The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
474 
1.2 meters across-track and 10 centimeters along-track spacing. 
The dataset is provided by the ISPRS Commission III Working 
group8 official web-site, and is available on-line at: 
http://isprs.ign.fr/packages/zone3/package3_en.htm 
An aerial image with 25 centimeters ground pixel size is also 
provided from the scene which is shown in Fig. 1. This image 
can be useful for visual comparisons. 
WKÊg 
Figure 1 - Aerial image of the study area 
3. IMPLEMENTATION 
The building detection system starts with a classification 
process which makes use of both FP and LP points. This 
classification divides the LIDAR points into “Rough” and 
“Smooth” classes. Of course as will be described later, many 
points within dense trees will be misclassified in “Smooth” 
point class. Then a simplified version of the so-called Sohn 
filter is used to extract on-terrain points from the points of 
“Smooth” class and the DTM is generated using these points. 
The normalized DSM can be computed by the DSM and DTM. 
Then a thresholding separates high-rise pixels from the nDSM. 
These pixels may belong either to building roofs or to dense 
vegetation covers. Then a slope thresholding applied on the 
slope map of the nDSM arranges the pixels of nDSM into either 
of the two classes of “Severely” and “Slightly” variable slope 
pixels. Finally building pixels are detected among the members 
of “Slightly Variable” class which simultaneously belong to the 
“High-rise” class. The whole procedure is described in details in 
the following subsections. 
3.1. DSM roughness analysis 
As mentioned before, in order to reduce the amount of 
calculations in the Sohn filter, our system tends to find the 
points belonging to “Rough” areas and filters them out. Such 
points in both FP and LP data have different heights due to the 
canopy penetration capability of laser pulse. So a simple way to 
detect these points is the subtraction of the heights of all points 
in last pulse return from corresponding points in first pulse 
return. The only problem is that the height differences from the 
first and last returns do not work for areas covered by dense 
trees where laser pulses cannot penetrate [Zhang et. al 2006]. 
This will cause many points of dense vegetated areas to remain 
among “Smooth” points. 
Often the points of first and last returns of laser do not 
necessarily have the same exact planar coordinates since the 
scan angle is not perpendicular to terrain. This case happens 
predominantly wherever the elevation changes abruptly like 
vegetated areas and near the walls of buildings. To tackle this 
problem we generate two Digital Surface Models (i.e. DSM) by 
interpolating FP and LP points individually. The height 
difference of corresponding pixels in these two models is stored 
in an image called the differential DSM (i.e. DDSM) image. 
The value of the pixels of DDSM is more wherever the pixels 
belong to vegetations or walls. 
A threshold equal to 15 centimeters is set to discriminate 
vegetation from other covers in the DDSM. Pixels with values 
more than the threshold are classified as “Rough” pixels and the 
rest of pixels will be assigned the “Smooth” label. The pixels of 
“Smooth” class then make a mask image (Fig. 2). Every LIDAR 
point which lies inside the mask should contribute in the 
generation of the Digital Terrain Model and hence these points 
are stored in an individual file labeled “Smooth points”. Fig. 2 
shows the classified DDSM on which the pixels of “Rough” 
class are assigned a green color, while yellow pixels represent 
the “Smooth” class. 
Vi 
Lei'll 
Figure 2 - The result of the classification of DDSM pixels into 
“Smooth” (light tone) and “Rough” (dark tone) 
3. 2. Filtering the LP data 
In order to generate the Digital Terrain Model from LIDAR data, 
a filtering process is implemented on the LP data. The result of 
this filtering is a set of points which lie on the terrain. A 
filtering method called the “Sohn filter” (G.Sohn, I.Dowman 
2002) -also called “Progressive TIN densification/ 
Regularization method” by some authors- is the basis of our 
filtering step. Their algorithm is based on a two-step progressive 
densification of a TIN; the Points in the TIN at the end of the 
densification are accepted as a representation of the bare earth, 
and the rest as object [Sithole 2005]. We have done our filtering 
based on a simplified version of their algorithm. The first step 
of densification in our filtering is somehow the same as Sohn’s. 
The only difference is that we select more than four points as 
initial on-terrain points. But we have made some simplifications 
in the second step, where we have ignored the MDL (i.e. 
Minimum Description Length) criterion. Our study area is 
almost a flat, smoothly sloped area with a few flat roofed 
buildings. Since there is no dominant topographic influence in 
the scene, investigating the MDL criterion is not a necessary 
task. That’s why we have made the aforementioned 
simplification. 
All the points inside the “Smooth points” file are the inputs to 
the filtering step. A set of initial on-terrain points including four 
points covering the study area, and a few points (three points in 
this case) at the middle of the scene are selected and the 
triangulation is triggered by them. The selection of these points 
is not a difficult task since they are members of the “Smooth” 
class of the DSM. Then lowest point in each triangle is found 
and added to the on-terrain points group and the triangulation is 
repeated again. This procedure is iterated until there is no point 
below any triangle. All the points of the last TIN are assigned 
an on-terrain label. 
The second step of densification starts with the final TIN made 
in the last step. A buffering space with a distance of 50 
centimeters is defined above each triangle. All of the points in 
the “Smooth points” file except for those used in the TIN are 
examined. Every point within the buffer is assigned an on-