Simultaneous localization and mapping (SLAM) algorithms have been widely studied for precise pose estimation and mapping in autonomous vehicles. However, most research has focused on small-scale platforms such as ground vehicles and UAVs, leaving a ga...
Simultaneous localization and mapping (SLAM) algorithms have been widely studied for precise pose estimation and mapping in autonomous vehicles. However, most research has focused on small-scale platforms such as ground vehicles and UAVs, leaving a gap in studies on large autonomous vehicles. To address this gap, this paper presents an enhanced LiDAR-based odometry estimation method designed for autonomous buses in urban environments. While LiDAR-inertial odometry (LIO) methods utilizing IMU sensors are commonly used to improve accuracy, IMU dependency is impractical for large vehicles as IMU measurements are easily affected by external factors. Instead, this study leverages vehicle kinematics and wheel speed to estimate velocity and yaw rate, resulting in improved robustness and accuracy compared to traditional LiDAR odometry methods. Following the odometry estimation, this thesis proposes smoothed neural implicit 3D reconstruction for a large-scale LiDAR map against multi-LiDAR calibration errors in autonomous buses. Recent advances in neural implicit representations offer several advantages over traditional explicit mapping approaches by training neural networks to fit scene observations, allowing for efficient querying of various properties at any location within the scene. These memory-efficient implicit neural maps provide rich map information. However, multi-LiDAR systems are essential for large-scale autonomous platforms but introduce extrinsic calibration challenges due to sensor spacing and self-occlusions. Traditional calibration methods relying on overlapping fields of view or rigid mounts are impractical, leading to calibration errors that cause undesirable thickness in the LiDAR point cloud. This degrades neural implicit mapping and mesh reconstruction, yet research on handling calibration-induced offsets in multi-LiDAR systems remains limited. To address the challenges, this thesis proposes a novel method combining offset correction and neural implicit reconstruction. The study presents results based on actual vehicle data collected on urban tracks and simulation dataset. Experimental results demonstrate that the proposed method effectively mitigates pose estimation errors observed in conventional LIO methods. In addition, by integrating an offset correction process with SDF-based neural implicit map learning, our approach mitigates mesh errors caused by multi-LiDAR calibration inaccuracies while improving reconstruction accuracy.