Akademik Veri Yönetim Sistemi
12. INTERNATIONAL SUMMIT SCIENTIFIC RESEARCH CONGRESS, Gaziantep, Türkiye, 29 - 31 Mayıs 2024, ss.546-547
In recent years, the utilization of 3D mapping has surged, driven by the increased affordability and enhanced precision of sensors such as 3D Lidars, lasers, and RGB-D cameras. 3D mapping has employed many diverse domains, including geometric information systems, building information modeling, as well as robotics and autonomous vehicles. In robotics, Simultaneous Localization and Mapping (SLAM) methods are generally utilized for the automated generation of 3D maps. The significance of SLAM lies in its capability to produce precise maps of unknown environments while concurrently determining the exact position of the robot within these maps. LiDAR Odometry and Mapping (LOAM) emerged as a pivotal technique leveraging point cloud data derived from a 3D LiDAR sensor to execute real-time SLAM. LOAM follows three steps: Scan registration, laser odometry, and mapping. In the scan registration step, consecutive scans are initially aligned, followed by the extraction of four distinct features. The feature extraction of LOAM mainly depends on the curvature value of points positioned within scan lines. Points with curvature surpassing a predetermined threshold are categorized as sharp or less sharp features based on their curvature. If the curvature of a point is less than the predetermined threshold that point is considered a flat or less flat feature depending on the curvature. The laser odometry step accepts these features from consecutive scans and computes the odometry at a high frequency. Finally, the laser mapping phase utilizes both the odometry and the map, iterating the laser odometry step at a lower frequency. In previous studies, various extensions, and improvements of LOAM, such as Advanced LOAM (A-LOAM), Fast LOAM (FLOAM), and Lightweight and Ground-Optimized LOAM (LeGO LOAM) have been developed to enhance its performance. FLOAM prioritizes accelerating computation while maintaining accuracy by employing a non-iterative two-stage distortion compensation approach. Conversely, LeGO-LOAM is specifically designed for ground-based vehicles, emphasizing computational efficiency and resilience on uneven terrain. While numerous endeavors have aimed to enhance the efficacy of LOAM, few have specifically focused on feature extraction, despite its critical role in LOAM's performance. Previous studies typically relied on the curvature values of points within the same scan line for feature extraction. However, this approach restricts LOAM's applicability to sensors other than capturing point clouds in scan lines with varying heights such as Velodyne VLP-16. In this study, we opted to employ the Intrinsic Shape Signature (ISS) method for feature extraction in LOAM. ISS identifies salient points based on their geometric attributes, ensuring a robust and consistent feature set. Through the analysis of eigenvalues within the covariance matrix surrounding each point, ISS ensures the selection of distinct and reproducible features. Our objective is to enhance LOAM's performance and broaden its applicability by integrating the ISS approach for feature extraction. Unlike previous methods, we utilized the entire point cloud rather than limiting feature extraction to points within the same scan line. This approach enables the utilization of point clouds acquired from both RGB-D cameras and 3D lasers with LOAM. Subsequently, the laser odometry and mapping stages leverage the features extracted via the ISS method. We conducted experiments on the KITTI Odometry dataset, which is a comprehensive benchmark for autonomous driving research to evaluate the performance of the proposed method. We used Absolute Pose Error (APE) and Relative Pose Error (RPE) metrics, which are commonly used metrics for measuring the performance of SLAM methods. APE measures the overall deviation of the estimated trajectory from the ground truth, while RPE evaluates the local accuracy of the estimated motion. We performed experiments on sequence 7 of the KITTI Odometry dataset. We also performed hyperparameter optimization for the ISS approach by adjusting parameters such as the non-maximal suppression radius and the salient radius to optimize feature extraction. The experimental results indicated that employing ISS for feature extraction significantly reduces both APE and RPE compared to the curvature-based feature extraction method. Specifically, the APE reduced from approximately 4.36m with A-LOAM to 3.35m with the proposed method. On the other hand, the RPE reduced from 0.0034rad with A-LOAM to 0.0025rad with our method.