-based method which uses a novel thresholding step and a scene depthmap derived from the horizon to emphasise local saliency and incorporate scene context, respectively. The second uses a deep semantic segmentation network to separate ‘sea’, ‘sky’ and ‘other’ regions. Contextual scene knowledge is then used to extract objects by applying rule-based reasoning. Evaluation on publicly available maritime surveillance datasets shows that the proposed methods address limitations of current approaches, particularly with regard to the detection of small, distant targets. The analysis also explores the key aspects of performance required to deploy an algorithm‘ in the field’ as part of a larger system for detection, tracking, situation awareness and threat detection. As well as the trade-off between real-time operation and performance, results on data collected from a real-world surveillance system show that the relationship between detection scores in images and tracking performance in the real-world is not trivial. Finally, contributions are made in the benchmarking and evaluation of image-based maritime object detection methods, including a novel dataset for counter-piracy surveillance, improvements to metrics for performance evaluation in the maritime domain, and comparison of the proposed approaches with state of the art object detection methods from other domains.