IaNDT-SLAM: intensity-augmented NDT for robust LiDAR-inertial SLAM in unstructured environments

Light detection and ranging (LiDAR) simultaneous localization and mapping (SLAM) has proven effective for robust perception in challenging environments. However, most existing LiDAR SLAM methods primarily rely on geometric features for pose estimation, often neglecting the intensity information of each point. This limitation often results in registration failures, especially in unstructured environments where geometric features are sparse or absent. To address this limitation, we propose IaNDT-SLAM, an intensity-augmented normal distribution transform (NDT)-based LiDAR-inertial SLAM system that integrates intensity information to enhance the robustness of both odometry and loop closure detection (LCD). In the odometry process, a dynamic geometric-intensity fusion model is developed to adaptively balance the contributions of geometric and intensity information during NDT registration. To enhance the robustness of LCD against viewpoint changes and dynamic objects, we design a novel global descriptor, the entropy-intensity scan context (EISC), which effectively captures environmental structure while employing correlation coefficients to quantify the similarity between descriptors. Extensive evaluations on public and self-collected datasets demonstrate that IaNDT-SLAM reduces the absolute translational error (ATE) by 60.5% and improves the maximum F1-score in LCD by 29.5%, outperforming state-of-the-art (SOTA) geometric-only and intensity-assisted LiDAR SLAM methods in both accuracy and robustness.