xity of the 
ion of fore 
em, but is 
ross them. 
c freedom. 
of building 
dition and 
ces, where 
lidar. 
  
yrovided by 
Greenwich 
lidar DSM 
20 airborne 
n OSGB36 
4. The lidar 
responds to 
it per 3.2 x 
m 1.4 m to 
1er than the 
be found in 
rich LIDAR 
3 number of 
d over the 
t density of 
ly represent 
e formed in 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
Figure 2. Greenwich lidar DSM 
3. BUILDING DETECTION 
The complexity of building extraction process can be 
reduced by a large amount if the process can be focused on 
single building object. This section presents a building 
detection method to localize individual buildings by 
sequentially removing dominant urban features which are 
not relevant to buildings. 
3.1 Terrain detection 
A lidar filter, called recursive terrain fragmentation (RTF) 
filter, was developed to distinguish between on-terrain 
points and off-terrain ones from a cloud of lidar points. The 
RTF filter was implemented by employing a hypothesis-test 
optimization in different scales. This filter assumes that 
generic terrain surface is a mosaic of planar terrain surfaces. 
The entire lidar space, initially hypothesized as a single 
planar terrain surface, is recursively fragmented with small 
sub-regions until the coexistence of different terrain slopes 
cannot be found over all fragmented regions. More detailed 
description of the RTF filter can be found in Sohn & 
Dowman (2002). Figure 3(a) shows the on-terrain points 
detected by the RTF filter from figure 2. In this figure, some 
terrain segments which are not densely covered by the 
filtered on-terrain points show poor quality of the Greenwich 
lidar DSM. 
32 High-rise and low-rise object detection 
With the on-terrain points detected by the RTF filter, a 
DTM is generated. Then, outlying points with a height less 
than a pre-defined height threshold (4m) from the generated 
DTM are classified as “high-rise” features, otherwise as the 
“low-rise” ones (see figure 3(b)). 
3.3 Tree detection 
Since the “high-rise” feature class generally contains trees 
as well as buildings, *vegetation" points must be identified. 
The normalized difference vegetation indices (NDVI) are 
computed by a combination of red and near-infrared 
channels of Ikonos. When the “high-rise points" are back- 
projected onto the NDVI map, a small mask (5x5 size) is 
constructed around them, and “vegetation” points are 
identified if any masked pixel has the NDVI value larger 
than a threshold value (70.8) (see figure 3(c)). 
34 Building blob detection 
Isolating the building label points and making them into 
individual building objects is rather straightforward. Those 
points classified into the on-terrain, low-rise, and tree 
objects are together assigned non-building labels. Then, 
building points surrounded by the non-building labels, are 
grouped as isolated objects. As a result, 28 building “blobs” 
can be found from figure 3(d) after removing small “blobs” 
whose member points are less than 30 points. Further 
processing allows the individual building “blobs” to be 
bounded with rectangle polygons, and these polygons are 
then fed into the building description process, which will be 
discussed in the next section. 
  
    
    
    
d med re i ACUARIO 
(b) “high-rise” point detection result 
€ 
Rie Pe beth 
(c) after removing “vegetation” points 
| (d) building “blob” detection result 
Figure 3. Building detection results