Automation in Construction 22 (2012) 298–305
Contents lists available at SciVerse ScienceDirect
Automation in Construction journal homepage: www.elsevier.com/locate/autcon
Automatic processing of Terrestrial Laser Scanning data of building façades Joaquín Martínez a,⁎, Alex Soria-Medina b, Pedro Arias a, Alzir Felippe Buffara-Antunes b a b
Natural Resources Department, Mining School, Univerity of Vigo, Rua Maxwell S/N, 36310 Vigo, Spain Geomatics Department, Federal University of Parana, PO. Box 19.001, 81.531-990, Curitiba, Parana, Brazil
a r t i c l e
i n f o
Article history: Accepted 16 September 2011 Available online 9 November 2011 Keywords: Terrestrial Laser Scanner Building measurement Segmentation Border extraction Façade
a b s t r a c t Feature extraction on façades from unstructured point clouds is a challenging work, especially in the presence of noise. Point cloud segmentation is one of the most important steps in this context. In this paper, a new approach for automatic processing of façade laser scanner data is introduced. Scanner orientation is partially known through the inclination sensors of the laser scanner used. Knowing these values allows us to reduce the point cloud data into a profile distribution function. After orientation, this distribution is a series of peaks and valleys suitable for segmentation. Each segmented layer is afterwards processed to find the façade contours. The results obtained prove that the approach may be successfully employed in building segmentation and extraction of planar features. Moreover, the accuracy of contours is very dependent on the resolution of the scan data. © 2011 Elsevier B.V. All rights reserved.
1. Introduction The integration of laser measuring devices with photogrammetry is proposed in [1,2] for the inventory of buildings under construction. In recent years great progress has been made in terms of accuracy and speed in order to obtain and render 3D models of buildings and façades. The automatic extraction of different features on a façade is a fundamental problem in point cloud processing. Segmentation of different façade features can be helpful for localization, classification and feature extraction. Automatic processing of point cloud data results in highly versatile systems that allow the automation of repetitive processes for the acquisition of massive data on large sites [3,4]. In this sense, different approaches for segmentation are suggested in the literature. They differ mainly in the criteria used for similarity measurement within a given set of points and, hence, to carry out decisions about point clustering. The accuracy of the segmentation is strongly linked with the segmentation method. There are segmentation methods that have a large number of parameters, whose meaning and effect on final segmentation are not always clear. Most of the comparisons used separate iterative optimization methods to find the best set of parameters which describe the feature of the façade. However, many techniques applied to photogrammetric, computer vision and signal processing fields have been used for classification and segmentation of the point clouds resulting from TLS [5]. Some of these techniques include transformations into a parameter space, such as Hough transform and Gaussian sphere. In [6], the authors try to
bring together common elements based on the surface parameters and normal surface information respectively. Techniques such as region growing have been applied to segmented data based on localized information [7,8]. Morphological approaches such as medial axis and skeletonisation have also been used by introducing diffusion equations, radial basis function and grass-fire techniques [9,10]. According to [11], the reliability and accuracy of façade models generated from terrestrial data depend on data quality in terms of coverage, resolution and accuracy. Façade parts for which only little or inaccurate 3D information is available, cannot be reconstructed at all or require manual pre- or post-processing. When time-consuming user interaction is to be avoided, automatic modeling algorithms which can handle heterogeneous data are a solution. In this paper, a new approach to automatic segmentation of TLS point clouds is introduced. The main aim of this work is to extract a set of features of building façades in different layers based on planar features. The first step is the orientation of the point cloud by using the RANSAC (RAndom SAmple and Consensus) [12] paradigm. After that, in the following step the segmentation is performed using a profile distribution of data and local maxima and minima information. Finally, façade contour points are labeled. These contour points are suitable for façade measurement and may be exported to CAD software in order to be annotated. A council museum in Vigo was chosen to test the process. “Pazo Quiñones de León”, dates from 1670 and is an example of urban renaissance palace. This building was chosen as a case study due to the presence of planar features on its façade. 2. Related work
⁎ Corresponding author. E-mail addresses:
[email protected] (J. Martínez),
[email protected] (A. Soria-Medina),
[email protected] (P. Arias),
[email protected] (A.F. Buffara-Antunes). 0926-5805/$ – see front matter © 2011 Elsevier B.V. All rights reserved. doi:10.1016/j.autcon.2011.09.005
Segments are geometrically continuous elements of a surface or objects that have some similarities. Segmentation is the process in which points that have similar features on a surface are labeled as
J. Martínez et al. / Automation in Construction 22 (2012) 298–305
299
belonging to a segment. A homogeneous segment is characterized by its homogeneities based on certain feature, such as geometric properties, reflectance i.e. intensity data and spectral properties. This is an important step in the creation of model documentation from 3D point cloud. Various algorithms are proposed in the literature for laser point cloud segmentation and contour extraction. The methods of segmentation which are described below are based on geometric criteria while in Section 2.2 some methods for contour extraction are introduced.
enforcing proximity and surface smoothness criteria. The authors in [8], proposed an approach to automatically extract planar surfaces from TLS point clouds following the region-growing segmentation method in [6]. In this approach, several parameters need to be specified for the planar surface-growing algorithm, such as the number of seeds, the surface-growing radius and the maximum distance between surfaces. Using different values for these parameters, it is easy to obtain bad segmentation such as over-segmentation, under-segmentation and no segmentation.
2.1. Segmentation methods
2.1.3. Segmentation based on fitting model This method is based on the observation that man-made objects can be decomposed into geometric primitive features like planes, cylinders or spheres. This process tries to fit primitive shapes in the point cloud data to describe the form of building façades. One of the methods that use fitting models is the RANdom SAmple Consensus (RANSAC). The RANSAC algorithm is used to detect mathematical features like straight lines, circles and plane. Its principle is fully explained in [12]. While the RANSAC algorithm has the great advantage of being robust, even in the presence of noise, there are also shortcomings which should not be overlooked like the appearing of spurious surfaces, especially in the case of parallel planar surfaces like stairs. This problem is solved in [15] by sequentially applying algorithm to data. The principle of RANSAC algorithm entails the search of the best plane among a 3D point cloud. At the same time, it reduces the number of iterations, even if the number of points is very large. For this purpose, it randomly selects three points and calculates the parameters of the corresponding plane. Then it detects all points of the original cloud belonging to the calculated plane, according to a given threshold. Afterwards, it repeats these procedures N times; in each, it compares the obtained result with the last saved one. If the new results is better than the last one it is replaced. To find the features the RANSAC needs four input data (1) the 3D point cloud which is a matrix of three coordinate columns X, Y and Z, (2) the tolerance threshold of distance between the chosen plane and the other points. Its value is related to the altimetric accuracy of the point cloud, (3) the foreseeable is the maximum probable number of points belonging to the same plane. It is deduced from the point density and the maximum foreseeable plane surface (4). The probability α is the minimum probability of finding at least one good set of observations in N trials. According to [16], it lies usually between 0.90 and 0.99.
The most relevant methods investigated for the proposal of detecting and extracting features from point cloud laser scanning data are explained in the next sections, and are divided in 3 main groups: clustering of features, region growing and fitting model methods. 2.1.1. Segmentation based on clustering of features The methods based on clustering of features are useful to identify homogeneous patterns in point cloud data. This process firstly identifies patterns in the data based on attribute and data clusters grouping. Afterwards, the points belonging to each cluster are labeled as a unique segment in the space object. The results of this technique depend on the selected features. This technique has been found very sensitive to the noise data and influenced by the definition of neighborhood. An example of a segmentation algorithm using this technique is described below. In [13], a clustering algorithm using an unsupervised classification technique is presented for extracting homogeneous segments in Airborne Laser Scanning (ALS) data from unorganized point cloud containing only a limited amount of information (x, y, z). The authors defined seven dimensional vectors for each point, that are coordinate position, the parameters of the fitted plane to the neighborhood and the relative difference height between the points and its neighbors. Then, the feature space is clustered using unsupervised techniques to identify surface classes. After extracting the surface classes, the points are grouped in object space using spatial proximity measurement. 2.1.2. Segmentation based on region growing In this process, the algorithm starts at a chosen point and grows around the neighboring points based on certain similarity criteria. In TLS data for façade segmentation the region algorithm is used to extract planar surfaces. It starts by determining a group of nearby points. If a plane is found to fit the points within some predefined threshold, then this plane is accepted as a seed surface. Afterwards, seed surfaces grow according to specific criteria.. As an example, neighboring points may be added to the surface if they meet some predefined threshold. After addition of new points to the surface, the equation of the plane is updated. Several extensions for surface growing methods have been suggested. The authors in [9], present a variation of a region-growing algorithm for ALS data. A triangulated irregular network — TIN is used to describe the basic elements of the surface. The merging of triangular elements is carried out by comparing the plane equation of neighboring triangles. In [14], a segmentation method for ALS data based on region growing is proposed in which the normal vector at each point is estimated using the k-nearest points. In the growing step, neighboring points are added to the segment based on criteria of similarity in normal vectors, distance to the growing plane and distance to the current point. A method to segment industrial scenes based on a smoothness constraint is proposed in [6]. Local surface normal estimation and the region-growing method are used. The residuals in a plane fitting are employed to approximate the local surface curvature. The growing of segments is performed by using previously estimated point normals and their residuals. Points are added to the segment by
2.2. Façade contour extraction Contour extraction may be performed from the segmented planar features. Usually, this operation is used to construct a vector model of the building façades and to export this model to computer aided design software. Some research has been conducted to extract the contours using an algorithm based on Delaunay triangulation. These works use the lengths of the edges of the triangulated network to automatically extract the outer and inner contours of façades. The threshold value to extract the contours must be higher than the resolution of the point clouds used [16,17,4]. Following a different method, the authors in [18], use the Triangular Irregular Network (TIN) to define the segment walls. Only long TIN edges appear at the outer boundary (wall outline) or inner boundary (holes) of a wall. Boundary points are just the end points of the long TIN edges. The geometric reconstruction can be seen as a process of polygon fitting plus the generation of knowledge based assumptions for occluded parts. Polygon fitting is achieved by directly applying least squares fitting, the Quick hull method, or the Hough transform to extracted feature segments. 3. Methodology In this section, the methodology used for the processing of terrestrial laser scanner (TLS) data is presented. The laser scanner used is a
300
J. Martínez et al. / Automation in Construction 22 (2012) 298–305
time-of-flight (TOF) Riegl LMS-390i, which is a hybrid laser scanner. This means that the scanning is unrestricted in the horizontal movement with a 360° field of view (HFOV) and has a restricted vertical field of view (VFOV) of 80°. Some technical data from Riegl specifications are shown in Table 1. Façade data processing is divided in four main procedures: data acquisition, data orientation, data cleaning and segmentation and, finally, contour extraction. These last three procedures are fully automated and performed with no need of user intervention. In the following subsections three coordinate systems are mentioned: the scanner own coordinate system, the up-right coordinate system and the final oriented coordinate system, of which YZ plane is parallel with the façade. These coordinate systems are depicted in Fig. 1. 3.1. Data acquisition One key feature of the Riegl LMS-390i TLS is the fact that it is equipped with inclination sensors. These inclination sensors permit us to obtain an up-right point cloud of the façade to be processed. The measurement of the sensors ranges from − 5° to +5° and will only provide valid values in the alignments labeled by the manufacturer as standard and lay back. In this work, we assume that the laser scanner is set in standard (vertical up-right position) alignment and near vertical. Thus, rotation angle values about X and Y axes are known and we can transform the scanner own coordinate system to the up-right coordinate system. In data acquisition, the angular step width of a façade scanning is defined as the angular interval between two consecutive measurements of the laser scanner. If the distance between the laser scanner and the façade is known, we can easily transform the angular value of the laser scanner into a distance step width interval δ(mm). The quality [19] in the detection of an object with minimum size D (mm) is:
Q ¼ 1−
δ : D
Fig. 1. Coordinate systems for façade processing: scanner own coordinate system (SOCS), up-right coordinate system (UP) and final oriented coordinate system (F). The last coordinate system is parallel to the façade.
transform the points in the up-right coordinate system to the final coordinate system is a Z axis rotation angle. In order to calculate the proper Z rotation angle, the histogram of the X coordinates is computed, obtaining a simple 1D signal with the profiles of the point cloud. The bin width of the histogram is set to δ = 5 mm, because this value is similar to the precision of the laser scanner distance meter. In case we had a perfect plane, an oriented profile would tend to be a Dirac delta distribution and an unoriented profile would tend to be a constant distribution. Thus, the profile distribution function gives us an idea of how much oriented to the façade we are. In Fig. 2 the difference between unoriented and oriented profile distribution is shown. Laser data are processed creating a histogram that is followed by a threshold in order to filter data. After this thresholding, only the bins with the larger number of points are processed. Next step consists of plane extraction using the RANSAC algorithm. In the up-right coordinate system, the equation of the plane is simplified:
ð1Þ A x⋅ þ B y⋅ þ C ¼ 0
Given that the distance step width and the minimum size are both distances, Q is dimensionless. The Q value is a measure of data quality and indicates the level to which the object has been scanned. Negative values of Q would be considered an unacceptable fit whereas positive values would show the percentage confidence that the object will be detectable. Decreasing the sample interval will lead to a more dense point cloud in which smaller objects are easier to detect but heavier to process. For every façade, three successive scannings with approximate step widths of 100, 50 and 10 mm have been performed. 3.2. Data orientation The façades of historic building to be processed are mainly composed of planar features, therefore, orientation is achieved by using the planar information of the façade. The orientation plane is computed using the RANSAC algorithm [12]. The only unknown angle to
ð2Þ
where (A, B, 0) is the normal vector of the plane. In this situation, the rotation matrix RZ is given by the expression 2
A 6 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 6 A2 þ B2 6 B RZ ¼ 6 6 − pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 4 2 A þ B2 0
B pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi A2 þ B2 A pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 A þ B2 0
3 07 7 7 7: 07 5 1
ð3Þ
The oriented coordinates of data are known by applying the rotation matrix RZ. This orientation procedure is repeated iteratively until the rotation angle is lower than a threshold. In this case, the threshold is set to the value of the precision of the inclination sensors of the Riegl TLS in Table 1, i.e. 0.05°. The work flow to get the façade oriented is shown in Fig. 3. 3.3. Data cleaning and segmentation
Table 1 Some technical data from Riegl LMS-390i specifications. Max angular resolution Measurement range Accuracy Repeatability Laser beam divergence (full angle, Inclination sensors precision
1 e2
value)
0.002° 1–400 m 6 mm (one sigma at 50 m) 4 mm (one sigma at 50 m) typ. 0.3 mrad 0.05°
The next automatic procedure is data cleaning and segmentation. After point cloud orientation is completed, the profile distribution is a series of peaks and valleys suitable for segmentation by finding local maxima and minima. Data are filtered for non valid point reduction, by applying a gate depending on the distance of the X coordinates of the points to the X coordinate of the maximum of the profile distribution. In our tests, this filter performed the deletion of the points more than 2 m in front and behind the façade. This threshold depends on the façade, thus, if a façade is more than 4 m long, this threshold
J. Martínez et al. / Automation in Construction 22 (2012) 298–305
301
Fig. 2. Comparison of the orientation of a scanning with 100 mm resolution. The unoriented profile distribution (left) of the X coordinates of points is a general random distribution. The oriented profile (right) distribution tends to be a Dirac's distribution.
should be changed. Fig. 4 shows the oriented profile distribution of Fig. 2 after applying the cleaning filter. After orientation, laser data is segmented into different classes that share some common characteristics. In this presentation such classes are defined as layers. These layers are obtained using the local maxima and minima from the profile distribution. Firstly, local maxima and minima coordinates are filtered in order to reject consecutive maxima or minima to obtain a series of consecutive maximum–minimum pairs. Segmentation layers are placed at each maximum position and are limited by the two adjacent minima. Fig. 5 shows a close-up image of the profile distribution function overlaid with the series of maxima and minima that define the segmented layers. Usually, a great number of layers may be found, hence a filter that reduces the number of layers was developed. This filter takes into account the statistical information of each layer derived from the profile distribution and selects only layers
RAW DATA
that have less points than the thirtieth percentile. Next, we look for the successive layers that can be grouped together to form a new distribution of layers. This algorithm is repeated until there are no changes in the distribution. In Fig. 6 the final segmentation layers are shown. The limits of the segmentation layers are scaled by its relative frequency. 3.4. Contour extraction The final automatic procedure in this work is the extraction of the contours of the façade as final result. The contours of the façade are obtained by processing each segmented layer as shown in the work flow in Fig. 3. The layer processing consists of three steps: 1. Point cloud triangulation. The layer points are orthogonally projected to the plane X = Ml, being Ml the X coordinate of the local
SEGMENTED LAYERS
Laser data profiling
Profile Thresholding
Detail layer?
Detail contour extraction
Export contours
RANSAC plane extraction
Point Cloud Orientation
Angle< Threshold ?
Fig. 3. Workflow of façade segmentation and contour extraction.
Facade extraction
302
J. Martínez et al. / Automation in Construction 22 (2012) 298–305
Fig. 4. Oriented profile distribution of Fig. 1 filtered by X distance gate. The different layers in the scanning data are visible in the profile distribution.
maximum of the layer, and triangulated using a 2D Delaunay triangulation. 2. The triangulation is filtered to remove the edges of the triangulation which are higher than a threshold. As a result, the layer is split in a group of polygons to be processed for border extraction. 3. Labeling and, optionally, refining the border points of each triangulation polygon as contour points of the layer. The threshold of point 2 and the border extraction in point 3 depend on the properties of the layer. We define a detail layer as a layer that contains several separate objects (such as windows or door objects) and a façade layer (wall) as a layer that contains a few objects that fill most of the layer. We classify each layer depending on its point density. Let us define the point density of a layer as the ratio dl ¼
Nl Al
ð4Þ
where Al is the area of the convex hull of the layer (m 2) and Nl is the number of points in the layer. Layers with lower density ratio are labeled as detail layer whereas layers with higher point density ratio are labeled as façade layer. If a layer is labeled as a façade, a lower
Fig. 6. Profile distribution after grouping of layers. The star values are the limits of the layers scaled by its relative frequency.
number of objects is expected to form the contour of the façade. The point cloud is decimated to a resolution of δd = 300 mm in order to speed up the process of finding out a rough version of the layer contours in the layer, while preserving the confidence given by Eq. (1) to find objects larger than 1 m. This rough version is segmented in different polygonal objects which are individually refined and processed in detail to arrive at the final results. However, if a layer is labeled as a detail layer, the threshold of the edge length in the triangulation is set to two times the resolution of the scanning δ, in order to obtain as many details of the border as possible. If a layer is misclassified as detail or façade layer, a speed down of the process is produced, especially in a façade layer misclassified as detail. The above steps can be summarized in the following algorithm, written in pseudo code: IMPORT 3d point data repeat Laser data PROFILING Profile THRESHOLDING PLANE extraction Point cloud ORIENTATION until orientation angle b inclination sensor precision for each layer if layer is a façade DECIMATE layer TRIANGULATE decimated layer SPLIT the decimated layer in polygons for each polygon REFINE the polygon in original layer LABEL the contour points end for else TRIANGULATE layer LABEL the contour points
Table 2 Some properties of the laser scanning of “Quiñones de León” façade using Riegl LMS390i TLS.
Fig. 5. Oriented Profile distribution of Fig. 1. The segmentation layers are centered at maxima point (negative star values) and limited by the two adjacent minima (positive star valued).
Scanning linear approx. step width Scanning angular resolution deg Scanning time Number of points of scan data
100 mm 0.135 25″ 57,949
50 mm 0.068 1′42″ 236,622
10 mm 0.014 19′48″ 5,888,344
J. Martínez et al. / Automation in Construction 22 (2012) 298–305
end if end for EXPORT contour points 4. Experiments and results The historic building that was chosen to test the process is the “Pazo Quiñones de León” that is a council museum in Vigo [20]. The building dates from 1670 when it was constructed after the former building was destroyed during the war. The building is an example of an urban Renaissance palace and is in the style of a main building flanked by two towers. In the late nineteenth and early twentieth centuries the building was renovated by the Marques de Alcedo who donated the building to the people of Vigo in 1924. The building was scanned with approximate step widths of 100, 50 and 10 mm as mentioned in Section 3 from an approximate distance of 40 m. In Table 2 some data from the scanning process are shown. This building was chosen as a case study due to the presence of planar features on its façade. All of the system procedures were implemented in Matlab. Point cloud data were exported as ascii files from and to Riegl Riscan Pro software that was used with presentation purposes. The processing time, refers to the execution on a PC with a Core i5 CPU.
303
4.2. Façade contour extraction In order to test the validity of the façade contours, the façade was manually measured using a Leica TCR1102 total station as a reference. A total of 81 distinguishable points on the borders of the façade were measured to form a set of 45 distances on the façade. These distances were compared with those measured from the extracted contour points. In Fig. 9 the results of contour extraction from the 10 mm scan are shown, showing some of the measured distances for testing. When comparing errors in measurements, it is necessary to take into account the restrictions implied by Eq. (1) and the different data processing steps. Among the processing steps, two of the procedures were particularly important. Firstly, orthogonal projection of points in contour extraction was especially important, given that it is the only procedure that affects the actual point position directly. Secondly, the final accuracy of the measurements was affected by the orientation process, given that errors in the orientation angles, introduced a scale factor to the measurements. In Table 3 some statistics of the errors are shown. As expected , there were minor errors as the step width is finer. The processing time for extracting the façade contours of the scans with 100 mm, 50 mm and 10 mm step width was respectively 20″, 71″, and about 2400″. The amount of information to be processed with very fine scans is evident with these last data. 5. Conclusions
4.1. Segmentation process The segmentation process also comprises the orientation and cleaning of scan data. In Fig. 7 the result of the segmentation process for the scan of 10 mm resolution overlayed with the original scanning is shown. For the scan with 100 mm resolution a total of 19 layers were created, whereas 21 layers were created in the processing of 50 mm and 10 mm scans. The distribution of the layers for the scans of 50 mm and 10 mm was very similar: the maximum distance between the locations of the layers is 21 mm. The processing time for the scans of 100 mm, 50 mm and 10 mm was respectively of 12.84″, 33.44″, and 125.28″. In Fig. 8, the final result of segmentation is shown.
In this paper a simple and effective approach to the segmentation and processing of buildings façades data is presented. This research is based on planar information on the façade and is therefore suitable for characterizing building façades. Vertical orientation of the scans is achieved through the angles provided by the inclination sensors of the Riegl laser scanner. The unknown angle is obtained automatically using RANSAC. 3D point cloud data are processed in order to obtain an unidimensional profile distribution function that permits the orientation , cleaning and segmentation of data. As a first step, the main plane of the façade is extracted using the RANSAC paradigm and the coordinate system is rotated in order to make the X coordinate axis perpendicular to this plane. A threshold for façade depth is
Fig. 7. Comparison of the original scan data (rotated on the left) and the oriented and segmented scan (right).
304
J. Martínez et al. / Automation in Construction 22 (2012) 298–305
Fig. 8. Results of segmentation. Different colors show the different achieved layers in the façade.
established in order to automatically remove points outside the façade. Automatic segmentation of façade points is achieved by using local maxima and minima in the profile distribution function and statistical layer grouping. Segmentation results are satisfactory and similar results are obtained for different scanning resolutions, having a maximum difference of 21 mm between the layer maximum position. Point cloud segmentation in different layers simplifies further data processing. In this research layers are used for automatic façade contour extraction. The quality and accuracy of contour
extraction are absolutely dependent on the scanning sample width, because borders are high frequency zones in which step width is essential. For a 10 mm step width scanning, a mean error of 7 mm is obtained when compared to total station measurements. The error standard deviation is 19 mm. Processing time for the Matlab implementation is acceptable, although for very dense point clouds a native implementation should be considered as a future work. Future work will focus on solving the segmentation on more complex building façades, on determining the threshold to generalize parameters
Fig. 9. Results of contour plotting and annotation with a CAD software.
J. Martínez et al. / Automation in Construction 22 (2012) 298–305 Table 3 Errors in the measurements of the façade contours for the different scanning step width. Scanning linear approx. step width Mean error (m) Standard deviation (m)
10 mm 0.007 0.019
50 mm 0.035 0.043
100 mm 0.083 0.074
and automatically export geometric primitives to computer aided design software. Acknowledgments Thanks to CAPES Foundation, Ministry of Education of Brazil (Process code 0064-10-6), the Regional Ministry of Industry of the Xunta de Galicia (Isabel Barreto Human Resources Program exp. 114) and the Ministry of Science and Innovation of Spain (Project code BIA200908012) for the financial support. References [1] S. El-Omari, O. Moselhi, Integrating 3D laser scanning and photogrammetry for progress measurement of construction work, Automation in Construction 18 (1) (2008) 1–9. [2] C. Ordóñez, J. Martínez, P. Arias, J. Armesto, Measuring building façades with a low-cost close-range photogrammetry system, Automation in Construction 19 (6) (2010) 742–749. [3] C. Brenner, Building reconstruction from images and laser scanning, International Journal of Applied Earth Observation and Geoinformation 6 (3–4) (2005) 187–198. [4] S. Becker, Generation and application of rules for quality dependent façade reconstruction, ISPRS Journal of Photogrammetry and Remote Sensing 64 (6) (2009) 640–653. [5] S. Filin, Surface classification from airborne laser scanning data, Computers and Geosciences 30 (9–10) (2004) 1033–1041. [6] T.M. Awwad, Q. Zhu, Z. Du, Y. Zhang, An improved segmentation approach for planar surfaces from unstructured 3D point clouds, The Photogrammetric Record 25 (129) (2010) 5–23.
305
[7] H. Boulaassal, T. Landes, P. Grussenmeyer, Automatic extraction of planar clusters and their contours on building façades recorded by terrestrial laser scanner, International Journal of Architectural Computing 7 (1) (2009) 1–20. [8] S. Pu, G. Vosselman, Knowledge based reconstruction of building models from terrestrial laser scanning data, ISPRS Journal of Photogrammetry and Remote Sensing 64 (6) (2009) 575–584. [9] M. Alshawa, H. Boulaassal, T. Landes, P. Grussenmeyer, Acquisition and Automatic Extraction Of Façade Elements on Large Sites from a Low Cost Laser Mobile Mapping System, ISPRS/3DARCH09, 2009. [10] D. Belton, D.D. Lichti, Classification and segmentation of terrestrial laser scanner point clouds using local variance information, ISPRS Image Engineering and Vision Metrology 61 (5) (2007) 307–324. [11] G. Vosselman, B. Gorte, G. Sithole, T. Rabbani, Recognising structure in laser scanner point clouds, The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 46 (8) (2004) 33–38. [12] S. Pu, G. Vosselman, Automatic extraction of building features from terrestrial laser scanning, The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 36 (part 5) (2006) 25–27. [13] B. Gorte, N. Pfeifer, Structuring laser-scanned trees using 3D mathematical morphology, International Archives of Photogrammetry and Remote Sensing 35 (B5) (2004) 929–933. [14] W. Ma, F. Wu, M. Ouhyoung, Skeleton extraction of 3d objects with radial basis functions, Shape Modeling International (2003) 207–215 IEEE. [15] M.A. Fischler, R.C. Bolles, Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography, Communications of the ACM 246 (1981) 381–395. [16] D. Tóvári, N. Pfeifer, Segmentation based robust interpolation-a new approach to laser data filtering. International archives of photogrammetry, Remote Sensing and Spatial Information Sciences 36 (part 3) (2005) 79–84. [17] H. Boulaassal, T. Landes, P. Grussenmeyer, F. Tarsha-Kurdi, Automatic segmentation of building façades using terrestrial laser data, Proceedings of the ISPRS Workshop ‘Laser Scanning 2007 and SilviLaser 2007’, Espoo, Finland, 12–14 Sept. 2007, 2007, pp. 65–70. [18] R.C. Gonzalez, R.E. Woods, Digital Image Processing, Prentice Hall, Upper Saddle River, N.J., 2008 [19] J. Mills, D. Barber, P. Bryan, An Addendum to the Metric Survey Specifications for English Heritage—the collection and archiving of point cloud data obtained by terrestrial laser scanning or other methods, http://www.heritage3d.org (Accessed 25 September 2010). [20] http://www.museodevigo.org Concello de Vigo. Museo Municipal de Vigo. (Accessed 23 November 2010).
