An Encyclopaedia-based Computer Vision Framework with Application to Visual Geolocalisation

Abstract

The popularity of drones (UAVs -- Unmanned Aerial Vehicles) seems to have no bounds with multiple applications including parcel delivery, aerial photography, farming, search and rescue, defence, and so on. Accurate positioning of the drone is key to many, if not all, of these applications. However, there are many reasons for drones being unable to obtain an accurate location. These include multi-path reflections in an urban canyon affecting a delivery drone, or spoofing and blocking of satellite signals in defence applications. For these reasons, a large amount of research is being performed into visual self-localisation and the related topic of visual object-localisation from a drone. Research by others into visual localisation concentrates on particular environments, using coastlines, rivers, roads, urban layouts, and so on. Although there have been successes in these areas, there is not a one-size-fits-all solution. It is now generally accepted that a combination of algorithms is required, with the accumulation of knowledge about the environment aiding the algorithms in their operation. However, some of these algorithms require special or powerful hardware and the increasing prevalence of drone-swarms adds to the processing requirements. Additionally, the fact that most drones are in the air for some video-monitoring purpose, the complexity and distributed nature of computer vision systems can become an issue. This thesis proposes a novel Encyclopaedia-based Framework (EbF) for distributed computer vision systems. Building on the concepts of the popular Robot Operating System (ROS), this framework encourages an open, modular and flexible approach to the capture and incorporation of domain knowledge for an application. A key benefit of the framework, shown within the chapters of the thesis, is that it opens up a system to its owner which frees them from the ties of the original developers. This enables the owner, users and other developers to easily add and modify functionality. This includes by adding new hardware infrastructure, to keep a system responding to changing requirements. It also increases the overall return on investment and the value and usefulness of a system to its users.

Date of Award	6 Sept 2022
Original language	English
Awarding Institution	The University of Manchester
Supervisor	David Morris (Supervisor) & Martin Turner (Supervisor)