Beijing 2008 
AUTOMATIC CLASSIFICATION OF LIDAR DATA INTO GROUND AND NON 
GROUND POINTS 
Yu-Chuan Chang a , Ayman F. Habib a , Dong Cheon Lee* 5 , and Jae-Hong Yom* 5 
department of Geomatics Engineering, University of Calgary, Calgary, 2500 University Drive NW, Alberta, 
Canada T2N 1N4 - (ycchang, habib)@geomatics.ucalgary.ca 
^Department of Geo-Informatics, Sejong University, Seoul, South Korea - (dclee, jhyom)@sejong.ac.kr 
Commission: WG IV/3 
KEY WORDS: LiDAR, DEM/DTM Extraction, Photogrammetry, Classification, Laser Scanning, Point Cloud 
ABSTRACT: 
Recently, automatic object extraction from Light Detection And Ranging (LiDAR) data has attracted great attention. The level of 
detail and the quality of the collected point cloud motivated the research community to investigate the possibility of automatic object 
extraction from such data. Prior accurate knowledge of terrain information is usually essential for the data to be usable in further 
processing, such as feature extraction, and to obtain better object detection results. In this paper, a new strategy for automatic terrain 
extraction from LiDAR data is presented. The proposed strategy is based on the fact that sudden elevation changes, which usually 
correspond to non-ground objects, will cause relief displacements in perspective views. The introduced relief displacements will occlude 
neighboring ground points. A Digital Surface Model (DSM) is first generated by resampling the irregular LiDAR point clouds to a 
regular grid. By using synthesized projection centers located above the DSM and analyzing the visibility maps in perspective images, we 
can classify the DSM into non-ground and ground hypotheses. Surface roughness and inherent noise in the point cloud will lead to some 
false hypotheses. By using a novel algorithm which combines plane fitting and statistical filtering to remove these false hypotheses, non 
ground and ground points can be separated. The algorithm has been tested using both simulated and real datasets. The results have 
demonstrated that our approach can perform well with highly complex data from an urban area. In a comparison with the results obtained 
with TerraScan software, our algorithm showed the capability of producing better results while being less sensitive to used parameters. 
1. INTRODUCTION 
LiDAR technology has been demonstrated in recent years to be 
a prominent technique for the acquisition of highly dense and 
accurate information for physical surfaces. As LiDAR is a non- 
selective mapping method, the acquired data consists of a point 
cloud that includes bare-ground and non-ground objects such as 
trees and buildings. Methods of removing non-ground points, 
also referred to as filtering techniques, have been the focus of 
many researchers. Many applications, for example, the 
generation of contour lines for topographic maps, road 
engineering projects, and the delineation of flooding zones, 
among others, require the generation of a DTM from the ground 
points. A DTM can be produced by resampling those extracted 
ground points from LiDAR data. The filtering step is also 
essential for the data to be usable in further processing, such as 
in feature extraction. Building detection and reconstruction 
procedures for the generation of 3D city models can be 
facilitated by first detecting the non-ground points. The feature 
extraction and modeling procedures are also beneficial to 
applications such as change detection and database updating. 
To satisfy the needs of these applications, the research 
community has been developing several techniques for the 
classification of LiDAR data. The first group of methods that 
can be identified in the literature are based on mathematical 
morphology. A method related to the erosion operator was 
proposed by Vosselman (2000). In this method, the acceptable 
height difference between two points is explicitly defined as a 
function of the distance between the points. Morphological 
filters have some drawbacks when certain features, such as large 
buildings and dense forest canopy, are involved. In such cases, a 
window size that is too small could be including only building 
points, thereby classifying them as ground. However, a larger 
window size can potentially chop off hills that have a significant 
slope. Strategies such as the use of multiple window sizes, as 
proposed by Kilian et al. (1996), and the one developed by 
Zhang et al. (2003), which gradually increases the window size, 
might help in overcoming these problems. However, the success 
of these types of filters is strongly dependent on the selection of 
the discriminant function parameters. The second group of 
filters are based on the progressive densification of a TIN 
(Triangulated Irregular Network). In Axelsson (2000), ground 
points are classified by iteratively building a triangulated 
surface model. The third group of methods are based on linear 
prediction and hierarchic robust interpolation (Kraus and Pfeifer, 
2001). The approach is based on a surface model defined for the 
entire point set that iteratively approaches the ground surface. 
However, these two groups of methodologies cannot handle the 
surface with low and complex objects very well, as reported by 
Sithole and Vosselman (2004). 
Approaches that rely on segmentation are also found in the 
literature. Jacobsen and Lohmann (2003) developed a method 
that first segments the data and then classifies the segments as 
either ground segments or off-terrain segments, based on 
neighborhood height differences. When dealing with large areas, 
segmentation techniques require expensive computation for 
processing. Other filtering algorithms are also described by 
Elmqvist et al. (2001), and Brovelli et al. (2002), among others. 
A detailed comparison of some filters is provided in Sithole and 
Vosselman (2004). The experimental study conducted shows 
that in flat and uncomplicated landscapes, all the algorithms 
give satisfactory results. However, significant differences in the 
accuracy of these methods appear when landscapes containing 
457