Deep learning for autonomous drone vision
File(s)
Author(s)
Kolbeinsson, Benedikt
Type
Thesis
Abstract
Autonomous drones have the potential to reshape numerous industries with the role of advanced drone vision being central to achieving operational autonomy. This thesis marks a significant advancement in autonomous drone vision, tackling key challenges such as data collection, thin structure detection and semantic segmentation. In response to the pressing need for comprehensive data in this domain, the Drone Depth and Obstacle Segmentation (DDOS) dataset is introduced, specifically designed for drone vision. Using this dataset, a state-of-the-art monocular wire segmentation and depth estimation model is developed to address the challenge of detecting thin structures, which is crucial for the safe flight of autonomous drones. Another major contribution is the development of recursive denoising, a novel diffusion-based approach to semantic segmentation, which greatly improves scene understanding from aerial perspectives. This enables autonomous drones to better interpret their environment, a critical capability for navigating complex scenarios. Together, these developments not only propel drone vision technology forward but also advance the broader disciplines of machine learning and computer vision. They showcase the potential of sophisticated data-driven methods to tackle complex real-world challenges, highlighting the evolving capabilities of autonomous drones in understanding and navigating their surroundings effectively.
Version
Open Access
Date Issued
2024-03-13
Date Awarded
01/02/2025
License URL
Advisor
Mikolajczyk, Krystian
Publisher Department
Department of Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)