guest ::
TY - THES AU - Zhang, Haoming TI - Handling GNSS-outliers in multisensor state estimation PB - Rheinisch-Westfälische Technische Hochschule Aachen VL - Dissertation CY - Aachen M1 - RWTH-2026-00971 SP - 1 Online-Ressource : Illustrationen PY - 2025 N1 - Veröffentlicht auf dem Publikationsserver der RWTH Aachen University 2026 N1 - Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2025 AB - Vehicle localization using classic state estimation approaches generally conducts least-square techniques to infer unknown state variables by fusing nonlinear system dynamics and sensor models. In this setting, model uncertainties are typically assumed to follow a zero-mean Gaussian distribution. An optimal state estimate can be achieved by minimizing covariances using redundant state representations (aka models), provided the predefined noise parameters match the actual uncertainties. However, the assumption of Gaussian noise has been shown to be overly conservative and unrealistic in real-world environments, where sensor failures and signal interference frequently result in measurement outliers. In this case, measurement outliers typically exhibit time-varying noise dynamics associated with heavy-tailed, skewed, and multimodal distributions that cannot be adequately modeled in advance. Consequently, a state estimator employing the least-square technique may degrade or even completely diverge once an inconsistent measurement model with overly confident hand-crafted noise parameters is used. Therefore, developing a vehicle localization approach that achieves robust long-term operation in diverse environments remains a challenging problem. This dissertation examines methods for handling outliers in a graph-optimization-based state estimator, specifically designed for vehicle localization using Global Navigation Satellite Systems (GNSS). In addition to developing novel state estimation techniques, this thesis places noticeable emphasis on two key areas: (a) a theoretical analysis of the state-of-the-art robust outlier rejection approaches, and (b) learning-based methods, supported by extensive ablation studies. In this regard, I explore two methods for handling GNSS outliers. The first considers the classic M-estimation techniques, which are commonly used to reduce the impact of outliers in state estimation and have proven effective in practice. However, their application to ranging sensors, particularly in terms of robustness and efficiency across different loss kernels, has received little attention. The second method leverages learning-based techniques to predict outlier measurements. Beyond traditional machine learning models, such as Support Vector Machines, I introduce a novel deep network architecture that performs spatiotemporal modeling, specifically considering the characteristics of GNSS noise propagation. In summary, this dissertation advances the understanding of outlier handling using M-estimation and learning-based methods to enhance the robustness of vehicle localization for long-term operation in challenging environments. By emphasizing these approaches, the thesis aims to provide critical insights and inspire future advancements in robust state estimation for vehicle localization. Additionally, I prioritize reproducibility and comparability by open-sourcing the code for all methods discussed. LB - PUB:(DE-HGF)11 DO - DOI:10.18154/RWTH-2026-00971 UR - https://publications.rwth-aachen.de/record/1026564 ER -