Loading...
Multi-sensor fusion and SLAM system for UAV navigation in GPS denied environments
Date
2025
Abstract
The primary objective of this dissertation is to develop a robust, multi-sensor fusion and vision based state estimation framework for autonomous Unmanned Aerial Vehicles (UAVs) to operate reliably in complex, dynamic, and visually degraded environments. In recent years, the increasing demand for autonomous UAVs in real world applications such as infrastructure inspection, search and rescue, and environmental monitoring has highlighted the limitations of conventional visual-inertial Simultaneous Localization and Mapping (VI-SLAM) systems, especially in outdoor, unstructured, and GPS degraded environments. This thesis addresses these challenges by integrating tightly coupled multi-sensor fusion, deep learning driven adaptive sensor weighting, and semantic feature masking into a unified localization and mapping framework, significantly improving both robustness and accuracy in diverse operational scenarios.
The first major contribution of this work introduces LGVINS, a LiDAR-GPS-Visual-Inertial navigation system that tightly couples multi-sensor measurements within a unified optimization framework, enabling drift free and globally consistent state estimation. By directly incorporating GPS, LiDAR, and visual-inertial constraints into the system’s factor graph, LGVINS achieves superior accuracy in environments where traditional GPS or visual odometry alone would fail, such as urban canyons, bridges, and forested areas.
The second contribution, LSAF-LSTM, proposes a Learning based Self-Adaptive Fusion mechanism where a Long Short-Term Memory (LSTM) network dynamically adjusts the sensor fusion weights based on environmental conditions, feature quality, and sensor reliability. This data driven adaptive weighting allows the system to intelligently down weight unreliable sensors during periods of degraded visibility, low texture, or rapid motion, ensuring optimal estimation robustness in rapidly changing environments.
The third contribution, Mask-vSLAM, integrates real-time semantic segmentation using YOLOv11 into the feature based SLAM pipeline to selectively mask out unreliable and non informative features, such as sky, water surfaces, and transient dynamic objects e.g., pedestrians. This targeted feature filtering ensures only stable and reliable features are used for pose estimation and mapping, reducing drift and improving long term localization accuracy.
The proposed framework is comprehensively validated through extensive experiments on both public benchmark datasets and real world UAV field experiments datasets collected in highly challenging environments, including UL Car Bridge, ULLiving Bridge, and UL Car Parking datasets. Comparative analysis against state-of-the-art SLAM systems, demonstrates the consistent superiority of the proposed approach across all performance metrics. By combining multi-sensor fusion, deep learning adaptability, and semantic feature awareness, this dissertation delivers a robust, adaptive, and context aware localization framework that advances the state of the art in UAV autonomy and resilient state estimation.
This Ph.D. thesis contributes novel methodologies that advance the field of resilient localization and mapping for autonomous UAVs by bridging the gap between classical multi-sensor fusion, deep learning based adaptability, and semantic scene understanding. The proposed research enhances not only the accuracy of UAV state estimation, but also its robustness and adaptability, thereby laying the foundation for reliable and autonomous UAV operations in complex and uncertain environments.
Supervisor
Dooly, Gerard
Trslic, Petar
Trslic, Petar
Description
Publisher
University of Limerick
Citation
Collections
Files
ULRR Identifiers
Funding code
Funding Information
Sustainable Development Goals
External Link
License
Attribution-NonCommercial-ShareAlike 4.0 International