Robust lidar-inertial SLAM by fusing multi-level geometric and intensity features

Abstract Conventional light detection and ranging (Lidar)-inertial simultaneous localization and mapping (SLAM) systems often exhibit degraded performance and poor robustness in complex scenarios characterized by structural degeneracy and perceptual aliasing. To address these limitations, this paper proposes an optimized Lidar SLAM framework that enhances localization accuracy through a synergistic fusion of multi-level geometric features and intensity information. Building upon the tightly-coupled architecture of LIO-SAM, our method introduces two primary contributions. In the front-end, we introduce a novel line feature extraction strategy based on local linearity analysis. Unlike conventional methods that rely solely on high-curvature corner points for edge constraints, our method explicitly identifies linear structures (e.g. curbs and railings) from non-corner points. These enriched line features are then integrated into a unified point-to-line and point-to-plane constraint model, significantly bolstering the resilience of the odometry against scene degradation. In the back-end, the Scan-Context descriptor is augmented with intensity data to mitigate perceptual aliasing, thereby improving the precision of loop closure detection in geometrically similar environments. The efficacy of the proposed method was validated through comprehensive experiments on the public KITTI dataset and a self-collected campus dataset featuring a large-scale loop, benchmarking against the state-of-the-art LIO-SAM. The results demonstrate substantial performance enhancements. On the KITTI 07 sequence, the absolute trajectory error (ATE) root mean square error (RMSE) was reduced by 46.7%. More notably, on our challenging self-collected dataset—where the original LIO-SAM suffered from catastrophic drift due to consistent loop closure failures—our approach reduced the ATE RMSE from 15.80 m to 3.84 m, a dramatic decrease of 75.7%. This study validates that our method effectively corrects accumulated error, significantly advancing the localization accuracy and robustness of SLAM systems in challenging real-world environments.