ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS’’, Bangkok, May 23-25, 2001 
3.1 Laser Range Data Imputing and Pre-Processing 
In the first step, we need translate the raw data got by the 
laser ranger to the file format that Softlmage|3D can read. 
Generally, there are two ways for this transformation. First one 
is generating Softlmage]3D graphic file based on raw laser 
range data directly. Second is translating DXF file (generated 
by MicroStation or AUTO-CAD) into Softlmage|3D. We have try 
these two methods for imputing laser range data and it will be 
useful for different laser range hardware (such as airborne or 
ground-type machines). Another important problem is to 
convert the raw data from their local coordinate system to the 
global one. The processing can be realized by making C 
program for 3D coordinate transformation based on given 
parameters. 
3.2 MM Filtering and 3D object segmentation 
For the purposes of range image filtering and related object 
segmentation, here we used Mathematical Morphology (MM) 
based approaches (Haralick et.al, 1987; chem., 1991). 
Generally, MM operators (such as dilation, erosion, opening, 
closing, hit or miss, thinning...) can be described as a kind of 
combination of shift and logic operations. Shifting operations 
are controlled by the given structuring elements (SEs) whose 
size, shape and orientation can be changed by the different 
applications. Different MM operators according for the different 
purposes organize logic operations. 
MM filtering is one of the most successful tools for MM 
applications. For range image processing, we used the opening 
filter to remove dirty voxels and small-connected volumes. We 
also used the closing filter to fill the small holes within surfaces 
and to link short gaps among objects. Here, the key problem is 
how to select suitable SEs, since the spatial features in range 
images are very dense in most cases, the direction-oriented 
SEs are suitable for many applications. How to select the 
parameters and algorithms for segmentation processing is 
highly dependent on different applications. Generally, before 
the real processing of whole large images we can process 
several selected typical small testing areas to find the suitable 
parameters and optimal processing procedures. We also can 
generate simple knowledge bases based on these selected 
parameters and procedures using for other range image 
processing. 
3D object segmentation is also based on open processing. In 
this case, the size of SEs should be a little larger than the 
segmented objects, the shape of SEs can be a simple convex 
object, such as a cubic box in 3D space or a rectangular area 
in the 2D horizontal plain. The basic idea of MM based object 
segmentation is firstly filtering all the parts smaller than the 
given SEs, then segmenting the objects by the logic difference 
operations between the original 3D data set and the filtered 3D 
data set. Since MM based filtering with a small SE in the first 
step processing also will damage the detail object features, a 
feature recovering is better to be added in the object segment 
procedure. The feature-recovering algorithm is based on the 
conditional dilation operations, in which the segmented object 
parts sever as the dilation seeds and the original 3D data set 
serves as the masking field for limiting dilated ranges. Weidner 
and Foerstner (1995) have also used the similar MM based 
methods for filtering and segmentation of 3D spatial objects 
from small scale DSMs. 
In large-scale laser range image processing, we think it is 
better to add the feature recovering procedures during 
segmentation of spatial objects for protecting the detail object 
structures. When we integrally using the 2D map data for range 
image processing, a feature protected filtering based on the 
conditional masking volumes generated by the 2D boundary 
lines of building or roads also can be realized. Figure 4 shows 
a simple procedure of MM filtering and object segmentation. 
Figure 5 and Figure 6 are MM Filtering and H result Filtering 
result of Kyoto train way station area. 
Filier 
¡y 
Open 
|f| 
Difference (p 
Final 
<3 
Recover 
<£] 
9# 
Figure4 3D MM filtering and object segmentation in a 3D space 
Figure5 MM Filtering 
Figure6 H filtering result 
3.3 Semi-automated Feature Extraction 
Based on the extracted ground truth data from 2D existing 
maps/images, we can generate the parametric or prismatic 
building and road models with the unknown heights firstly. In 
this way, we can also generate a little detail object features, 
such as the roof structures of buildings by using the vectored 
feature lines. Then we can estimate the height parameters of 
generated spatial objects based on the processed laser range 
images, which can be thought as a kind of model based 
matching between the pre-interpreted models (generated from 
2D vectors) and objected data sets (processed range images). 
The matching operation will cause four kinds of results as 
described following: