用语义分区管理3D地图，让机器人省内存、快导航
Semantic Zone based 3D Map Management for Mobile Robot
- 按功能区域划分空间，以语义区为单位管理地图数据
- 内存占用降低40%以上，加载次数减少60%以上
- 适合医院、仓库等大型室内场景的智能导航
在医院和物流中心等大规模室内环境中，移动机器人需要精确的3D空间表示。然而，3D地图占用大量内存，难以在有限计算资源下保存完整数据。现有SLAM框架通常依赖几何距离或时间度量进行内存管理，导致在空间分隔明显的环境中数据检索效率低下。为此，我们提出一种基于语义区的3D地图管理方法，将管理范式从几何中心转向语义中心。该方法将环境划分为有意义的空间单元（如大厅、走廊），并以这些区域作为内存管理的基本单位。通过动态将任务相关的区域载入工作内存（WM），并将非活跃区域移至长期记忆（LTM），系统严格遵守用户定义的内存阈值。在RTAB-Map框架中实现后，相比标准方法，本方法显著减少了不必要的特征载入/卸载周期和累积内存使用量。结果表明，语义区管理可确保稳定的内存使用，同时保持导航所需的地图可用性。代码已开源：https://github.com/huichangs/rtabmap/tree/segment
原文摘要 · Abstract (English)
Mobile robots in large-scale indoor environments, such as hospitals and logistics centers, require accurate 3D spatial representations. However, 3D maps consume substantial memory, making it difficult to maintain complete map data within limited computational resources. Existing SLAM frameworks typically rely on geometric distance or temporal metrics for memory management, often resulting in inefficient data retrieval in spatially compartmentalized environments. To address this, we propose a semantic zone-based 3D map management method that shifts the paradigm from geometry-centric to semantics-centric control. Our approach partitions the environment into meaningful spatial units (e.g., lobbies, hallways) and designates these zones as the primary unit for memory management. By dynamically loading only task-relevant zones into Working Memory (WM) and offloading inactive zones to Long-Term Memory (LTM), the system strictly enforces user-defined memory thresholds. Implemented within the RTAB-Map framework, our method demonstrates substantial reductions in unnecessary signature load/unload cycles and cumulative memory utilization compared to standard approaches. The results confirm that semantic zone-based management ensures stable, predictable memory usage while preserving map availability for navigation. Code is available at: https://github.com/huichangs/rtabmap/tree/segment
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。