Enhanced LiDAR-inertial SLAM with adaptive intensity feature extraction and fusion

• An adaptive intensity edge feature extraction algorithm is proposed. • It’s suitable for mechanical and solid-state light detection and ranging (LiDAR). • An adaptive weighted update scheme of geometric and intensity features is proposed. • The proposed method improved the accuracy and robustness of LiDAR-inertial odometry. Light detection and ranging (LiDAR) simultaneous localization and mapping (SLAM) demonstrates superior accuracy and robustness compared to visual SLAM. Currently, LiDAR SLAM predominantly relies on the curvature of scan points to extract geometric features from the environment, enabling effective reconstruction and localization in scenes with abundant geometric features. However, it encounters considerable challenges in degenerate scenarios such as long straight corridors or tunnels where the scarcity of distinctive geometric features poses significant difficulties. In this study, an enhanced LiDAR-inertial odometry (LIO) system with intensity feature adaptive extracting and fusion is introduced to address the issue of LIO in degraded environments. It is implemented within an iterative error-state Kalman filter (IESKF) framework. Firstly, point clouds are refined by filtering out outliers, and state prediction is conducted using observations of inertial measurement unit (IMU). Subsequently, building on the original extraction of geometric features from the point cloud, intensity edge features are adaptively extracted by setting thresholds based on intensity values of the point cloud. Distortion correction of the point cloud is performed using the predicted state from the IMU. Finally, the proposed method performs an adaptive weighted update with consideration of the observation quality of geometric and intensity features. Experiments were conducted using datasets from both conventional and challenging environments and compared with LIO-SAM, FAST-LIO2 and Point-LIO. The intensity-enhanced LIO demonstrated significant improvements in both precision and robustness with improvements of 1.44 %∼91.08 % in conventional scenarios and 1.98 %∼83.44 % in challenging scenarios compared with the existing systems.