massive aerial LiDAR point clouds

Transcription

1 Visualization, storage, analysis and distribution of massive aerial LiDAR point clouds Gerwin de Haan and Hugo Ledoux Data Visualization group GIS technology group GeoWeb 2010, Vancouver July / 30

2 AHN 2 : A dataset covering totally the Netherlands compiled by the government (x, y, z) points raster (50cm grid available) at least 4 pts/m 2 > 150 billions points 2 / 30

3 AHN 2 : A tiled dataset 3 / 30

4 AHN 2 : A tiled dataset 3 / 30

5 File g37en1 08.xyz , , , , , , , , , , , , , , , , , , , , / 30

6 Our challenges when dealing with massive point clouds 1 Real-time visualisation 2 Storage + spatial analysis 3 Distribution of such datasets Can our tools/workflows scale to massive datasets? 5 / 30

7 Real-time Visualisation 6 / 30

8 Why visualise point clouds? Because it s a nice challenge! Because we need to: Perform visual inspection of raw measurements Gain insights from numbers Rendering Visualisation Interaction 7 / 30

9 From DEM terrain... (AHN2 DEM 0.5m grid) 8 / 30

10 ... to point cloud ( 2 millions points from AHN2 ) 9 / 30

11 Our tools of choice OpenSceneGraph most popular FOSS scene graph engine based on flight-simulator software (not GIS) OpenSceneGraph-based VRMeer software Python abstraction layers 10 / 30

12 Demonstration video 11 / 30

13 Visualisation pipeline overview Point Cloud Server Real-Time Visualization Client LiDAR Offline LOD generator Data Server LOD tiles network Database Pager Thread display Visualization Thread 12 / 30

14 Multi-resolution data struture L1 L1 L0 L0 L0 L0 L1 L2 L1 Ln 13 / 30

15 Multi-resolution data struture Split LOD0 LOD1 LOD2 Merge LOD3... LOD N 13 / 30

16 LOD effects 14 / 30

17 Interaction Stereoscopic display 3D interaction with space-mouse, Wii balance board interaction More info: PhD thesis Techniques and Architectures for 3D Interaction, Gerwin de Haan, TU Delft, / 30

18 Interaction 16 / 30

19 Analysis & Storage 17 / 30

20 Point clouds are often 2.5D surface LiDAR datasets are formed by scattered points in 3D space, which are the samples of a surface that can be projected on the horizontal plan. 18 / 30

21 Reconstruction of the surface Original LiDAR points 19 / 30

22 Reconstruction of the surface Raster representation 19 / 30

23 Reconstruction of the surface Raster representation 19 / 30

24 Reconstruction of the surface Original LiDAR points 19 / 30

25 Reconstruction of the surface TIN (with Delaunay triangles) 19 / 30

26 Reconstruction of the surface p TIN (with Delaunay triangles) 19 / 30

27 Reconstruction of the surface a f b c p e d TIN (with Delaunay triangles) 19 / 30

28 A star-based data structure Goes beyond the usual store points and edges/triangles Ideas come from data structures for compression of graphs p 20 / 30

29 A star-based data structure Goes beyond the usual store points and edges/triangles Ideas come from data structures for compression of graphs a f b c p e d 20 / 30

30 A star-based data structure Goes beyond the usual store points and edges/triangles Ideas come from data structures for compression of graphs a f b c p e d 20 / 30

31 A star-based data structure Goes beyond the usual store points and edges/triangles Ideas come from data structures for compression of graphs a f b c p e star(p) = abcdef d 20 / 30

32 Every star(v) is stored implicit triangles 21 / 30

33 Every star(v) is stored implicit triangles 21 / 30

34 Every star(v) is stored implicit triangles 21 / 30

35 Every star(v) is stored implicit triangles 21 / 30

36 Stars in a DBMS ID x y z star Advantages: 1 Only one table with id x y z star 2 No spatial index needed: fetching of triangles based on walking 3 Star column need not be filled ( Simple Features) 4 Local updates are possible (insertion and removals) 5 Ideas are readily extensible to 3D for storing manipulating tetrahedra 22 / 30

37 Point Location = Walking in the triangulation starting triangle p (Can be made efficient with some tricks [MSZ99]) 23 / 30

38 Range Queries: also uses the triangulation 24 / 30

39 One problem: how to create that DT in the first place? Figure from Martin Isenburg s presentation at GIScience 2006 [ILSS06] 25 / 30

40 Distribution of the Data 26 / 30

41 Velas3D: An online AHN 2 selector 27 / 30

42 Velas3D: An online AHN 2 selector 27 / 30

43 Velas3D: An online AHN 2 selector 28 / 30

44 Thanks for your attention Hugo Ledoux Gerwin de Haan 29 / 30

45 References Martin Isenburg, Yuanxin Liu, Jonathan Richard Shewchuk, and Jack Snoeyink. Streaming computation of Delaunay triangulations. ACM Transactions on Graphics, 25(3): , Ernst P. Mücke, Isaac Saias, and Binhai Zhu. Fast randomized point location without preprocessing in two- and three-dimensional Delaunay triangulations. Computational Geometry Theory and Applications, 12:63 83, / 30