A Long-Term SLAM System With Graph-Structured Maps and Online Maintenance Using Dual-Layer Octree Change Detection

Environmental changes over extended periods pose significant challenges for map-based navigation applications, such as map volume expansion, localization errors caused by environmental changes or incomplete map coverage, and the inability to update maps in real time. This paper introduces a novel long-term LiDAR SLAM system that organizes maps into a graph structure, combines map-based localization with odometry data through pose fusion, and implements an efficient real-time change detection and map update mechanism, ensuring continuous positioning and map maintenance. To maintain manageable map volumes and simplify usage, key frames from laser scans are aggregated into sub-maps, connected via traversable paths within the graph. The system temporarily relies on odometry for short-term localization, re-optimizing when map conditions improve to achieve more accurate and robust positioning. A dual-layer octree with a ray-casting method is also integrated for rapid real-time change detection and map updates. Extensive testing on the NCLT long-term dataset and our self-collected factory dataset demonstrates that the proposed algorithm effectively controls map size, enhances localization accuracy and robustness, and supports real-time map updates.