Advanced visual odometry and depth estimation techniques for unmanned aerial systems (UAS) in U-Space environments

Abstract

This thesis explores advanced techniques in visual odometry (VO) and depth estimation for Unmanned Aerial Systems (UAS), specifically within the context of U-Space environments. U-Space, as a European initiative, aims to ensure the safe, efficient, and secure integration of UAS into airspace. This work contributes to this goal by addressing two critical aspects of UAS navigation: precise visual odometry and reliable depth estimation. Chapter 1 - Introduction: The introduction presents the context of U-Space, outlining its evolution and the services it offers, with a focus on emergency management. The challenges of autonomous contingency planning in UAS operations are highlighted, particularly in relation to visual odometry and depth estimation. Chapter 2 - Canonical Trinocular Feature-Based Visual Odometry: This chapter proposes a novel trinocular camera configuration to enhance VO for UAS. The research compares two trinocular setups—inline and 45-degree—with traditional binocular setups, testing them in various scenarios (horizontal, vertical, hybrid, and long). The results demonstrate that the 45-degree trinocular configuration with a standard lens offers significant improvements in both accuracy and computational efficiency, reducing the computational effort to 40\% of that required by binocular systems while delivering more accurate results. However, when a fisheye lens is used, the benefits are less pronounced, particularly in vertical and long scenarios. Chapter 3 - Fine-Tuning Monocular Depth-Estimator Artificial Neural Networks Trained on Synthetic RGB-D Datasets for Real Scenes: This chapter addresses the challenge of depth estimation for UAS using monocular cameras, which are cost-effective but typically less reliable than stereo cameras. The research investigates the effectiveness of fine-tuning deep-learning models trained on synthetic data with small real-world datasets. The results show that complete fine-tuning of all model parameters, as opposed to just the decoder, yields the best performance, especially when the available real data is limited to less than 12.5\% of the data required for optimal model performance. This finding is crucial for applications where only limited real-world data is available. Conclusion: The thesis concludes that the proposed trinocular VO configuration significantly enhances the accuracy and efficiency of UAS navigation, particularly in complex U-Space environments. Additionally, it establishes the importance of fine-tuning depth estimation models with real-world data, even when such data is scarce, to improve the reliability of UAS in operational scenarios. These advancements contribute to the broader goal of integrating UAS into airspace, ensuring they can operate safely and effectively under various conditions.