Over the last decade, advancements in Unmanned Aerial Vehicle (UAV) technology and Artificial Intelligence (AI) have offered promising solutions for enhancing navigation and positioning capabilities under challenging conditions. Despite significant progress, widespread adoption in the field has been limited due to the complexities of integrating various technologies and ensuring real-time, reliable data processing. This thesis addresses these challenges by developing a comprehensive framework that combines advanced data collection platforms, deep learning algorithms, and novel fusion methods to improve UAV positioning accuracy and reliability.

The core of this research is the development of the Quality Driven Drone (Q-Drone) UltraWideband (UWB) benchmark dataset, which includes data collected from a UAV equipped with five UWB sensors across five diverse sites, covering indoor, outdoor, and semi-outdoor environments over a distance of 4 km. This dataset provides a standardized benchmark for testing UAV positioning systems, allowing researchers to develop and validate their algorithms under varied conditions. The publicly available dataset supports advancements in UAV navigation and positioning research.

Building on this foundation, the thesis introduces an incremental smoothing approach that integrates high-rate and low-rate UWB measurements with inertial data within a unified pose graph framework. This method, featuring an "add-after-eliminating" strategy, reduces Mean Absolute Error (MAE) by approximately 0.2 meters compared to baseline multilateration methods and achieves a 0.3-meter MAE reduction compared to two-factor pose graph methods.

The research further advances positioning accuracy with the development of the DeepLearning-based Dynamic Covariance Prediction in Pose Graphs (DeepCovPG) framework, which combines a Variational Autoencoder (VAE) with a Long Short-Term Memory (LSTM) network to predict and incorporate dynamic covariances into the pose graph. This deep learning-based approach results in a 48% reduction in The Root Mean Squared Error (RMSE) and a 51% reduction in Range Covariance RMSE, with significant improvements of 0.41 meters in tunnels and 0.23 meters in fields. Additionally, the framework achieves a 26% reduction in multilateration RMSE and a 32% reduction in multilateration Covariance RMSE, demonstrating substantial improvements over baseline methods.

The thesis also explores Light Detection and Ranging (LiDAR)-based positioning and mapping and proposes the Inferred Attention Fusion (INAF) fusion method, which dynamically selects relevant information from geometric and AI-based odometry techniques. This method excels in managing large angle rotations and motion distortions, achieving a 3.90% improvement over direct fusion methods and a 0.25% improvement over attention-based fusion methods. The INAF fusion method’s adaptability to various driving conditions enhances accuracy in both straight and dynamic environments, showcasing the effectiveness of integrating LiDAR technology with advanced fusion techniques.

The research conducted in this thesis addresses several scientific challenges, each aimed at improving navigation and positioning systems. By developing a comprehensive dataset, refining multilateration techniques, and integrating advanced deep-learning models, this work makes substantial contributions to UAV positioning. Additionally, the novel INAF fusion method enhances vehicular positioning using LiDAR and AI-based odometry techniques.