Extended Abstract
Introduction
Extraction of two-dimensional floor plans of residential buildings is a fundamental requirement in architectural, civil engineering, and surveying applications. A 2D floor plan provides a vector-based representation of a building’s interior layout, accurately depicting the geometric configuration and spatial arrangement of walls and rooms. The process of extracting residential floor plans is typically time-consuming and requires extensive human involvement. The implementation of semi-automated approaches can significantly facilitate this process, improving both the accuracy and efficiency of floor plan extraction. Point cloud data consist of a discrete set of three-dimensional points that accurately describe the geometric and spatial characteristics of an environment. With the advancement of 3D scanning technologies, rapid, precise, and comprehensive acquisition of indoor building spaces in the form of three-dimensional point cloud data has become increasingly feasible. Accordingly, the use of point cloud data as input for the problem of 2D floor plan extraction in indoor surveying has been steadily increasing. Despite recent advances in deep learning methods, most existing approaches have been developed for structured data or outdoor environments, and their performance remains challenging in enclosed indoor spaces with complex structural layouts. Therefore, the extraction of two-dimensional floor plans of residential buildings from three-dimensional point cloud data—particularly with an emphasis on corner detection, wall boundary delineation, and room topology reconstruction—remains a challenging and open problem. The objective of this study is to propose a novel approach for the vector-based representation of indoor residential spaces from point cloud data, grounded in principal component analysis, deep learning techniques, and morphological operators.
Materials & Methods
The main contribution of this research lies in the development of an effective framework for extracting two-dimensional floor plans of indoor residential spaces from three-dimensional laser-scanned point clouds. The proposed floor plan extraction process consists of three principal stages: data preprocessing, segmentation, and vector-based reconstruction. In the first stage, the normal vector of each point is estimated using the Principal Component Analysis (PCA) algorithm. The normal vector at a point is defined as the vector perpendicular to the surface at that location. The main idea is to utilize the orientation of the normal vectors as a criterion for surface type identification. Points are classified based on the nz component of their normal vectors in order to accurately distinguish between horizontal and vertical surfaces. By removing points corresponding to horizontal surfaces—such as floors, ceilings, and horizontal furniture elements—only vertical surfaces are preserved for subsequent processing. to reduce computational complexity in terms of execution time and memory consumption, the processed point cloud is transformed into a density image, and the floor plan extraction procedure is subsequently carried out in the two-dimensional image domain. Focusing on points associated with vertical wall surfaces, which define the structural boundaries of the space, enables more accurate and efficient reconstruction of the floor plan. The density image represents the spatial distribution of points projected onto the horizontal X–Y plane and serves as the basis for subsequent data analysis stages.
In the second stage, two deep learning models—U-Net and SegFormer—are employed to extract a pixel-based two-dimensional floor plan from the density image. In the third stage, a set of morphological operations is applied to convert the pixel-wise model output into vector-based representations.
To evaluate the performance of the proposed method and to gain deeper insight into its capabilities and limitations, the Structured3D point cloud dataset was employed.
Results & Discussion
The U‑Net and SegFormer models were trained to segment the density images. In this stage, model performance was evaluated on the test dataset using the Dice Score and Mean Intersection over Union (Mean IoU) metrics. The U-Net and SegFormer models achieved Dice Scores of 97% and 94%, respectively, and Mean IoU values of 94% and 90%, respectively. The results of the study indicated that, although SegFormer performs well in capturing overall structural regions, it exhibits lower accuracy in reconstructing thin lines and precise wall boundaries compared to the U-Net model. Ultimately, U-Net was selected as the final model due to its superior performance across the evaluation metrics.
Morphological operators were used for the vector representation of the maps. To evaluate the accuracy of the vector reconstruction of the building maps, the standard metrics Precision and Recall were measured at the corner and room levels and were compared with the results of the RoomFormer model. The research findings showed that the proposed method achieved accuracies of 84% and 85% in the Precision and Recall metrics at the room level, and 75% and 82% at the corner level. The RoomFormer model achieved accuracies of 97.9% and 96.7% in the Precision and Recall metrics at the room level, and 89.1% and 85.3% at the corner level. Unlike previous methods that rely on the two‑stage training of Mask R‑CNN and transformer models to generate vector maps, the proposed method has lower computational complexity and simultaneously provides stable results with desirable quality.
Conclusion
In this study, a semi-automated method for extracting two-dimensional floor plans from laser-scanned point cloud data is presented. In the first stage, the model output is generated as a pixel-based image, which is then converted into a vector representation suitable for CAD software using morphological operations. The subsequent floor plan generation process still requires human-guided drafting and revision. This approach, however, provides significant time savings and simplifies the subsequent stages of map refinement. This approach enables easy editing and refinement of terrestrial scan results by surveying engineers. The main advantage of the proposed method lies in its semi-automated workflow, user interactivity, and substantial reduction of the manual processing time required to generate floor plans from raw surveying scanner data.
The experimental results demonstrated that the proposed method, with its simpler architecture, lower computational resource requirements, and satisfactory quality in vector-based floor plan representation, can serve as a practical and cost-effective alternative to transformer-based approaches.