Robust LiDAR-based SLAM with real-time loop closure and adaptive mapping in complex environments

Abstract

LiDAR-based Simultaneous Localization and Mapping (SLAM) leverages advanced light detection technology to deliver highly accurate, real-time 3D mapping crucial for autonomous navigation. This technology is particularly effective in guiding robots and vehicles through complex urban environments and is integral to the development of safety features in autonomous vehicles. LiDAR's robustness to variations in lighting and environmental appearance allows it to operate effectively across diverse conditions, making it well-suited for a wide range of environments. Additionally, LiDAR SLAM plays a pivotal role in 3D reconstruction, aiding in architectural planning and the digital preservation of historical sites by creating precise and detailed models. The precision and adaptability of LiDAR-based SLAM make it an essential tool in advancing autonomous technologies and enhancing interactive robotics across various sectors. This thesis specifically focuses on adaptive mapping techniques and loop closure detection, which are critical for maintaining consistent accuracy and reducing cumulative errors in complex environments.

The initial problem addressed in this thesis is improving LiDAR-Odometry mapping for detailed and accurate environmental modeling. Recognizing that LiDAR scanning involves a coarse-to-fine process, we introduced a VoxelMap approach that adapts to scene dimensions and accounts for LiDAR measurement noise. Building on this foundational mapping strategy, we developed ImMesh, an innovative framework designed for online adaptive meshing of the surrounding environment. This system dynamically adjusts mesh construction in real-time, enhancing the resolution and fidelity of the 3D models according to environmental complexity.

In the latter sections of this thesis, we tackle the critical issue of loop closure detection in SLAM, which is essential for ensuring long-term map consistency and accuracy. We present the Stable Triangle Descriptor (STD), a global descriptor that leverages the geometric stability of triangles invariant under rigid transformations to enhance place recognition. While STD provides strong global consistency and robust matching across large-scale environments, it falls short in capturing detailed local point cloud information. To address this limitation, we developed the Binary and Triangle Combined (BTC) descriptor. BTC augments STD by integrating a binary descriptor that captures local point distributions, resulting in a more detailed and comprehensive descriptor. This combination significantly improves the discriminative power and matching accuracy, particularly in complex environments where local details are crucial for reliable loop closure detection. By merging global and local descriptors, BTC ensures the SLAM system remains reliable over extended operational periods and across diverse environments, effectively preventing drift and preserving the integrity of the environmental model.

A key direction for future work involves integrating additional sensors, particularly cameras, into the SLAM system. Incorporating camera data will enable the generation of texture-mapped meshes, enriching the visual quality of 3D models. Moreover, combining camera information with the BTC descriptor can overcome challenges in degenerate or structurally similar environments. To this end, we have developed a novel camera-LiDAR extrinsic calibration algorithm in the final chapter, ensuring precise sensor alignment and enabling the fusion of depth and visual information for more complex applications.