Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Mechanical Engineering | en_US |
| dc.contributor.advisor | Su, Zhongqing (ME) | en_US |
| dc.creator | Ma, Weixin | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/13745 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | GNSS-free global localization in urban environments | en_US |
| dcterms.abstract | Global localization provides robots with their current positions relative to the environment, enabling them to perform more complex downstream tasks such as path planning and navigation. One of the most widely adopted solutions, the Global Navigation Satellite System (GNSS), has demonstrated significant advantages over the past decades. Its robustness and accuracy may be compromised by weak signals, multipath effects, or signal obstruction, especially in densely built urban areas with tall buildings. We therefore identify a need to advance techniques on global localization in GNSS-denied environments. The goal of this thesis is to develop real-time solutions with high memory efficiency for GNSS-free global localization in urban environments, with a focus on autonomous vehicles. | en_US |
| dcterms.abstract | Our first contribution focuses on global localization in high-rise environments using a publicly available map, i.e., OpenStreetMap. We observe that the building roof outline captured by a sky-looking fish-eye camera shares similarities with the building outline featured in the OpenStreetMap. Based on this observation, we propose a descriptor that incorporates both topological (i.e., junction types) and geometric (i.e., building outline) information to bridge fish-eye camera images with OpenStreetMap data. To handle the challenge of similar or repeated building outlines across different images, we formulate our method as a Bayesian Filtering problem using Monte Carlo localization, which leverages multiple consecutive fish-eye images for robust localization. The sky-looking fish-eye camera also naturally avoids disturbances from dynamic objects such as vehicles and pedestrians, making it particularly suitable for urban environments. | en_US |
| dcterms.abstract | While our first solution enables real-time localization in high-rise environments without the need for manually collecting and maintaining reference maps, its long-term reliability can be impacted by the quality of captured images, which are sensitive to changes in illumination, weather, and seasons. In contrast, range sensors such as LiDAR and radar are more resilient to these environmental factors. To investigate the influence of these factors on range sensing based global localization methods, we conduct a comprehensive evaluation of current techniques in long-term scenarios with significant seasonal variations and adverse weather conditions. In addition, we design a novel metric to evaluate the influence of matching thresholds on place recognition performance for long term localization. | en_US |
| dcterms.abstract | Based on our previous evaluation study, LiDAR-based place recognition shows good robustness under long-term conditions. However, it typically only identifies whether the current place has been visited before. Determining the vehicle's pose relative to the environment requires additional point cloud or feature point based registration, which might demand significant memory to store these data. Our third contribution addresses this challenge by integrating pose estimation with place recognition to improve LiDAR-based global localization performance. Rather than relying on abstract environmental representations like 3D points, we use lightweight semantic features (e.g., traffic signs, trees, poles, and so on) to represent both on-vehicle LiDAR scans and reference scans in the databases, significantly reducing memory requirements. We propose a novel semantic histogram descriptor to represent each semantic instance, which is used for instance association in pose estimation and aggregated into a global descriptor for place recognition. | en_US |
| dcterms.abstract | While our previous contribution show competitive localization accuracy and memory efficiency, the scan-to-scan framework still redundantly stores some of semantic instances multiple times across keyframes. Meanwhile, its RANSAC-based pose estimation might fail when outliers dominate. Our fourth contribution shifts to a scan-to-map localization manner, where each semantic instance is stored only once in the reference map, further improving the memory efficiency. We improve the semantic histogram descriptor in our previous work to achieve more robust and effective instance-to-instance correspondences. In addition, we propose a novel Road Surface Normal (RSN) map to provide a prior rotational constraint, enhancing pose estimation. We then apply graph-theoretic outlier pruning to extract inlier correspondences for robust 6-DoF pose estimation. | en_US |
| dcterms.abstract | Finally, we develop a multi-robot localization system, building upon our previous scan-to-map localization approach. Unlike single-robot localization, multi-robot systems typically lack a prior map, and the initial relative poses between local reference frames of robots are unknown. To this end, we formulate the multi-robot localization as an optimization problem and incorporate one-shot registration-based localization into the optimization process. Specifically, lightweight semantic instances from different robots are transmitted to a central server, which constructs instance maps for inter-robot localization. We propose a dual-metric validation strategy to confirm the validity of pairwise localization results from our previous scan-to-map localization solution to reduce the risk of involving incorrect localization results. A pose averaging based optimization method is used to obtain reliable alignment estimations between local reference frames of different robots. A shortest transformation chain searching method is used to align all robots into a shared reference frame. Our preliminary results demonstrate the feasibility and effectiveness of the proposed system, as well as its promising potential in communication bandwidth-limited conditions. | en_US |
| dcterms.extent | xxi, 165 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2025 | en_US |
| dcterms.educationalLevel | Ph.D. | en_US |
| dcterms.educationalLevel | All Doctorate | en_US |
| dcterms.LCSH | Location-based services | en_US |
| dcterms.LCSH | Automated vehicles | en_US |
| dcterms.LCSH | Remote sensing | en_US |
| dcterms.LCSH | Hong Kong Polytechnic University -- Dissertations | en_US |
| dcterms.accessRights | open access | en_US |
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/200/13745