Continuous and ultra-compact LiDAR mapping and localisation

Abstract

Recently, Light Detection and Ranging (LiDAR) has gained prominence in robotics and autonomous driving for capturing precise environmental geometry, essential for simultaneous localisation and mapping (SLAM) and navigation. However, storing and updating large-scale high-definition maps presents significant challenges, making the development of memory-efficient, updatable, and localisable map representations crucial for advancing SLAM systems in dynamic environments. In this thesis, we present a series of advancements in LiDAR-based SLAM systems aimed at achieving memory-efficient representation, continuous reconstruction, and global mapping. The term continuous reconstruction refers to a continuous function that allows sampling 3D points at any resolution. We firstly introduce CURL (Continuous, Ultra-compact Representation for LiDAR), which leverages spherical harmonics (SPHARM) basis functions to encode point clouds, achieving effective compression and continuous reconstruction while outperforming contemporary deep learning methods. Building on this foundation, we propose CURL-MAP, an extension of CURL for mapping with pose estimation capabilities, utilising a set of bounding boxes containing SPHARM-encoded patches to construct a global map. To build a globally consistent map, we develop CURL-SLAM, which contains a customised CURL-based bundle adjustment (BA) with pose graph to ensure global consistency, even in large-scale scenarios and further increase the efficiency of the system by replacing quasi-conformal mapping with a mask-based method for identifying valid regions. These modifications reduce computational requirements and storage space, facilitating the integration of all newly observed patches and enhancing system robustness. Overall, CURL, CURL-MAP, and CURL-SLAM collectively offer a memory-efficient, updatable, and localisable 3D dense map representation, supporting continuous reconstruction for robust SLAM applications.