Comparative analysis of clutter removal methods and accuracy assessment in scan-to-BIM using smartphone LiDAR

다국어 초록 (Multilingual Abstract)

The use of Scan-to-BIM workflows in the Architecture, Engineering, and Construction （AEC） industry has increased significantly. However, indoor point clouds frequently contain non-structural clutter that requires time-consuming manual removal. Smartphone-based LiDAR provides a cost-effective alternative to traditional terrestrial laser scanning, but its uneven point density and higher noise levels introduce additional challenges for reliable clutter removal.
This study evaluates three approaches for clutter removal and segmentation of smartphone-acquired indoor point clouds: a rule-based filtering method, a commercial machine learning–based classification tool （Leica Cyclone 3DR Auto-Classify）, and a deep learning semantic segmentation model （PointNeXt）. Performance was assessed using point-wise accuracy, precision, recall, and F1-score against manually annotated ground truth data from five indoor environments.
In addition to clutter removal performance, geometric accuracy was evaluated using cloud-to-mesh distance analysis and Level of Accuracy （LOA） assessment to determine the reliability of the processed point clouds for Scan-to-BIM applications.
The results indicate that all methods improved point cloud usability for preprocessing. PointNeXt achieved the highest classification and geometric accuracy, demonstrating more reliable preservation of structural elements. Auto-Classify produced smoother and more visually coherent segmentation but showed inconsistencies in distinguishing detailed architectural components. The rule-based method provided a simple and interpretable approach but exhibited limitations in complex environments due to its sensitivity to data variability.
The findings highlight the importance of balancing clutter removal and structural preservation, while also demonstrating that geometric accuracy is a critical factor in evaluating preprocessing effectiveness. Overall, the study confirms the potential of smartphone LiDAR as an accessible solution for indoor Scan-to-BIM workflows when supported by appropriate preprocessing strategies.

목차 (Table of Contents)

• I. Introduction 1
• 1. Background and Research Context 1
• 2. Research Trends and Gaps 5
• 3. Research Motivation and Objectives 10
• 4. Research Contributions 13
• 5. Research Scope and Limitations 14
• II. Literature Review 17
• 1. Overview of Scan-to-BIM Preprocessing 17
• 2. Rule-Based and Geometry-Driven Segmentation Methods 19
• 3. Commercial and Open-Source Software 22
• 4. Deep Learning-Based Semantic Segmentation 24
• 5. Smartphone LiDAR in Indoor Scanning 26
• III. Research Methodology 28
• 1. Overall Workflow and Study Framework 28
• 2. Data Collection 30
• 3. Dataset Description 32
• 4. Experimental Setup and Computing Environment 35
• 5. Clutter Removal Methods 36
• IV. Results 40
• 1. Comparison of Clutter-Removal Outputs 40
• 2. Visualization of Clutter-Removed Point Clouds 45
• V. Evaluation and Validation 49
• 1. Ground Truth Generation and Annotation 49
• 2. Point-Wise Classification Metrics 50
• 3. Room-Level Performance Evaluation 51
• 4. Overall Average Performance Comparison 53
• VI. Accuracy Assessment 55
• 1. Cloud-to-Mesh Distance Analysis 55
• 2. Level of Accuracy Assessment for Scan-to-BIM 59
• 3. Spatial Error Visualization Using Scalar Field 65
• VII. Discussion 70
• 1. Method Comparison Summary 70
• 2. Impact of Clutter Removal on Scan-to-BIM Preprocessing 74
• 3. Effect of Dataset on Method Performance 75
• 4. Implications for Improving Smartphone-Based Scan-to-BIM Segmentation 76
• 5. Geometric Accuracy and LOA Evaluation 78
• VIII. Conclusion 80
• 1. Key Findings 80
• 2. Academic Contribution 81
• 3. Practical Contribution 82
• 4. Limitations of the Study 83
• 5. Future Research Directions 84
• References 85
• Appendix 97
• Appendix A. Histogram Distributions of Cloud-to-Mesh Deviations 97
• A.1 Rule-Based Method. 97
• A.2 Auto-Classify (Leica Cyclone 3DR) 100
• A.3 PointNeXt 103
• 국문초록 106
• Acknowledgment 108