| Author: | Xu, Penghui |
| Title: | Learning-based approaches for global navigation satellite system positioning in urban area |
| Advisors: | Hsu, Li-ta (AAE) Yang, Bo (COMP) |
| Degree: | Ph.D. |
| Year: | 2025 |
| Department: | Department of Aeronautical and Aviation Engineering |
| Pages: | xvii, 140 pages : color illustrations |
| Language: | English |
| Abstract: | GNSS plays an important role in global localization for robotics and autonomous vehicles. However, the performance of the GNSS in urban suffers from non-light-of-sight (NLOS) and multipath reception due to the reflection. This thesis is systematically organized to explore advancements in GNSS positioning through deep learning (DL). Initially, PositionNet, a DL-based direct positioning method is developed, focusing on how neural network (NN) can be trained to directly estimate positions from raw GNSS data, thereby bypassing traditional approaches. Relying on the proposed single-differenced residual map (SDRes Map), PositionNet can greatly mitigate the positioning error due to NLOS/multipath, achieving a 7-meter level of accuracy in the deep urban of HK with 90% of the epochs. The mechanism of PositionNet is also invested, showing that the neural network is capable of assigning different importance to different satellites based on the SDRes Map consistency and their auxiliary input, such as signal strength. With the graphical properties of the SDRes Map, four main focuses of the NLOS/multipath problems, including positioning, signal status prediction, measurement weighting estimation, and NLOS/multipath error calculation, are able to be tackled simultaneously and demonstrated to have superior performance against the conventional or even state-of-the-art method methods. To enhance the robustness of the DL-based method, the differentiable estimation method is explored, which integrates end-to-end learning with estimation techniques, endowing a better modeling ability to the robust estimation framework. Currently, the prevalent variance models for GNSS state estimation are based on various statistical models, such as the C/N0-based Sigma-ε model. However, the performance of these variance models could be suboptimal in the presence of unknown errors and noises. In this thesis, a differentiable FGO (DFGO) for intelligent weighting adaptation for state estimation is proposed. Adaptive weightings are obtained via the NN, which is trained with ground truth location via backpropagation. The weighting from the DFGO outperforms the weighting calculated from the measurement uncertainties by 41.8%, which shows the feasibility and robustness of the DFGO approach. Further, AutoW, the DFGO in a self-supervised manner is also investigated, aiming to reduce the need for the labeled data to enhance the generalization ability. The new loss functions are designed to couple with the DFGO training pipeline based on two priors of the static experiment: clustering and zero-velocity. The performance of AutoW is evaluated using massive data collected with u-blox F9P and OPPO Find X6, which demonstrates superior performance against the conventional statistical models. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13752