idow 
ipport of the 
Engineering 
ammetry and 
998. Iterative 
Int. Conf. on 
ry 4-7. 
ne, J. Chris, 
John Wiley & 
servation and 
ew York. 
M., 1980. 
, Publishers, 
., 1999. Self- 
Varying and 
ial Journal of 
and Exterior 
on, NAS-99- 
ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision", Graz, 2002 
  
HEURISTIC FILTERING AND 3D FEATURE EXTRACTION FROM LIDAR DATA 
Abdullatif Alharthy, James Bethel 
School of Civil Engineering, Purdue University, 1284 Civil Engineering Building, West Lafayette, IN 47907 
alharthy@ecn.purdue.edu, bethel@ecn.purdue.edu 
KEY WORDS: LIDAR, feature extraction, building extraction, LIDAR data filtering, tree removal, plane fitting 
ABSTRACT: 
The need for a fast, efficient and low cost algorithm for extracting 3D features in urban areas is increasing. Consequently, research 
in feature extraction has intensified. In this paper we present a new technique to reconstruct buildings and other 3D features in 
urban areas using LIDAR data only. We have tried to show that dense LIDAR (Light detection and ranging) data is very suitable 
for 3D reconstruction of urban features such as buildings. This concept is based on local statistical interpretations of fitting 
surfaces over small windows of LIDAR derived points. The consistency of the data with surfaces determines how they will be 
modeled. Initially, the data has been filtered to remove extraneous objects such as trees and undesired small features. Then, 
boundaries will be extracted for each facet using statistical information from the surface fitting procedure, and using inferences 
about the dominant direction. Building features extracted from actual dense LIDAR collected over the Purdue campus are 
presented in the paper. 
1. INTRODUCTION 
With the availability of many sources of data such as 
conventional imagery, SAR imaging, IFSAR DEMs, and 
LIDAR DEMs, there are many avenues open to derive terrain 
and feature data in urban areas. Through much research, it has 
been shown that laser scanning data has the potential to 
support 3D feature extraction, especially if combined with 
other types of data such as 2D GIS ground plans (Maas, 1999; 
Brenner and Haala, 1999; Weidner and Fórstner, 1995). 
Despite the fact that LIDAR data is attractive in terms of cost 
per high quality data point, the quantity of the data makes a 
challenge for storage and display (Vosselman, 1999). 
Acquiring 3D object descriptions from such data is a difficult 
problem and many approaches have been tried to solve it. 
Several of them have succeeded with some limitations. The 
principle idea of this research is to detect and reconstruct 
buildings form laser altimetry data exclusively. 
The building detection procedure described here includes 
detecting and excluding other natural features such as trees. 
The type and quality of object description depends on the goal 
of the research. Our aim is to detect and reconstruct buildings 
in urban areas. Filtering to identify candidate buildings from 
other urban features was the first step in this work. Many 
segmentation techniques such as thresholding determined by 
histogram analysis, the use 2D GIS data, and multispectral 
inference have been tested together with LIDAR heights 
(Mass, 1999; Brunn and Weidner, 1997). Direct thresholding 
works by extracting the terrain surface using a filtering 
technique such as the morphological filter. Consequently, all 
objects above the terrain including buildings, trees and others 
will be detected. This approach should be followed by a 
refinement step to get the desired results. Using the second 
strategy, 2D GIS ground maps give the building footprints. 
However, the LIDAR data gives heights on the building roofs, 
and roofs can be larger than the ground footprints (in case the 
footprint represents the building structure rather than the 
overhanging eaves). Furthermore, most of the roof details are 
not shown in the ground plans. The classification based on the 
multispectral response is limited since it requires the 
availability of such data, and its accuracy can depend on the 
complexity of the scene. 
In this paper, besides the first and last return analysis, we used 
the local statistical variation as a key to discriminate buildings 
from other extraneous objects. Low variation indicates smooth 
surfaces and, on the other hand, high variation is an indication 
of inhomogeneous surfaces or, in other words, extraneous 
objects. Those extraneous objects such as trees have been 
detected and removed through an iterative procedure. The 
result of this filtering process is a modified DSM (digital 
surface model), which represents only terrain and buildings. 
Moreover, the DEM (digital elevation model) was extracted by 
applying a region growing segmentation technique to 
discriminate continuous surfaces from other objects and 
applying a local minimum filter on other regions. Finally, 
buildings were detected using the normalized DSM, and their 
primitive descriptors were extracted. Then, the two dominant 
directions for each building were computed and its polygon 
was constructed. A developed shape generalization procedure 
was applied to the extracted polygons and the 3D model of 
buildings was constructed. We have tested our approach on the 
data set that has been collected over the Purdue university 
campus in spring 2001 with an approximate density of one data 
point per square meter. 
2. DATA FILTERING 
Laser point clouds should be segmented in order to use them in 
3D building reconstruction in urban areas. The segmentation 
procedure is mandatory to differentiate among diverse objects 
in the scene. Extraneous objects such as trees, and any other 
object above the ground that does not belong to the building 
category should be detected and removed from the scene. The 
filtering process we used in this project to segment buildings 
from other undesired objects consists of two steps. The first