XXIX-B3, 2012 
> was applied to the 
surface variation and 
10.10m, respectively, 
for verification of 
lly outlined building 
eference wall facades 
s and completeness. 
1gth of the extracted 
| of all extracted lines. 
igth of the extracted 
igth of the reference 
cm which we believe 
le in this research. 
  
2 
0.65 
0.32 
s of verification 
  
  
  
  
  
10ds described in the 
oud of the Enschede 
s were extracted and 
ian be seen that the 
. as planar segments 
s non-planar (shown 
nschede is quite flat, 
several large planar 
in Figure 6. On the 
) while the 3D wall 
exist in wall facades. 
5. It was also noticed 
>getation area. These 
1e to their sizes are 
  
schede site. 
ed in white) 
  
Enschede data. 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
Finally, the wall facades were modelled using the information 
of building roof structures. The roof segments, eaves and roof 
ridges were explored to determine the wall extent and corners. 
Several examples are given in Figure 7. The initial wall outlines 
in Figure 7(b) were derived by detected wall points. Initial 
outlines were incomplete and incorrect. After modelling using 
building roof structure information, the wall boundary and 
corners were determined in Figure 7(c). The 3D models of the 
full buildings are shown in Figure 7(d). 
  
   
    
  
  
  
       
  
  
  
  
  
  
(a) (b) (c) (d) 
Figure 7. Detection of cubic buildings. (a) point cloud of 
original building; (b) detected wall outlines; (c) 
constraint wall outlines; (d) wall facade cubes 
  
Figure 8. Result of Melbourne campus site. (a) raw data, (b) 
detected planar points (red) and non-planar points 
(green), (c) generated segments, (d) reconstructed 
walls. 
The wall reconstruction performance for the Melbourne data is 
shown in Figure 8. The algorithms worked equally well, even 
considering that the point density was lower than in the 
Enschede data. Once again, the planar points (roof points, 
terrain points) and non-planar points (vegetation points, roof 
ridge points) were successfully detected, as indicated in Figure 
8(b). Planar segments, including roof façades and ground 
surfaces were derived by region growing using the planar seed 
points, as seen in Figure 8(c). Although only a small number of 
walls were illuminated by the LIDAR sensor due to the flight 
pattern, these walls were successfully extracted. An example of 
reconstructed walls is shown in Figure 8(d). 
5. DISCUSSIONS AND CONCLUSIONS 
This paper has presented a methodology for automated 
reconstruction of building walls from airborne LIDAR data. All 
procedures have been detailed, including point cloud 
segmentation and classification, wall reconstruction and 
modelling. The developed approaches have been tested using 
different datasets. Experimental results are presented. 
Segmentation plays a critical role in point cloud processing, 
particularly for object reconstruction. To achieve high quality 
segmentation, new approaches to search range determination 
and seed point selection have been proposed and implemented. 
Adaptive determination of search range can efficiently 
accommodate varying point cloud densities. Results show that 
PCA is an effective method to select planar points for 
segmentation. Thus, non-planar points, such as vegetation 
points, can be avoided from beginning. In both test sites, in 
Europe and Australia, all the roof segments, wall segments and 
planar ground segments were correctly extracted and modelled 
from the LIDAR point cloud, even though the point density was 
very different in each case. Thus, the developed segmentation 
method can be also used for roof reconstruction and terrain 
extraction. This method may also be applicable for tree 
detection upon further refinement. 
The experiments conducted have also shown that the wall plane 
can be determined from LIDAR points. However, LIDAR 
points alone are not sufficient to decide the wall boundaries. 
The extent and corners of extracted wall planes can be 
reconstructed with geometrical and topological relations 
between the wall and the roof structures. This modelling process 
proved to be powerful. Verification of correctness and 
completeness is preformed. Even though correctness is 
relatively higher than completeness, both are low due to point 
distribution. 
The reconstructed walls together with the 3D roofs generate 
complete 3D building models. Unfortunately, many walls 
cannot be reconstructed from the LIDAR point cloud since they 
are not ‘seen’ by the sensor. With the decreasing cost of 
airborne LIDAR, oblique scanning for dense wall point cloud 
coverage may well be more practical in the future. 
The sensitive of parameter setting and accuracy of segmentation 
result should be further investigated. Future research will refine 
the method for wall reconstruction and in general for building 
reconstruction. For instance, the current method reconstructs 
roofs and walls separately. New approaches may be researched 
to more efficiently explore the inherent relationships between 
different parts of a building so as to generate comprehensive 
building models with simultaneous roof and wall extraction and 
modelling.