Building facade structure extraction method based on three-dimensional laser point cloud by considering semantic information

Figures & data

Figure 1. Flow chart of the proposed method.

Figure 2. The key to original and improving the CANUPO method (a) local dimensionality properties and (b) angular features.

Figure 3. The flow chart of building facade semantic segmentation based on improved CANUPO method.

Figure 4. Blocks of 3D LPCD.

Figure 5. The flow chart of point cloud segmenting.

Figure 6. RB-RANSAC line recognizing and regularized.

Figure 7. The flow chart of structure line optimization method.

Figure 8. Main structure optimization (a) structure lines before optimizing (b) structure lines after optimizing.

Table 1. Confusion matrix for binary classification evaluation.

		Truth value
		1	0
Predicted Value	1	True Positive (TP)	False Positive (FP)
	0	False Negative (FN)	True Negative (TN)

Table 2. Training data.

Number	Name	3D LPCD	Ground truth
Group 1	The front of CSMECB
Group 2	Teaching Building
Group 3	Inpatient

Table 3. Test data.

Number	Name	3D LPCD	Ground truth
Group 4	Administration Building
Group 5	The side of CSMECB
Group 6	Stgallen3

Table 4. Experimental data details.

Data	Name	Number of stations	Resolution (m)	Number of point cloud
Training data	The front of CSMECB	4	0.01	14,623,508
	Teaching Building	3		11,552,517
	Inpatient	2		9,624,730
Test data	Administration Building	3		7,173,799
	The side of CSMECB	1		1,825,375
	Stgallen3	1		4,927,719

Figure 9. Facade structure extraction results: (a) administrative building, (b) Stgallen3 and (c) side of the CSMECB.

Table 5. Evaluation metrics of extraction results.

Index	Administrative building			Average	Stgallen3			Average	Side of the CSMECB	Average
	1	2	3		4	5	6		7
Precision	0.97	0.92	0.89	0.92	0.90	0.90	0.91	0.90	0.90	0.90
Recall	0.96	0.94	0.98	0.96	0.85	0.90	0.95	0.90	0.94	0.94
RAL	0.99	0.90	0.84	0.91	0.90	0.95	0.98	0.94	0.96	0.96

Figure 10. Partial extraction results of building facade by different methods (the red lines represent the extracted structural lines): (a) Wang method, (b) Lu method and (c) proposed method.

Table 6. Evaluation metrics of local extraction results.

Metrics	Lu method	Wang method	Proposed method
Precision	0.59	0.71	0.90
Recall	0.75	0.56	0.94
RAL	0.58	0.51	0.96

Figure 11. Extraction result of Stgallen3: (a) Lu method and (b) proposed method.

Figure 12. Comparison between the original and improved CANUPO method: (a) classification results of the original CANUPO and (b) classification results of the improved CANUPO.

Table 7. Extraction results by the improved RB-RANSAC and original RANSAC.

Method	Results	Point cloud in red area	Time
RB-RANSAC			28.937s
RANSAC			130.672s

Figure 13. Training set (outside wall, windows, stairs and pillars).

Figure 14. Semantic segmentation results (windows-wall, pillars-wall and stairs-wall).

Table 8. Semantic segmentation evaluation of classifiers with different train sets.

Train sets	Density ($\mathit{\rho }$)	Number	ADPM
windows-wall	0.0128191	273233	3502.60
pillars-wall	0.0085137	150994	1285.51
stairs-wall	0.0125541	145936	1832.09

Data availability statement

The data and codes that support the findings of this study are available from the corresponding author upon reasonable request.