




Automatic Detection and Elimination of Vegetation from 3D Laser Scan Data


Advanced Software Practical



Comparision of the Algorithms
Chainsaw Algorithm
 very fast
 works locally and provides good results even with slight uphill grades
 2D Tree is needed
2DFretsaw Algorithm
 good performance
 moderate recognition of objects
 little additional costs (only 2DTree is needed)
3DFretsaw Algorithm
 computationally intensive
 improved recognition of objects
 2DTree and 3DTree are needed
Performance Comparision
In the diagram above you see the results of performance measurements we did
on a laptop with a Pentium Core 2 Duo Processor (1.6 GHz).
We applied our three algorithms on two different data sets with three different patch (cell) sizes.
The times we called Init 2D and Init 3D give the time needed to build up the KDtrees (one tree for 2D and
two trees for 3D).
In the first scan, our 3DFretsawAlgorithm clearly outperforms the 2Dvariant. We guess here the data is sorted
in a better way, so that the 3Dalgorithm doesn't need to switch between branches as frequently as in the 2D variant.
2D Fretsaw Algorithm
vs.
3D Fretsaw Algorithm
In the pictures above deleted points are marked in margenta. As you can see, important objects (gravestones) are preserved by the 3D Fretsaw Algorithm.
Results
Scan from Crete  Doline
In the following scan from Crete, you can clearly see the benefit of our algorithm in the basement of the doline, where the scanner was positioned.
Before FretsawAlgorithm
left: point cloud
right: generated mesh from point cloud
After FretsawAlgorithm
left: point cloud
right: generated mesh from point cloud
Old Castle
In the following, you see a terrain scan inside a thick forest, where only the surface is of interest.
In this case the 2DFretsaw Algorithm is the first choice to delete the foilage.
Unmodified Point Cloud
After 2DFretsaw Algorithm
Resulting Mesh (colored)
Resulting Mesh (noncolored with diffuse shading)






