LiDAR-Based Large-Scale Indoor Environment Mapping Incorporating Prior Scene Information

Abstract

Objective Point cloud maps serve as a crucial digital foundation for smart city development, underpinning applications such as autonomous driving, 3D reconstruction, and intelligent logistics. Compared to traditional static laser scanning, mobile platforms equipped with perception sensors offer a flexible solution for acquiring point cloud maps. However, mobile mapping in large-scale indoor environments remains challenging due to three key issues: (1) the large number of complex architectural structures, (2) insufficient registration accuracy between sparse scans, and (3) significant pose drift accumulated during long-term mapping. To address these problems, this paper proposes a novel LiDAR-based indoor positioning and mapping method that incorporates prior scene information. Methods The proposed method consists of four stages: feature extraction and organization of prior scene information, indoor global positioning, scan pose determination and optimization, and point cloud map generation. (1) Feature extraction and organization of prior scene information. We extract robust features with stable shapes and positions (e. g., walls, beams, and columns) from various prior data sources. For different types of prior information, we adopt customized extraction strategies: large language models (LLMs) for text-type data, edge and circle detection algorithms for image-type data, and file parsing for model-type data. Finally, all extracted features are organized into feature attribute files and grid maps to facilitate easy access (see Fig. 2). (2) Indoor global positioning. First, we extract similar robust features from real-time scans and establish architectural skeleton feature patterns, including line pairs and cylindrical shapes. Next, we improve the matrix descriptor to characterize line-pair features. Coarse localization is achieved by matching descriptors using Manhattan distance, followed by fine localization using point hit ratios on the grid map. (3) Scan pose determination and optimization. Scans are classified into three categories: feature-rich, regular, and feature-sparse. For feature-rich scans, we use our global positioning method to determine poses by matching architectural skeleton feature patterns. For regular scans, we adopt an improved ICP-based method that focuses on indexed skeleton features, while standard ICP is used for feature-sparse scans. Meanwhile, we detect loop closures using the similarity of architectural skeleton feature patterns and optimize poses for looped areas. For scenarios without loop closures, we employ key scan-based global localization and bundle adjustment to correct pose drift. (4) Point cloud map generation. Scans are stitched using optimal poses to generate a conventional point cloud map. A pre-trained RandLA-Net is applied for semantic segmentation, and the extracted semantic information is combined with architectural skeleton indices to produce a lightweight map. This map provides a compact and efficient representation suitable for large-scale indoor applications. Results and Discussions The proposed LiDAR-based mapping method was validated in three typical large-scale indoor environments: an office building, an exhibition hall, and an underground parking lot. Since most public datasets lack prior information, we collected scene data using a self-developed low-cost mobile LiDAR system (see Fig. 8). Additionally, we gathered various types of prior scene information for the experimental scenes, including building floor plans, fire evacuation diagrams, and architectural design specifications. For qualitative evaluation of global positioning, we visualized two randomly selected scan locations overlaid on floor plans across all scenes (see Fig. 10), which demonstrated accurate positioning. For quantitative evaluation, we compared positioning accuracy and efficiency with three existing methods (see Fig. 11). Our method achieved a positioning error within 0.10 m, an orientation error within 1°, and an average time cost of approximately 7 ms per scan—showing a clear performance advantage over the compared methods. Regarding mapping quality, visualization results (see Fig. 12) confirmed that the generated maps aligned with floor plans. Quantitative comparisons with LOAM, F-LOAM, and Traj-LO (see Fig. 13 and Table 3) indicated that our method achieved a mapping error within 0.1 m and outperformed these comparison algorithms in both accuracy and efficiency. Both qualitative and quantitative results verify the effectiveness of our method for mapping large-scale indoor environments. Furthermore, we confirmed that the offline lightweight point cloud map significantly improved localization efficiency while maintaining positioning accuracy across three tested localization methods (see Table 5). Conclusions This study presents an innovative LiDAR-based indoor mapping approach that integrates prior scene information to address the challenges of complex large-scale indoor environments. The proposed method effectively improves mapping accuracy, reduces computational complexity, and enhances the usability of point cloud maps for real-time applications. Experimental validations confirm that it can meet the indoor localization and environmental perception needs of mobile robots, highlighting its significant engineering value. Future research will focus on exploring lightweight semantic segmentation and online 3D model construction methods suitable for understanding large-scale indoor environments.