ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision‘, Graz, 2002 
  
  
Figure 6 A test area including a road, street lamps, an 
underpass and a small vegetation area with small pine 
trees. Top: raw laser data. Bottom: the estimated 
ground surface 
80r 
  
  
  
| —— | Optimized mesh 
79 | : 5 Segmented raw data 
| 
78 } 
77+ 
76 : 
75 À s 4 j Y 
74 
73H i 
faa pala, “es 
EM » 
72 Pt i 
Ta -— 
71 i i L ii 1 i i 
1.4868 1.4868 1.4868 1.4868 1.4868 1.4868 1.4868 1.4868 1.4868 1.4868 
x 10° 
Figure 7 Estimated ground surface for a single laser 
radar swath. From left: the road, a ditch and a slope 
with trees. 
5. DISCUSSION 
When the surface estimation is done it is possible to add a 
classification step. By comparing the raw data set with the 
model, the points close enough to this surface are labeled as 
ground points, Figure 8. Then by picking out the ground 
points from the data set, a TIN is built using the original 
values from the data set. 
When running the algorithm on more sparse data sets some 
problems occur due to the re sampling of the data into a 
regular grid. When the re sampling is done the true X,Y- 
values of the points are lost. The height value of the point is 
moved to the centre of the nearest mesh and it is not good if 
this is outside the original footprint of the laser. On the other 
hand, if the data is re sampled into a fine grid most of 
A - 
  
Raw Data 
Resample 
Raw Data 
Sampled 
Grid 
Optimize Segmented 
Active Contour 
Ground Compare with 
Estimation 
Raw Data 
  
Figure 8.The classification chain of the laser scanner 
data 
the grid points lack data, this makes the algorithm slow. It is 
preferred that the grid points are approximately of the same 
number as the raw data points. In the data sets the algorithm 
was designed for, the laser footprint is about 0.3-0.5 m. The 
grid size has been set to 0.25 m. This gives a maximum error 
of displacement in X,Y of a laser point to less than 18 cm. 
One way to handle the displacement of points would be to 
adjust the algorithm to work with a TIN model instead of a 
mesh. The most straight forward method would be to let the 
model have the same number of vertices as the number of 
laser data points. A better way is probably to let the model 
have a variable number of control points, adding vertices 
near edges. 
REFERENCES 
Ahlberg S. Elmqvist M.,Hermansson P.,acobsson J., 
Persson À. and Sóderman U.,2001. Synthetic Environments 
and Sensor Simulation — Progress Reports 2001. FOI-R— 
0292—SE, Linkóping, Sweden 
Persson À.,2001.Extraction of Individual Trees Using Laser 
Radar Data. FOI-R—0236— SE, Linkóping, Sweden 
Axelsson, P., 1999, Processing of laser scanner data — 
algorithms and | applications, | ISPRS Journal of 
Photogrammetry & Remote Sensing 54 (2). 
Kass M., A. Witkin, and Terzopoulos D. /998.Snakes: active 
contour models, Int. J. of Computer Vision, 1:321-331. 
Pfeifer N.,Kóstli A., Kraus K.,/nterpolation of Laser Scanner 
Data — Implementation and first results, Vienna University 
of Technology, Austria, 1998 
TerraScan, TerraScan for  microStation user's guide, 
TerraSolid Ltd, 1999