Line-LIO: high-precision line feature integration and quality-driven keyframe optimization for LiDAR-inertial odometry

Abstract LiDAR simultaneous localization and mapping systems face significant challenges in complex environments due to unreliable feature extraction and suboptimal keyframe selection strategies. This paper presents Line-LIO, a high-precision LiDAR-inertial odometry system that addresses these limitations through two technical innovations. First, we introduce a cross-scan line search-based line feature extraction method that employs a triple verification framework (linearity, verticality, and spatial continuity) to distinguish genuine line features from noise while using line center points as compact representations. Second, we develop a matching quality-driven keyframe selection strategy that forms a closed-loop feedback system by dynamically preventing the integration of poorly matched frames. Comprehensive evaluations on the public M2DGR dataset and a self-collected dataset demonstrate that Line-LIO significantly outperforms state-of-the-art methods without increasing computational burden. Line-LIO reduces the root mean square error of absolute trajectory error by up to 90.32% compared to LIO-SAM and 91.85% compared to FAST-LIO2 across diverse environments. The performance improvements are particularly pronounced in scenarios with rapid motion changes and complex geometric structures, confirming that Line-LIO effectively enhances feature extraction reliability while preventing cumulative errors caused by low-quality keyframes.