ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002 
  
Figure 2-d. The Candidate Building Points 
(Building 2) 
For each building candidate point a window is constructed 
around the point. All candidate points in the window are used 
to fit a plane using Equation (1). For each candidate point 
nine positions were tried and the position with the minimum 
residual was selected. This helps to include corner points and 
edge points, where the candidate point is either in the 
window corner or on its edge. A linear least squares 
estimation approach (Mikhail, er. a]. 2001) is used to find the 
plane parameters at each point. If a statistic representing 
misclosure is small, the plane parameters at this point are 
used to vote in the 2D parameter space. Figure 3a and b 
show the parameter space for the two buildings previously 
illustrated. 
The window size and the parameter space cell size are based 
on the quality of the LIDAR data and the building 
characteristics. Poor quality data forces the cell size to be 
large in order to compensate for variation in the evaluated 
parameters of different points in the same plane. The 
relationship between points and their associated parameter 
cell is preserved for later use. The approach is similar to that 
used in Davies et. al. (1988) to extract straight lines. 
  
Figure 3-a. The Parameter Space (Building 1) 
A- 104 
  
Figure 3-b. The Parameter Space (Building 2) 
The parameter space is then searched to find peak cells. Cells 
with a high number of contributing points are identified and 
used as the building planes. The minimum number of points 
in the cell to be used as the threshold varies depending on the 
data quality, building size, and DEM resolution. Each cell is 
then given an identification number to identify this plane. 
Points that contribute to a certain plane are then categorized 
using their plane identification number. Plane regions are 
then extracted using the identification number. Each point in 
the DEM has its plane identification number. We used a 
region- growing algorithm (Jain, 1989) to connect points with 
the same identification number. Regions with small areas are 
then eliminated, while other regions are kept as the building 
roof regions. Holes in the roof regions are then filled. Figure 
4-a and b show the extracted roof regions. The roof regions 
are used to extract the roof boundaries in the next section. 
  
Figure 4-a. Extracted Roof Regions (Building 1)