The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
methodology is discussed in section 3. Then, concluding 
remarks are mentioned in section 4. 
2. METHODOLOGY 
As abovementioned the proposed methodology consists of a 
sequence of four steps: ground/non-ground point separation; 
building hypothesis generation; segmentation of planar patches 
and intermediate boundary generation; and boundary 
refinement and 3D wire frame generation. In this section, 
detailed explanation and experimental results are presented for 
each step. 
2.1 Ground/non-ground Point Separation 
The developed methodology for ground/non-ground separation 
is based on the assumption that non-ground points cause 
occlusions under perspective projection. In a perspective 
projection, the top and bottom of a structure are projected as 
two different points. These points are spatially separated by a 
distance referred to as the relief displacement. This 
displacement takes place along a radial direction from the 
image space nadir point, and is the cause of occlusions in 
perspective imagery. In this work, the presence of occlusions is 
detected by sequentially checking the off-nadir angles of the 
lines of sight connecting the perspective center and the DSM 
points, along the radial direction starting from the object space 
nadir point (Habib et al., 2007). Several synthesized perspective 
centers with heights close to the maximum elevation of the 
entire study area are introduced to more thoroughly detect the 
points causing occlusions. Figure 1 illustrates the basic concept 
of detecting non-ground points along a profile using a 
synthesized perspective center. By scanning for occlusions from 
different radial directions with multiple synthesized perspective 
centers, ground points are well-identified from the DSM. The 
DSM used in the analysis is generated by resampling the 
irregular LiDAR point cloud to a regular grid, using the nearest 
neighbor method to increase computational speed. After 
removing the effects caused by the roughness of the terrain, the 
non-ground points and ground points can be separated from one 
another. For more detailed explanation and experimental 
verification of this novel ground/non-ground point classification 
technique, please refer to [Habib et al., 2008]. 
Figure 1. Basic concept of detecting non-ground points. 
In this paper, raw LiDAR point data over University of Calgary, 
Canada, is introduced. The study area includes building 
structures which are connected with other complex buildings. 
Figure 2 and 3 shows aerial photos and the LiDAR point cloud 
over the area of interest, respectively. 
Figure 2. Aerial photo over the area of interest. 
Figure 3. LiDAR points over the area of interest (colors are 
assigned according to their elevations). 
The ground/non-ground separation algorithm is applied to the 
LiDAR points over the area of interest. Figure 4 shows the 
points which are classified into ground and non-ground points. 
The points in blue and red indicate ground and non-ground 
points, respectively. 
Figure 4. Separated ground and non-ground points. 
2.2 Building Hypothesis Generation 
Once the LiDAR point cloud has been classified into ground 
and non-ground points, non-ground points are further analyzed