3d surface reconstruction of terrestrial laser scanner data for forestry
3D SURFACE RECONSTRUCTION OF TERRESTRIAL LASER SCANNER DATA FOR FORESTRY Hongjoo Park*, Samsung Lim*, John Trinder* and Russell Turner** * School of Surveying and Spatial Information Systems, UNSW, Australia Email: [email protected], [email protected], [email protected] ** Industry and Investment NSW, Australia Email: [email protected] Abstract Recently, Terrestrial Laser Scanners (TLS) have received considerable attention for their potential applications in forest management, archaeology, ecology as well as remote sensing and urban planning applications. Although TLS is limited in its use in small areas, it is feasible to be applied to forest inventory and deliver better sampling accuracy, objectivity, and can enhance or replace field surveys in forestry. This paper presents a framework for using TLS measurements taken by Leica HDS6000 TLS produce 3D point cloud data and to model individual trees. In particular, the quantitative and qualitative analysis of the 3D point cloud data for four different types of trees derived by TLS is discussed and the processing steps are presented. The Crust algorithm is used for the reconstruction of surfaces of arbitrary topology from the 3D point cloud data. The four individual tree models derived from the TLS system and their 3D surface reconstruction by the field survey of individual tree surfaces are possible with this technology. Keywords: TLS, Crust Algorithm, 3D Tree Surface Modelling, Forestry 1. Introduction For the past decades, analytic forest research has been conducted using laser-scanning techniques. Terrestrial Laser Scanners (TLS) is thought to be especially valuable in the 3D representation of individual trees because TLS data have a much higher resolution than the conventional field surveying methods (Bae, 2008). Many researchers have attempted the reconstruction of forestry surface in 3D using different methods, for example, Ivanov et al. in 1995 used close-range photogrammetry in modelling the maize canopy. Similarly, Hosoi (2006) made use of high-resolution portable scanning lidar when estimating leaf area density by voxel-based 3D modelling of individual trees, whereas Wang et al. (2007) combined laser scanning, CAD and crop growth mathematical models for crop modelling. Aguilar et al. (2008) has proposed close-range photogrammetry for 3D surface modelling of tomato plants.
Despite the tremendous attention given to the 3D modelling of individual tree surfaces, it remains challenging to obtain an accurate 3D surface reconstruction of these structures. This paper aims to utilise a simple computer graphics algorithm in delivering an objective reproduction of individual tree surfaces by triangulating the surface by approximating an indicator function of the model. It is believed that the use of this algorithm will also enable us to reduce the total data processing time. 2. Crust Algorithm for 3D Surface Reconstruction Three-dimensional Voronoi diagrams and Delaunay triangulations are the basis of the crust algorithm which produces a set of triangles called the crust of sample points. All crust triangles are based on a Delaunay triangulation using sample points as the vertices of crust triangles (Amenta, 1998). The crust algorithm may fail when the number of samples is small. The three-dimensional crust is a set of triangles resembling the geometrical surface. If S is a set of r-samples from a smooth surface F, then for r 0.06 the crust of S contains a subset of triangles forming a mesh topologically equivalent to F and for every point on the crust lies within a distance 5r * d(p) of some point p on F, where d(p) is the distance from p to the medial axis (Amenta, 1998). 3. Data Acquisition The TLS data was collected on the 21st of May 2009 using a Leica HDS 6000 system. In Figure 1, the imagery on the left-hand side outlines the study area where red-coloured circles showing four different TLS Ground Control Points (GCPs). The imagery on the right-hand side illustrates the TLS systems and targets. The study area was chosen to be 0.5 km by 0.5 km of the Centennial Park in Sydney, Australia as it contains a variety of different tree types. The data acquired using this system gave access to greater number of returns per laser shot. A phase shift method was used to obtain the measurements. Because processing of the TLS data is a demanding and time-consuming task, we
selected a small sample area covering only four different types of trees with different sizes. Table 1 shows the specifications of the scanner used in this experiment.
Figure 2. The processing result of TLS data
Figure 1. Study area for the TLS system: left) Ground Control Point (GCP) location, right) TLS and targets
Table 1. Leica HDS 6000 Performance specification Parameter
Specification
Measuring Type
Phase-Shift
Range
79m ambiguity interval 79m @90%: 50m @18% albedo
Data sampling rate
up to 5000pts/sec
Scan resolution Scanner field of view Single point Accuracy
The TLS data were classified and filtered according to the height information. The filtering was carried out manually to eliminate possible outliers. Figure 2 shows the initial classification results over the test area. Once the processing has been completed, the outcome was compared with the digital photos of actual trees, which were of different heights and shapes. The digital image and the corresponding TLS data are illustrated in Figure 3.
spot size: 3mm @ 50m (based on Gaussian definition) 360° (Horizontal, Maximum), 310° (Vertical, Maximum) Position= 10 mm to 50m range, distance= 5 mm @ 50m (90%)
Modelled surface precision
±2 mm at 25 m, ±4 mm at 50 m
Laser class
3R (IEC 60825-1) Figure 3. True Image vs. TLS Data result
4. Data Processing In this study the post-processing of the 3D point clouds was performed with Leica’s Cyclone 6.0. This software offers registration and geo-referencing of point clouds as well as multiple options for post-processing. Four GCPs were surveyed with Global Positioning System (GPS) receivers in a static mode over 30-minute observation sessions. The root mean squares errors (RMSE) of the geo-referencing were 0.003m horizontally and 0.004m vertically.
Finally, these trees were reconstructed in 3D using the crust algorithm based on Matlab software to perform each processing step of the procedure. Figure 4 compares the TLS points cloud with the 3D tree model which has been reconstructed using the crust algorithm. The total processing time was 5.271 seconds for Tree 1 and 2.379 seconds for Tree 3. These two samples were chosen as they varied greatly in shape and height.
Tree No1 No2 No3 No4 RMSE
Figure 4. 3D tree points cloud vs. 3D tree model
5. Data Analysis
No1 No2 No3 No4 RMSE
Tree No1 No2 No3 No4 RMSE
Table 2. Tree Crown Diameter Tree Total Crown Diameter Reference (m) TLS model (m) Difference (m) 2.34 2.29 0.05 2.44 2.48 0.03 3.59 3.54 0.05 2.91 2.97 -0.06 0.05
Table 3. Tree Height Tree Height Reference (m) TLS model (m) 7.22 7.57 11.12 11.35 4.82 4.69 10.61 10.41
The RMSE of total crown diameter, the tree height and the tree trunk size were 0.05 m, 0.22 m and 0.16 m, respectively, which demonstrates good comparison between the two sets of measurements. This method has shown to be effective for use in direct size measurement which provides a better reproduction of individual trees in 3D compared to conventional GIS software available on the market. Since it was hard to distinguish tree leaves from branches, they were not processed separately but as a whole, which in turn could have served as a possible source of error. 7. Conclusions
In analysing the 3D tree modelling accuracy, individual trees from the TLS data were compared to those from the field surveying. Three parameters were taken into account for the comparison: the total crown diameter, the tree height and the tree trunk size measured at the height of 1m from the ground (Chen et al., 1992, Wilhelm et al., 2000). Differences between the ground survey and the modelling by the crust algorithm for the tree parameters of crown diameter, height and trunk size measured at 1m from the ground are shown in Tables 2, 3 and 4 respectively. Since the crown diameter of a tree varies at different points, the average was used for comparison.
Tree
Table 4. Tree Trunk Size Tree Trunk Size (h=1m) Reference (m) TLS model (m) Difference (m) 1.01 1.29 -0.28 1.69 1.82 -0.13 0.68 0.85 -0.17 1.55 1.38 0.17 0.16
Difference (m) -0.35 -0.23 0.13 0.2 0.22
Today, researchers use the TLS modelling in many fields of study including civil engineering, building modelling, city planning, and 3D modelling of the forestry area, in particular. The aim of this paper was to apply 3D surface reconstruction algorithm to the TLS data for the reconstruction of individual tree surfaces. When the result of the TLS data was compared to that of the field survey, the total tree crown diameter the difference was 0.05m. The difference in tree height between the measurement taken using the TLS and the field survey was 0.22m while for the tree trunk size taken at 1m from the ground the difference was 0.16m. It was shown that this algorithm is useful for estimating the overall size of the trees. A possible source of error arises from the fact that tree leaves and branches cannot be clearly distinguished from one another. Future work should focus on delivering a more accurate classification system enabling users to separate the leaves from the branches and obtain a 3D reproduction that is more representative of the actual objects. Acknowledgements The authors would like to thank Sinclair Knight Mertz Pty Ltd for providing TLS data for this research. The authors also wish to thank Richard Sugandha who surveyed the TLS data. References [1] Aguilar M.A., Pozo J.L., Aguilar F.J., SanchezHermosilla J., Páez F.C. and Negreiros J., 2008, “3D surface modelling of tomato plants using close-range
photogrammetry” The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing, pp.139-144. [2] Amenta N., Bern M. and Kamvysselis M., 1998, “A new Voronoi-based surface reconstruction algorithm”, Siggraph pp. 415-421. [3] Amenta N. and Bern M., 1999, “Surface reconstruction by Voronoi filtering”, Discrete and Computational Geometry, 22 pp. 481-504. [4] Bae K.H. and Derek D.L., 2008, “A method for automated registration of unorganised point clouds” ISPRS Journal of Photogrammetry & Remote Sensing, Volume 63. Issue 1 pp. 36-54. [5] Chen J.M., and Black T.A., 1992, “Defining leaf area index for non-flat leaves” Agricultural and Forest Meteorology 57, pp. 1–12. [6] Hosoi F. and Omasa K., 2006, “Voxel-based 3D modeling of individual trees for estimating leaf area density using high-resolution portable scanning lidar” IEEE TRANSCTIONS ON GEOSCIENCE AND REMOTE SENSING, Vol. 44, No. 12 December 2006. [7] Ivanov N., Boissard P., Chapron, M. and Andrieu, B., 1995, “Computer stereo plotting for 3-D reconstruction of a maize Canopy” Agricultural and Forest Meteorology, 75(1), pp. 85-102. [8] Wang T., Dickinson J., Lang S., Khosla S. and Wu J., 2007, “Building a parametric 3D tomato plant model using laser scans and field data” In: Proceedings of the 18th IASTED International Conference MODELLING AND SIMULATION, Montreal, Quebec, Canada, unpaginated CD ROM. [9] Wilhelm W.W., Ruwe K. and Schlemmer M.R., 2000, “Comparisons of three Leaf Area Index Meters in a Corn Canopy”, Crop Science 40, pp. 1179-1183.
Despite the tremendous attention given to the 3D modelling of individual tree surfaces, it remains challenging to obtain an accurate 3D surface reconstruction of these structures. This paper aims to utilise a simple computer graphics algorithm in delivering an objective reproduction of individual tree surfaces by triangulating the surface by approximating an indicator function of the model. It is believed that the use of this algorithm will also enable us to reduce the total data processing time. 2. Crust Algorithm for 3D Surface Reconstruction Three-dimensional Voronoi diagrams and Delaunay triangulations are the basis of the crust algorithm which produces a set of triangles called the crust of sample points. All crust triangles are based on a Delaunay triangulation using sample points as the vertices of crust triangles (Amenta, 1998). The crust algorithm may fail when the number of samples is small. The three-dimensional crust is a set of triangles resembling the geometrical surface. If S is a set of r-samples from a smooth surface F, then for r 0.06 the crust of S contains a subset of triangles forming a mesh topologically equivalent to F and for every point on the crust lies within a distance 5r * d(p) of some point p on F, where d(p) is the distance from p to the medial axis (Amenta, 1998). 3. Data Acquisition The TLS data was collected on the 21st of May 2009 using a Leica HDS 6000 system. In Figure 1, the imagery on the left-hand side outlines the study area where red-coloured circles showing four different TLS Ground Control Points (GCPs). The imagery on the right-hand side illustrates the TLS systems and targets. The study area was chosen to be 0.5 km by 0.5 km of the Centennial Park in Sydney, Australia as it contains a variety of different tree types. The data acquired using this system gave access to greater number of returns per laser shot. A phase shift method was used to obtain the measurements. Because processing of the TLS data is a demanding and time-consuming task, we
selected a small sample area covering only four different types of trees with different sizes. Table 1 shows the specifications of the scanner used in this experiment.
Figure 2. The processing result of TLS data
Figure 1. Study area for the TLS system: left) Ground Control Point (GCP) location, right) TLS and targets
Table 1. Leica HDS 6000 Performance specification Parameter
Specification
Measuring Type
Phase-Shift
Range
79m ambiguity interval 79m @90%: 50m @18% albedo
Data sampling rate
up to 5000pts/sec
Scan resolution Scanner field of view Single point Accuracy
The TLS data were classified and filtered according to the height information. The filtering was carried out manually to eliminate possible outliers. Figure 2 shows the initial classification results over the test area. Once the processing has been completed, the outcome was compared with the digital photos of actual trees, which were of different heights and shapes. The digital image and the corresponding TLS data are illustrated in Figure 3.
spot size: 3mm @ 50m (based on Gaussian definition) 360° (Horizontal, Maximum), 310° (Vertical, Maximum) Position= 10 mm to 50m range, distance= 5 mm @ 50m (90%)
Modelled surface precision
±2 mm at 25 m, ±4 mm at 50 m
Laser class
3R (IEC 60825-1) Figure 3. True Image vs. TLS Data result
4. Data Processing In this study the post-processing of the 3D point clouds was performed with Leica’s Cyclone 6.0. This software offers registration and geo-referencing of point clouds as well as multiple options for post-processing. Four GCPs were surveyed with Global Positioning System (GPS) receivers in a static mode over 30-minute observation sessions. The root mean squares errors (RMSE) of the geo-referencing were 0.003m horizontally and 0.004m vertically.
Finally, these trees were reconstructed in 3D using the crust algorithm based on Matlab software to perform each processing step of the procedure. Figure 4 compares the TLS points cloud with the 3D tree model which has been reconstructed using the crust algorithm. The total processing time was 5.271 seconds for Tree 1 and 2.379 seconds for Tree 3. These two samples were chosen as they varied greatly in shape and height.
Tree No1 No2 No3 No4 RMSE
Figure 4. 3D tree points cloud vs. 3D tree model
5. Data Analysis
No1 No2 No3 No4 RMSE
Tree No1 No2 No3 No4 RMSE
Table 2. Tree Crown Diameter Tree Total Crown Diameter Reference (m) TLS model (m) Difference (m) 2.34 2.29 0.05 2.44 2.48 0.03 3.59 3.54 0.05 2.91 2.97 -0.06 0.05
Table 3. Tree Height Tree Height Reference (m) TLS model (m) 7.22 7.57 11.12 11.35 4.82 4.69 10.61 10.41
The RMSE of total crown diameter, the tree height and the tree trunk size were 0.05 m, 0.22 m and 0.16 m, respectively, which demonstrates good comparison between the two sets of measurements. This method has shown to be effective for use in direct size measurement which provides a better reproduction of individual trees in 3D compared to conventional GIS software available on the market. Since it was hard to distinguish tree leaves from branches, they were not processed separately but as a whole, which in turn could have served as a possible source of error. 7. Conclusions
In analysing the 3D tree modelling accuracy, individual trees from the TLS data were compared to those from the field surveying. Three parameters were taken into account for the comparison: the total crown diameter, the tree height and the tree trunk size measured at the height of 1m from the ground (Chen et al., 1992, Wilhelm et al., 2000). Differences between the ground survey and the modelling by the crust algorithm for the tree parameters of crown diameter, height and trunk size measured at 1m from the ground are shown in Tables 2, 3 and 4 respectively. Since the crown diameter of a tree varies at different points, the average was used for comparison.
Tree
Table 4. Tree Trunk Size Tree Trunk Size (h=1m) Reference (m) TLS model (m) Difference (m) 1.01 1.29 -0.28 1.69 1.82 -0.13 0.68 0.85 -0.17 1.55 1.38 0.17 0.16
Difference (m) -0.35 -0.23 0.13 0.2 0.22
Today, researchers use the TLS modelling in many fields of study including civil engineering, building modelling, city planning, and 3D modelling of the forestry area, in particular. The aim of this paper was to apply 3D surface reconstruction algorithm to the TLS data for the reconstruction of individual tree surfaces. When the result of the TLS data was compared to that of the field survey, the total tree crown diameter the difference was 0.05m. The difference in tree height between the measurement taken using the TLS and the field survey was 0.22m while for the tree trunk size taken at 1m from the ground the difference was 0.16m. It was shown that this algorithm is useful for estimating the overall size of the trees. A possible source of error arises from the fact that tree leaves and branches cannot be clearly distinguished from one another. Future work should focus on delivering a more accurate classification system enabling users to separate the leaves from the branches and obtain a 3D reproduction that is more representative of the actual objects. Acknowledgements The authors would like to thank Sinclair Knight Mertz Pty Ltd for providing TLS data for this research. The authors also wish to thank Richard Sugandha who surveyed the TLS data. References [1] Aguilar M.A., Pozo J.L., Aguilar F.J., SanchezHermosilla J., Páez F.C. and Negreiros J., 2008, “3D surface modelling of tomato plants using close-range
photogrammetry” The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing, pp.139-144. [2] Amenta N., Bern M. and Kamvysselis M., 1998, “A new Voronoi-based surface reconstruction algorithm”, Siggraph pp. 415-421. [3] Amenta N. and Bern M., 1999, “Surface reconstruction by Voronoi filtering”, Discrete and Computational Geometry, 22 pp. 481-504. [4] Bae K.H. and Derek D.L., 2008, “A method for automated registration of unorganised point clouds” ISPRS Journal of Photogrammetry & Remote Sensing, Volume 63. Issue 1 pp. 36-54. [5] Chen J.M., and Black T.A., 1992, “Defining leaf area index for non-flat leaves” Agricultural and Forest Meteorology 57, pp. 1–12. [6] Hosoi F. and Omasa K., 2006, “Voxel-based 3D modeling of individual trees for estimating leaf area density using high-resolution portable scanning lidar” IEEE TRANSCTIONS ON GEOSCIENCE AND REMOTE SENSING, Vol. 44, No. 12 December 2006. [7] Ivanov N., Boissard P., Chapron, M. and Andrieu, B., 1995, “Computer stereo plotting for 3-D reconstruction of a maize Canopy” Agricultural and Forest Meteorology, 75(1), pp. 85-102. [8] Wang T., Dickinson J., Lang S., Khosla S. and Wu J., 2007, “Building a parametric 3D tomato plant model using laser scans and field data” In: Proceedings of the 18th IASTED International Conference MODELLING AND SIMULATION, Montreal, Quebec, Canada, unpaginated CD ROM. [9] Wilhelm W.W., Ruwe K. and Schlemmer M.R., 2000, “Comparisons of three Leaf Area Index Meters in a Corn Canopy”, Crop Science 40, pp. 1179-1183.
