LiDAR-based SLAM provides robust and precise 3D structural information regardless of illumination changes; however, it still faces several limitations, including non-uniform point distribution, high computational cost, accumulated odometry drift, and ...
LiDAR-based SLAM provides robust and precise 3D structural information regardless of illumination changes; however, it still faces several limitations, including non-uniform point distribution, high computational cost, accumulated odometry drift, and numerical instability caused by incorrect loop closures. Existing SLAM approaches often rely on simple downsampling to reduce computation or utilize semantic information only as auxiliary cues, resulting in restricted performance improvement.
To address these issues, this study proposes Light3D SLAM, a lightweight LiDAR SLAM framework that integrates semantic information with geometry-driven reliability evaluation. The proposed method employs a PointNet-based semantic segmentation model that directly processes raw point clouds to extract semantic features, and defines a 5D Reliability Score by combining semantic class, geometric structure, height, intensity, and local density, enabling reliable point-wise importance estimation. Based on this score, Semantic-aware Compression selectively retains structurally meaningful points, reducing the overall point cloud size by approximately 46% while minimizing information loss.
High-precision odometry is then obtained using CUDA-accelerated Fast-GICP, and loop closure candidates detected by Scan Context are geometrically validated through GPU-based Multi-Hypothesis ICP (MH-ICP) and a Rotation Filter to remove false positives. Finally, global drift is corrected through Pose Graph Optimization (PGO) using g2o with an Adaptive Information Matrix.
Experiments on the KITTI Odometry benchmark (Sequences 00–10) demonstrate that Light3D SLAM achieves high registration stability in urban environments, maintaining an average Fast-GICP fitness above 0.97. When loop closures are present, the system reduces Z-axis drift by an average of 58.1% and ATE by 21.5%, confirming the effectiveness of global optimization. Compared to LOAM, Light3D SLAM shows significantly improved ATE in urban environments, and the proposed Semantic-aware Compression reduces computation without degrading registration quality. However, in highway environments with limited structural landmarks, loop closure is rarely triggered, highlighting the inherent limitations of LiDAR-only SLAM.
This study presents a practical SLAM framework that mitigates the structural limitations of LiDAR-only systems by combining semantic scene understanding with reliability-based point selection. Light3D SLAM demonstrates robust performance in structurally rich urban environments and provides a strong foundation for future extensions involving IMU–Vision sensor fusion and advanced neural registration techniques.