| Author: | Fang, Kaiduo |
| Title: | Multi-sensor fusion LiDAR SLAM for advanced robotics |
| Advisors: | Ho, Ivan (EEE) |
| Degree: | Ph.D. |
| Year: | 2026 |
| Department: | Department of Electrical and Electronic Engineering |
| Pages: | xv, 112 pages : color illustrations |
| Language: | English |
| Abstract: | LiDAR Odometry (LO) and LiDAR-Inertial Odometry (LIO) are crucial for autonomous robotics but face challenges in accuracy (especially in dynamic or feature-scarce environments), computational efficiency for low-cost platforms, generalization across LiDAR types, and inconsistency in filter-based methods. While optimization-based LIO offers accuracy, it is computationally expensive. Extended Kalman Filter (EKF)-based LIO is more efficient but suffers from inconsistent estimation due to linearization. Invariant EKF (InEKF) addresses inconsistency through Lie group theory but remains under-explored for LIO. This thesis presents novel frameworks addressing these gaps. First, we propose Inc-DLOM, a generalized, efficient LO framework using Voxel-to-Voxel (V2V) direct registration. It optimizes Generalized ICP (GICP) by downsampling point clouds into voxels storing centroids and covariances. This V2V formulation significantly improves computational efficiency while retaining essential information. An Incremental Voxel Mapping (IVM) method, based on spatial hashing, enables O(1) neighbor search and efficient map updates with single-scan insertion. Inc-DLOM flexibly adapts to point cloud data from various LiDAR architectures. Our evaluation primarily focuses on both mechanical rotating and solid-state LiDARs. Building on Inc-DLOM, we introduce Invariant-DLIO, a tightly coupled LIO system resolving EKF inconsistency using Lie group-based InEKF. The state (rotation, velocity, translation) is defined on the SE₂(3) manifold, extended with IMU biases (SE₂(3) × R6). This structure yields approximately log-linear, trajectory-independent error dynamics, enhancing consistency. The correction step leverages direct registration in an iterated optimization scheme. Combined with IVM, this forms a lightweight, fast, and robust LIO system. Both frameworks were extensively validated on public datasets and real-world robotic platforms across diverse environments, scenarios, and motion dynamics. Inc-DLOM demonstrated superior accuracy and computational efficiency compared to state-of-the-art open-source LO methods (e.g., HDL-SLAM, KISS-ICP). Invariant-DLIO outperformed leading open-source LIO methods in both accuracy and efficiency, confirming the benefits of InEKF for consistent estimation. Finally, recognizing deployment challenges on resource-constrained hardware (e.g., Jetson Orin Nano), we explored CUDA-accelerated deployment of Inc-DLOM for autonomous driving digital twin applications, paving the way for real-time performance on low-cost platforms. This work delivers significant advancements in LO/LIO through: 1) A general, efficient V2V direct registration LO (Inc-DLOM); 2) A consistent, tightly-coupled InEKF-based LIO (Invariant-DLIO); 3) Comprehensive validation showing state-of-the-art performance; 4) Initial exploration of GPU acceleration for broader deployment. |
| 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/200/14529