Oskar Branzell presenterar sitt examensarbete "Video Object Detection for Maritime Navigation"

Svensk Titel: Objektdetektion i video för navigation till sjöss

Handledare: Magnus Oskarsson och Mårten Lager

Examinator: Mikael Nilsson

Abstract:

The purpose of this thesis was to investigate whether video object detection methods can outperform single-frame object detection methods in the context of maritime navigation. Object detection is a well-explored area within machine learning with many high performing published methods. Video object detection however, where information from previous video frames is considered in order to capture historical data such as movement in the detections, is a relatively recent area.

In this thesis I have researched current state-of-the-art within video object detection and applied these methods to the data provided by Saab Kockums. The video object detection methods can successfully detect navigation objects in video, even small and far away objects that single-frame object detection methods have difficulties finding. I have tested the methods for different settings, backbones and datasets, and evaluated performance using both an annotated test set as well as visual comparisons in video.

The final result of the tests show that video object detection is a very promising area for maritime navigation, even though further testing is necessary. Tests show that video object detection methods can detect objects further away, with higher accuracy, and are less sensitive to the quality of the dataset than single-frame methods, though at a much slower speed. The tests also show that video object detection can additionally be very useful for post-analysis of video data.

Keywords: Video Object Detection, Object Detection, Maritime Navigation, Machine Learning