Efficient and accurate LiDAR mapping : from extrinsic calibration to global optimization

Abstract

Three-dimensional (3D) maps play a vital role in our daily lives. They can assist urban planning and smart transportation, facilitate the preservation of historic buildings, and provide virtual tourism with augmented and virtual reality technologies. Among several 3D mapping methods, Simultaneous Locomotion and Mapping (SLAM) based approaches stand out due to their cost-efficiency, easy deployment, and compatible accuracy with sensors such as Light Detection and Ranging (LiDAR), camera, and Inertial Measurement Unit (IMU). This thesis focuses on efficiently producing highly accurate LiDAR point cloud maps from two perspectives: extrinsic calibration and global optimization.

The first problem addressed in this thesis is the targetless extrinsic calibration between multiple LiDARs with few or even no field-of-view (FoV) overlap. Existing methods either rely on external calibration targets or assume FoV overlap between LiDARs, which are not always feasible due to dedicated mounting positions. To address this issue, this thesis proposed a motion and appearance-based method, which creates co-visible features through rotational movement and extracts LiDAR plane and edge features for correspondence matching. This method is targetless and concurrently optimizes the LiDAR odometry and extrinsic parameters, meanwhile producing an accurate environmental point cloud map.

The second problem addressed in this thesis is a faster and more accurate extrinsic calibration pipeline for multiple LiDARs and cameras with minimal or no FoV overlap. To address this issue, this thesis formulated the LiDAR extrinsic calibration problem into a LiDAR bundle adjustment (BA) problem and derived the cost function to second order. This thesis also implemented an adaptive voxel map to extract LiDAR plane and edge features for LiDAR-LiDAR and LiDAR-camera correspondence matching. Both approaches have boosted the computation speed to an order of magnitude faster than existing baselines yet achieved a more precise colorized point cloud mapping result.

The third problem addressed in this thesis is point cloud global consistency optimization. To address this issue, this thesis proposed a hierarchical LiDAR BA framework that utilizes LiDAR BA to directly optimize the point cloud consistency and the PGO to update the LiDAR pose efficiently. The hierarchical structure is highly efficient as it divides the original BA problem into multiple sub-problems and is ideal for parallel computation. This proposed hierarchical framework is tested under various indoor and outdoor LiDAR datasets with structured and unstructured environments where it stably produces globally consistent point cloud maps.

The last problem addressed in this thesis is the integration of IMU measurements with the hierarchical BA framework and the joint optimization of multiple LiDAR point cloud maps. To address this issue, this thesis fully utilized the IMU measurements for point cloud undistortion and constructed a pre-integration factor with global factor graph optimization (FGO). To avoid redundant calculation, this thesis proposed an IMU pre-integration merging theory. To jointly optimize multiple point cloud maps, the extrinsic parameters are used to align the trajectories into the same coordinate frame. This augmented framework consumes slightly more time in processing IMU measurements but achieves optimal performance and produces a more complete and consistent point cloud map.