Year: 2023
Type: dissertation
Abstract: This paper presents a method for tackling erroneous odometry estimation results from LiDAR-based simultaneous localization and mapping (SLAM) techniques on complex urban roads. Most SLAM techniques estimate sensor odometry through a comparison between measurements from the current and the previous step. As such, a static environment is generally more advantageous for SLAM systems. However, urban environments contain a significant number of dynamic objects, the point clouds of which can noticeably hinder the performance of SLAM systems. As a countermeasure, this paper proposes a 3D LiDAR SLAM system based on static LiDAR point clouds for use in dynamic outdoor urban environments. The proposed method is primarily composed of two parts, moving object detection and pose estimation through 3D LiDAR SLAM. First, moving objects in the vicinity of the ego-vehicle are detected from a referred algorithm based on a geometric model-free approach (GMFA) and a static obstacle map (STOM). GMFA works in conjunction with STOM to estimate the state of moving objects in real-time. The bounding boxes occupied by these moving objects are utilized to remove points corresponding to dynamic objects in the raw LiDAR point clouds. The remaining static points are applied to LiDAR SLAM. The second part of the proposed method describes odometry estimation through referred LiDAR SLAM, LeGO-LOAM. The LeGO-LOAM, a feature-based LiDAR SLAM framework, converts LiDAR point clouds into range images, from which edge and planar points are extracted as features. The range images are further utilized in a preprocessing stage to improve the computation efficiency of the overall algorithm. Additionally, a 6-DOF transformation is utilized, the model equation of which can be obtained by setting a residual to be the distance between an extracted feature of the current step and the corresponding feature geometry of the previous step. The equation is optimized through the Levenberg-Marquardt method. Furthermore, GMFA and LeGO-LOAM operate in parallel to resolve computational delays associated with GMFA. Actual vehicle tests were conducted on urban roads through a test vehicle equipped with a 32-channel 3D LiDAR and a real-time kinematics GPS (RTK GPS). Validations results have shown the proposed method to significantly decrease estimation errors related to moving feature points while securing target output frequency.
Author: 권우진
Institution: Seoul National University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7061683356
