Document Type : Original Research Paper
Authors
1 Department of Civil Engineering, Faculty of Engineering, Bu-Ali Sina University, Hamadan, Iran
2 Department of Computer Science, Faculty of Engineering, Bu-Ali Sina University, Hamedan, Iran
Abstract
Background and Objectives: The extraction of 2D floorplans of building interiors plays a vital role in various domains, including architecture, surveying, building information modeling (BIM), robotics, and virtual reality. Mobile laser scanners capture the geometric structure of indoor environments with millimeter-level accuracy and record the results as point cloud data. Point clouds are a rich source of information for generating 2D floorplans of indoor spaces. However, surface reflection noise, occlusions caused by indoor objects, and the non-uniform density of points pose significant challenges for processing such data. Initially, 2D floorplan extraction relied on classical geometric methods. In recent years, however, deep learning-based approaches have gained increasing attention due to their strong ability to understand complex patterns and their robustness to noise. The main objective of this study is to present an effective framework for extracting 2D floorplans of building interiors from point cloud data using deep learning methods and to compare its performance with that of classical techniques.
Methods: In this study, an effective framework is proposed for extracting 2D floorplans of indoor building spaces from point cloud data, consisting of three sequential steps: data preprocessing, model implementation, and final evaluation. This framework enables a direct comparison between classical methods and deep learning approaches within a unified setting. Point cloud data are inherently discrete and unstructured, making direct processing challenging. In the preprocessing step, point clouds were projected onto a 2D space to generate density images, thereby reducing computational complexity. In the second step, two deep learning models, U-Net and Pix2Pix, as well as the classical Hough Transform algorithm, were implemented, with the density images serving as a common input for all methods. In the third step, the proposed framework was evaluated using publicly available datasets, including FloorNet and Structure3D. The input data were split into training, validation, and test sets, and data augmentation techniques were applied to improve model generalization. The performance of the models was assessed using the Dice Score and Intersection over Union (IoU) metrics.
Findings: Deep learning models demonstrated satisfactory performance on samples without occlusions, achieving accuracy levels above 90%. In particular, the U-Net model achieved a Dice Score of 97% on the Structure3D dataset. However, in samples containing occlusions, the models were unable to fully extract the floorplans. In contrast, the Hough Transform algorithm performed reasonably well in line detection but exhibited limitations in generating coherent and topologically valid outputs suitable for indoor map modeling due to its inability to capture topological structure. Moreover, the trial-and-error process required to tune the algorithm’s parameters significantly increased its runtime.
Conclusion: The findings of this study indicate that deep learning methods, when provided with complete data, are capable of accurately and structurally extracting 2D floorplans from point clouds. However, under real-world conditions where occlusion is inevitable, developing models that are robust to incomplete data becomes essential. To address this challenge, future research directions include employing hybrid architectures and incorporating complementary data sources such as RGB images or depth maps. The proposed framework in this study serves as an effective step toward the systematic comparison of 2D floorplan extraction methods and provides a foundation for developing more advanced models suitable for real-world applications.
Keywords
Main Subjects
COPYRIGHTS
© 2025 The Author(s). This is an open-access article distributed under the terms and conditions of the Creative Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
)