Title Page
ABSTRACT
국문 초록
Contents
NOMENCLATURE 15
1. Introduction 16
1.1. Research background 16
1.2. Problem statement 17
1.3. Research objective 17
1.4. Research flow 19
2. Preliminary Study 20
2.1. Research Trends on PCD-based Building Component Detection Algorithms 20
2.1.1. Three-dimensional space measuring method using LiDAR 20
2.1.2. Downsampling 22
2.1.3. Literature review 24
3. Development of an algorithm based on the geometric features by positional difference in pipe object 29
3.1. Outline 29
3.2. Algorithm process 31
3.2.1. Data preprocessing 31
3.2.2. Normal estimating with downsampling 32
3.2.3. Selecting bottom of pipe (BOP) PCD 35
3.2.4. Calculating 3D coordinates of points on the pipe centerline 39
3.2.5. Clustering PCD with pipe centerline 44
3.3. Determination of setting conditions and development of optimal algorithms 45
4. Applicability test 49
4.1. Outline 49
4.2. Algorithm performance evaluation method 51
4.2.1. F1 score 51
4.2.2. Test result 53
5. Conclusion 58
Reference 61
Table 1. Prior research on PCD-based object recognition 27
Table 2. Feature of pipe data according to cross-section level differences in individual object 41
Table 3. Algorithm setting conditions and functions 45
Table 4. Fundamental information of building for applicability test 49
Table 5. Four concepts of the confusion matrix and their meaning 51
Table 6. Confusion matrix per zone 55
Table 7. Precision, recall and F1 score per zone 55
Table 8. Comparison of algorithm in this study and widely used algorithm 56
Figure 1. Point cloud data described by Python Numpy 22
Figure 2. MEP detection algorithm flowchart 30
Figure 3. Comparison of PCD according to the difference in SOR set value 31
Figure 3-1. (a, b)=(6.0, 1.0) 31
Figure 3-2. (a, b)=(20.0, 1.0) 31
Figure 4. Normal estimation by variable value difference 34
Figure 4-1. (Voxel size, radius, maximum number of chosen data points)=(150.100.500) 34
Figure 4-2. (Voxel size, radius, maximum number of chosen data points)=(60.100.500) 34
Figure 4-3. (Voxel size, radius, maximum number of chosen data points)=(30.100.30) 34
Figure 5. Pipe PCD obtained through MLS method 36
Figure 6. Positional relationship between two normal vectors of data points in pipe 36
Figure 7. Positional relationship between normal vectors of data points around the joint where vertical and horizontal members meet 38
Figure 8. BOP PCD distribution influenced by voxel size and outer diameter 40
Figure 8-1. BOP data points distribution when voxel size is p and outer diameter is 2 × R1 39
Figure 8-2. BOP data points distribution when voxel size is q and outer diameter is 2 × R1 40
Figure 8-3. BOP data points distribution when voxel size is q and outer diameter is 2 × R2 40
Figure 9. Case of pipe PCD detection using BOP 42
Figure 10. Calculating 3D coordinates of pipe center point 43
Figure 11. Appearance of normal vectors at the ceiling corners 48
Figure 12. Differences in appearance of normal vectors according to proximity to edges 48
Figure 13. Applicability test data 50
Figure 14. Result of calculating the 3D coordinates of the pipe center point 53
Figure 14-1. Pipe center point-1 53
Figure 14-2. Pipe center point-2 53
Figure 15. MEP component visualizations detected by algorithm in zone A 57
Figure 16. MEP component visualizations detected by algorithm in zone B 57