ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision‘, Graz, 2002 
recognition is then achieved based on image features that 
include points, edges, and lines. Selection of type and level 
of feature extraction is related to the desired accuracy and 
method of reconstruction. Also straight lines are usually a 
basic characteristic of conventional buildings with regular 
edges. Buildings are then reconstructed from the listed 
features. 
In Wang (2000) Building Extraction from a high quality 
terrain surface is presented. The approach takes terrain 
surface data as input and goes through edge detection, edge 
classification, building point extraction, TIN model 
generation, and building reconstruction to extract and 
reconstruct buildings and building related information. For 
building detection, the presented algorithm detects edges 
from the surface data and classifies edges to distinguish 
building edges from other edges based on their geometry and 
shapes, including orthogonality, parallelism, circularity and 
symmetry. The classified building edges are then used as 
boundaries to extract building points and TIN models are 
generated with the extracted points. Each building has its 
own TIN model and its surfaces are derived from the TIN 
model. 
In Zhao and Trinder (2000) building extraction from aerial 
images and DEM is presented. In this research, a building is 
modeled as a polyhedron, comprising planes that are 
connected to form a solid volume. The intersections of 
adjacent planes are straight lines. The polyhedron has a set of 
attributes describing its geometry, radiometry, texture, 
topology, and context. From this model in order to address 
the complexity of the problem, the system consists of three 
parts: building detection, building segment extraction, and 
3D segment matching and building modeling. The detection 
process starts with segmentation of the DSM (Digital Surface 
Model) to derive regions of interest (ROT) that have high 
expectation of representing individual buildings. Texture and 
shadow information are extracted and used to refine and 
verify the ROI. Buildings are constructed in a bottom-up 
approach. Primitive linear features are first derived, and 
relevant building polygons are extracted by grouping and 
filtering these primitive features within individual building 
regions. 3D lines are then generated by feature matching of 
these segments. Based on the matched lines, buildings are 
reconstructed by piecewise plane formation and plane 
intersection. 
3. EXTRACTING ROOF REGIONS 
In this section the process of extracting the building roof 
regions is described. The first step is to find the building 
candidate points; this is done by convolving the LIDAR 
DEM with a minimum filter. The second step is to extract the 
roof planes from the LIDAR DEM. This is done by voting in 
a plane parameter space and finding the cells in the 
parameter space with large numbers of points. LIDAR DEM 
points are then classified based on the plane to which they 
contribute. A region-growing algorithm is then used to 
complete the roof region extraction. 
A minimum filter is first used for the process of finding 
building candidate points. First the DEM is convolved with 
the filter. The second step is to calculate the difference in the 
elevations between the original DEM and its filtered version. 
The differences in the elevations are used to select the 
candidate building points. All points with a height of 5.0 
meters or more above the surrounding terrain are classified 
as building points. Figure 2-a,b,c, and d show the LIDAR 
DEM s and the candidate building points for two buildings. 
A - 103 
The next step is to vote in the plane parameter space. The 
plane equation is presented by Equation (1). The two 
parameters a, b, represent two slopes in the X and Y 
directions and the parameter c is an offset parameter. In this 
research, the DEM grid was oriented nearly parallel with the 
building orientation and only one of the slopes, at most, was 
significantly different from zero. This allowed reduction of 
parametric space from 3 to 2. We used a 2D parameter space 
for the plane detection step after defining the main direction 
of the building, i.e. one of the two slope parameters is pre- 
selected to be zero. 
Z=aX+bY +c 
  
Figure 2-a. The Elevation Shaded LIDAR DEM 
(Building 1) 
  
Figure 2-b. The Candidate Building Points 
(Building 1) 
      
  
Figure 2-c. The Elevation Shaded LIDAR DEM 
(Building 2)