ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision“, Graz, 2002 
  
selected as the optimized one, and its on-terrain points are 
stored in the “on-terrain point stack”. This process continues 
until the upward divide-and-conquer triangulation is performed 
over all models stored in the “current model stack” (see Figure 
4). 
5. TEST DATA & RESULTS 
We tested our suggested filtering technique with several 
different LIDAR dataset. Figure 8 (a) shows a test area located 
in east London with an Ikonos panchromatic imagery and the 
off-terrain points are extracted by our filtering algorithm from a 
LIDAR data, which was collected over the same area by the 
Optech 1020 sensor with 3 metre planimetric resolution (Figure 
8 (b)). This area was selected since it contains a “good” mixture 
of different features and slopes, i.e., residential area, flat grass, 
knolls, forest and hills; it is suitable to validate how this 
filtering technique is continuously adaptive to terrain surface 
variations, especially for gently sloped terrain. Although overall 
the terrain is not flat and there are several gentle hills, our 
technique clearly extracted off-terrain points, while any points 
on the hills are not labelled as the off-terrain (see middle of the 
bottom in Figure 8 (b)). 
Figure 9 (a) shows the Shrewsbury dataset in UK, which was 
acquired by the Optec 2033 sensor with 2 metre post spacing. 
As a result of the terrain surface reconstructed by our filter, 
Figure 9 (b) shows how the definition of terrain surface 
established in our filtering technique works in order to deal with 
terrain surface variations. In this example, one can see that our 
filtering algorithm recognized a railway embankment in the 
middle of figure as on-terrain points even though it has 
relatively steep slope (about 21?). This result is reasonable; if 
LIDAR points are located consecutively along the side of the 
railway embankment and one cannot observe the “emptiness” 
within the “buffer space” generated between the railway 
embankment and neighboring meadow, our terrain fragment 
process continues within that area and finally, the railway 
embankment is recognized as on-terrain points. For the same 
reason, but as a different result, a railway bridge located along 
the railway dam is detected as off-terrain points (see bottom of 
right side in Figure (b)). 
Another result processed using the sub-area of the OEEPE data 
set of Vaihingen with the density of 0.18 points per square 
metre is shown in Figure 10. In this result, houses and a group 
of trees are removed, but geomorphologic features are well 
preserved after applying our filter. 
Even though we used different test data set in terms of 
resolution and terrain type, we fixed our parameters as 1 metre 
for &, 0.1 and 45 for a and f respectively in Eq. (13), but the 
results are robust for the parameter settings. 
6. CONCLUSIONS AND FUTURE WORK 
We have shown that by explicitly selecting the criterion to 
differentiate on-terrain points from off-terrain ones, a LIDAR 
filtering technique, which is continuously adaptive to terrain 
surface variations, may be developed. This method aims to 
recursively fragment the entire LIDAR data domain into a set of 
piecewise planar surface models in order to make underlying 
terrain slope variations regularized into homogeneous plane 
terrain. To this end, two characteristics of plane terrain surface 
are defined; i) there is an empty "buffer space" in which any 
LIDAR point cannot be located over plane terrain, and ii) a 
"terrain polarity" made of a contextual information of on- and 
off-terrain points is augmented when it is measured from plane 
terrain. These characteristics are estimated over local areas 
reconstructed by a hypothesized planar surface model. By this 
means, our terrain reconstruction process is recursively 
triggered and an optimised planar model is selected. Since only 
one criterion for this method is explicitly required, our method 
can easily reflect the user requirement for the generic purpose of 
LIDAR filtering. 
Although our algorithm is not optimized yet in terms of the 
computational speed, it demonstrated promising results of the 
terrain surface reconstructed using real LIDAR data. Based 
upon this result, our future work will seek to classify building 
and tree objects from the off-terrain points. This will enable an 
object to be classified as a building and it will serve as an 
efficient tool for building detection and model generation. 
7. REFERECES 
Axelsson, P., 2000. DEM generation from laser scanner data 
using adaptive TIN models. In: The International Archives of 
the Photogrammetry, Remote Sensing and Spatial Information 
Sciences, Annapolis, MD, Vol. XXXIII, Part B3/1, pp. 119-126. 
Baillard, C. and Maitre, H., 1999. 3-D reconstruction of urban 
scenes from aerial stereo imagery: a focusing strategy. 
Computer Vision and Image Understanding, 76(3), pp.244-258. 
Cham, T. J. and Cipolla, R., 1999. Automated B-spline curve 
representation incorporation MDL and error-minimizing control 
point insertion strategies. [EEE Transactions on Pattern 
Analysis and Machine Intelligence, 21(1), pp. 49-53. 
Cobby, D. M., Mason, D. C. and Davenport I. J., 2001. Image 
processing of airborne scanning laser altimetry data for 
improved river flood modelling. ISPRS Journal of 
Photogrammetry & Remote Sensing, 56(2001), pp. 121-138. 
Flood, M., 2001. LIDAR activities and research priorities in the 
commercial sector. In: The International Archives of the 
Photogrammetry, Remote Sensing and Spatial Information 
Sciences, Annapolis, MD, Vol. XXXIV, Part 3/WA, pp. 3-7. 
Li, M., 1993. Minimum description length based 2-D shape 
description. In: Proceedings of 4" International Conference on 
Computer Vision, Berlin, Germany, pp. 512-517. 
Li, S. Z, 2001. Markov Random Field Modeling in Image 
Analysis. Springer-Verlag, Tokyo, pp. 3-8. 
Pfeifer, N. and Kraus, K., 1998. Interpolation and filtering of 
laser scanner data — implementation and first results. In: The 
International Archives of the Photogrammetry, Remote Sensing 
and Spatial Information Sciences, Columbus, Vol. XXXII, Part 
3/1, pp. 153-159. 
Rissanen, J., 1984. Universal coding, information, prediction, 
and estimation. IEEE Transaction of Information Theory, 30(4), 
pp. 629-636. 
Vosselman, G., 2000. Slope based filtering of laser altimetry 
data. In: The International Archives of the Photogrammetry, 
Remote Sensing and Spatial Information Sciences, Annapolis, 
MD, Vol. XXXIII, Part 3/2, pp. 935-942. 
A - 343