My PhD
Title:
Coastal Navigation in Autonomous Vessels.
External Stakeholder:
UK Hydrographic Office (UKHO)
The Research:
The project will look to utilize image data in training a model for location recognition purposes about a coastline. The project will look to answer the questions of whether sensor location can be reliably retrieved given a dataset of observed points and camera locations. The project will also look to answer the question of whether such an observation can be distilled into key informative points, in order to reduce the impact of noisy imaging and to further generalize the outputs for a user. Active learning methods will also be explored to identify the benefit of user interaction for problems that involve explaining image-based information.
The project will first implement a system for retrieving the top likely locations of a vessel, given an observed sample and a model trained on a dataset containing observations of the coastline. The project will first utilize observed images to train a model to perform feature extraction, embedding the observations into a space that can be queried. A subsequent component of the pipeline will then perform retrieval of the possible positive locations given a query sample. The project will then utilize embedding models to generalize the observations and produce representations of the observed data which is more intuitive to humans, specifically via the use of image based saliency visualizations.
The project novelty comes from the development of models which combine information retrieval with active learning. Further novelty also comes in the development of models which are able to leverage local and global information to provide users with a simplified representation of their observation and why the model has provided a given prediction.