The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
458 
steep slopes and discontinuities. These differences are the result 
of the differing abilities of the algorithms to preserve 
discontinuities while detecting large and low objects. 
In this paper, a new approach for the automatic extraction of 
terrain points from LiDAR data is presented. The next section 
will briefly describe our approach. This discussion will be 
followed by the proposed methodology for extracting non 
ground points. The Experimental Results section presents the 
descriptions of our datasets and results. Finally, concluding 
remarks regarding the performance of the proposed technique, 
together with future research directions are summarized. 
2. METHODOLOGY 
In a perspective image, we can see relief displacement caused 
by the height of the corresponding object point above or below 
the datum. Relief displacement is directly proportional to the 
radial distance and the object height above the datum. However, 
relief displacement is inversely proportional to the flying height 
above the datum. A larger radial distance and a perspective 
center with a lower height can cause more occlusions in the 
image. The concept of our new approach is based on occlusion 
detection. Non-ground points can cause occlusions in 
perspective views. Therefore, if occlusions can be detected, and 
we can find out which points are causing the occlusions, then 
these points would be identified as non-ground points. In our 
approach, we generate a DSM grid from irregular LiDAR point 
cloud. Using this DSM, once the occlusions are detected using 
synthesized perspective centers, the points producing the 
occlusions are identified. After removing the effects of the 
roughness of the terrain, non-ground points and ground points 
can be separated from one another. Figure 1 summarizes the 
procedure. 
/ Irregular 
LiDAR Points 
/ Ground and Non 
ground points 
Figure 1. Flowchart of LiDAR data classification. 
2.1. DSM Generation 
A LiDAR point cloud is obtained as an irregularly spaced set of 
points. For most analytical processes, processing this irregular 
data format is time-consuming, and converting the points to a 
regular grid for analysis and visualization increases the 
efficiency. The pixel size has to be determined before 
resampling. Reducing information loss is important, as is 
keeping the redundancy at a minimum, while resampling. A 
very large ground sampling distance (GSD) for the resampled 
DSM will increase the information loss. However, the 
redundancy increases, as do the storage requirements, if the 
GSD is very small. To satisfy these requirements, the optimum 
GSD for resampling can be estimated to be equal to the average 
point density of the LiDAR data. In order to keep the edges 
from being blurred by some low pass filters, we use the nearest 
neighbor method for resampling. The elevation of each grid 
point is assigned the elevation of the closest original LiDAR 
point. If there is more than one point located in a pixel, we pick 
the one with the lowest height and assign its height to the pixel. 
2.2. Identification of the Points Causing Occlusion 
In this paper, the off-nadir angle to the line of sight will be 
denoted as the a angle, as in Figure 2. As we move away from 
the nadir point, the off-nadir angle a is supposed to increase 
(Habib et al., 2007). As long as the a angle increases while 
moving away from the nadir point, the DSM cells along the 
radial direction will be visible in the image in question. For 
example, points A and B are visible in Fig. 2 since their 
corresponding off-nadir angles increase as we move away from 
the nadir point. Occlusions, on the other hand, can be detected 
whenever there is an apparent decrease in the off-nadir angle a 
while proceeding away from the nadir point. This occlusion will 
persist until the off-nadir angle a exceeds the angle associated 
with the last visible point. In Fig. 2, we find that because 
Ot D < (X c , an occlusion is detected at point D. This occlusion 
will persist until point E, at which OC F > CC c . After an 
occlusion has been detected, the points causing the occlusion 
can be identified while tracing a path toward the nadir point. 
Figure 2 shows how the points causing an occlusion can be 
determined using the triangle composed of the last visible point, 
the first occluded point, and the perspective center. The non 
ground points can be traced until the off-nadir angle a is equal 
to the angle associated with the first occluded point. In Figure 2, 
an occlusion is detected because (X D < CC c . Point C is defined 
as the last visible point for this search, and point D is taken to 
be the first occluded point. A backward search for points 
causing occlusions can then be carried out. Those points with 
off-nadir angles larger than OC D are defined as points causing 
occlusion. For example, point B is taken to be a non-ground 
point since CC B > Oi D . The backward tracing stops when we 
find an off-nadir angle smaller than (X D . In Figure 2, the tracing 
stops at point A because Ci A < Ci D . 
In order to obtain a more complete list of the points causing 
occlusions, we need to enhance the procedure’s capability of 
detecting these points. Adjusting the locations of the 
synthesized projection centers relative to the DSM can 
maximize the introduced occlusions. If the elevation of the 
perspective center can be adjusted to be as close as possible to 
the height of the non-ground object, then the capability of 
detecting the points causing occlusions can be improved. A