tanbul 2004
g Data
China,
eying
ne and high
>, especially
canned data.
h a process,
ital study is
yEye system
1 Optech in
| E. Chance,
Airborne
n in surveys
of coastlines
areas [Kraus
| capture are
a capture, it
le-scanning.
serials from
nple vehicle
001], which
as, is mainly
uildings and
am of China
ehicle-borne
e Image" or
points. The
:h-resolution
and ground
(tract feature
constructing
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part BS. Istanbul 2004
high-quality DEM directly [Weidner 1996, Brunn 1997].
Because the range image consists of discrete points, without
topological relation, boundary attribute and object feature, it
brings about the uncertainty of extracted object. The key to the
problem focuses on how to distinguish different objects and
construct model respectively. Manandhar (2001) has classified
scanning points according to the spatial distribution feature of
laser points in each scan line. Although the data processing is
complex, it can separate buildings, roads and trees etc. Li (2003)
researched linear building feature extraction from range images.
But all those are limited in extracting only one lateral feature of
buildings. Because hitherto there are no matured feasible
methods of segmentation and feature extraction from range
image, current laser-scanning systems are all integrated with
CCD or similar image acquisition devices. The range images
are mainly used as a supplement of photogrammetry or be
constructed high-quality DEM/DSM, and the CCD image data
for image segmentation and feature extraction [Ackermann
1999]. This collaborative mode has the character with the high
cost of time, the large quantities of data-storage and the
complex processing and integration of multi-source data.
Segmentation is the base of identification, location and
modeling of objects. This paper mainly researches the
technology and method of segmentation and extraction from
range images captured by vehicle-borne laser scanning system.
2. WORKFLOW OF OBJECT ACQUISITION AND
SEGMENTATION
2. 1 Acquisition Of Range Image
Vehicle-borne laser scanning system realizes the
combination of multi-sensors such as 2D Laser Scanner, GPS,
attitude-measure device (IMU, INS or multi-antenna GPS) and
Odometer onto the automobile platform. Under the control of
the vehicle-borne computer, the automobile runs along building
at the normal speed (Fig.1), the surface geometric information
of the objects is acquired in the real-time. In the system, GPS
gives the precise 3D position of the LS in the space,
attitude-measure device gives the attitude parameter of
the LS in the space, and laser scanner precisely
determines the distance from the scanner center to the
object at high frequency. According to geometry theory,
the 3D coordinates of the sampling points can be calculated to
%
935
form points cloud easily and quickly.
A
4
%
N
+ N
NS d : Trajectory \
Figure 1: The principle of scanning a building object and other
visible objects around the building
2.2 Principles and Workflow Of Data Processing
Image segmentation is one kind of fundamental technology of
computer vision and image analysis. To 2D grey image, the
procedure for image segmentation is to classify the image into
distinct groups and to extract the interested groups or objects
from image. From 1960s plenty of researches have been done
on grey image segmentation and formed about 1000 kinds of
segmentation algorithms [Pal 1993]. Most of algorithms are
aim at special problem and there is not suitable for all images.
Even given a practical case of image segmentation, there is no
standard for selecting suitable segmentation algorithm [Zhang
1996]. Usually some important features are selected to make
the main objects of image achieve the best and the most marked,
at the same time discard the irrespective or minor information
in order to reduce the complexity of classification [Vailaya
2001].
Range image represents the surface geographic information of
the objects in the form of 3D discrete coordinates, which has no
description of attribute and there does not exist topological
relation among the data points. Compared with normal 2D grey
image, range image does not exist visible boundary or the
possibility of “see” the objects. Current algorithms for image
segmentation and identification are all aiming at 2D grey image,
which cannot be used for the classification of 3D range image.
This paper presents the segmentation method of different
objects (buildings, roads, trees, independent objects such as
lamp-pole etc.) from range images through the spatial feature
analysis of the objects. The principle of object segmentation is:
1) to form horizontal mesh grids; 2) to project all the data
points on the grids and calculate the number of points of each
e