Versatile and accurate LiDAR-inertial SLAM with efficient LiDAR bundle adjustment

Abstract

Simultaneous localization and mapping (SLAM) play a prominent role in mobile robots by providing state feedback and environment perception for navigation. Compared with visual SLAM, LiDAR (Light Detection and Ranging) SLAM has the ability to accurately build 3D dense point cloud maps and is insensitive to illumination and fast motion.

This thesis focuses on addressing three key aspects of LiDAR SLAM: the design of a low-cost and compact LiDAR, the bundle adjustment of LiDAR point clouds, and the development of a comprehensive LiDAR-inertial SLAM system. The first issue tackled in this thesis involves proposing a novel type of LiDAR design. Conventional mechanical LiDARs are expensive and bulky, hindering their widespread applications in robotics. Conversely, current solid-state LiDARs suffer from a limited detection range. To overcome these problems, we propose a novel low-cost robotic LiDAR based on incommensurable scanning. It offers the advantages of solid-state LiDARs while achieving comparable ranging distances to mechanical LiDARs. Several applications of this LiDAR are demonstrated to validate its advantages.

The second challenge addressed is the multi-view registration, also known as bundle adjustment (BA), for LiDAR point clouds. Existing point cloud registration methods typically work only for pair-wise registration, leading to cumulative drift in LiDAR odometry and restricting the efficiency of multiple LiDAR agent tasks. Drawing upon the concept of BA in visual SLAM, the thesis formulates an efficient and consistent BA for LiDAR point clouds in two steps. The first step is to formulate the BA problem as minimizing the distance from each point to its corresponding feature (plane or edge) and prove the feature parameters can be solved analytically in closed form. The analytical derivatives of the cost function, up to second order, are derived to expedite the optimization. An adaptive voxelization method is further introduced for points association. The second step is to propose a novel compact data structure, point cluster, to avoid enumerating each point at the same pose within a feature in LiDAR BA. The solver also exploits the second-order information to estimate the pose uncertainty stemming from measurement noises, leading to consistent estimates of LiDAR poses. Benchmark evaluations and several extended applications are conducted to demonstrate the efficiency and effectiveness of the proposed approach.

This thesis lastly presents Voxel-SLAM, a complete, accurate, and versatile LiDAR-inertial SLAM system. Current LiDAR-inertial SLAM systems generally comprise scan-to-map odometry and loop closure, lacking some essential functions commonly found in visual-inertial SLAM. Voxel-SLAM is proposed to fill this gap, including initialization, odometry, local mapping, loop closure, and global mapping, all employing a unified voxel map structure. The initialization is fast and robust, enabling the system to start with a motional initial state. The odometry estimates current states and detects potential divergence. The local mapping employs local BA to refine the map and states within a sliding window. The loop closure can detect the previously visited places across multiple sessions and optimize the pose graph accordingly. The global mapping culminates in a global BA optimization to ensure map consistency across the entire environment.