A LiDAR SLAM System With Geometry Feature Group-Based Stable Feature Selection and Three-Stage Loop Closure Optimization

Nowadays, light detection and ranging (LiDAR) sensors have been increasingly used in robotics, particularly in autonomous vehicles, for localization and mapping tasks. However, the use of LiDAR simultaneous localization and mapping (SLAM) in various scenarios is still limited. In this article, we propose a LiDAR SLAM system that addresses this issue by grouping consistent and stable geometry feature to better express the environmental properties in both odometry and loop closure detection. Specifically, to achieve stable geometry feature extraction in the LiDAR odometry component, we adapt an adaptive feature extraction technique that extracts planar, linear, and point features. In addition, we cluster the extracted geometry features to filter out noise. We also analyze the geometry feature matching error and constraint consistency to ensure that the constraints built from these features are stable and repeatable. For the global optimization component, we construct a three-stage loop closure detection approach based on the distribution of geometry feature groups and their corresponding relationships. Quantitative and qualitative experiments on the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) dataset, MulRan dataset, and a dataset collected on university campus demonstrate the adaptability, accuracy, and repeatability of our method. In conclusion, our proposed LiDAR SLAM system improves the performance in complex and diverse scenarios by implementing stable geometry feature extraction, effective feature constraint classification, and accurate loop closure detection. The source code of our approach is available at https://github.com/qq1962572025/GeometrySLAM.