| Author: | Hu, Runzhi |
| Title: | Tightly coupled deep learning and GNSS integration for positioning in urban canyons |
| Advisors: | Wen, Weisong (AAE) Hsu, Li-ta (AAE) |
| Degree: | Ph.D. |
| Year: | 2026 |
| Department: | Department of Aeronautical and Aviation Engineering |
| Pages: | xviii, 152 pages : color illustrations |
| Language: | English |
| Abstract: | Global positioning has been fundamental to industrial and civilian applications, as the Global Navigation Satellite System (GNSS) technique, including software and hardware, has rapidly developed in recent years. Unmanned aerial vehicles (UAV), autonomous driving, and robots are coming true, while abundant location-based services (LBS) have played an important role in people's daily lives through mobile phones. Urban areas are the main situation for the above applications. However, compared to open areas under open skies without signal blockage, GNSS signals suffer from obstacles, reflection, and diffraction in urban scenarios. These unexpected impacts are mainly caused by high-rise buildings, which are named urban canyons. Due to the affected signals, GNSS positioning accuracy could be degraded or even unavailable. Deep learning has been a representation of data-driven approaches, as it can numerically model complex processes. In this thesis, we explored using deep learning in the GNSS positioning process, especially in a tightly coupled way, to improve positioning accuracy. We first used a fisheye camera to capture the environment and used the Swin Transformer to segment the sky and identify the satellite, which was obstructed, but its signal was received. This phenomenon is called non-line-of-sight (NLOS). Then, the NLOS flag was contributed as a feature along with pseudorange residual, elevation angle, and carrier-to-noise ratio (C/N0) in a neural network to predict the double differenced pseudorange bias. The results showed improvements in our work. Then, in the continuous work, to directly and deeply capture the environment that causes the NLOS and multipath, we extended our framework to handle images as the direct representation of the environment to input to the neural network. In this work, we concatenated the GNSS features, image features, and positional encoding features to predict the double-differenced pseudorange bias. The ablation experiments proved that all three kinds of features contributed to the accuracy improvement. Meanwhile, we found that we need a method when no reference station is available; thus, in the next work, we considered simultaneously estimating the weight and pseudorange bias in the weighted least squares process of single point positioning (SPP). This work demonstrated that deep learning can be deeply integrated into the positioning process and improve positioning accuracy by data-driven methods. Furthermore, as we want to involve the environment in our model, a semantic high-definition (HD) map construction method was proposed to build a semantic map based on sensor fusion of light detection and ranging (LiDAR), camera, and GNSS. After evaluation, the mean error of our method indoors and outdoors is 0.156 meters and 0.332 meters, respectively. |
| 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/200/14621