AUTOMATIC ROAD EXTRACTION FROM DENSE URBAN AREA BY INTEGRATED
PROCESSING OF HIGH RESOLUTION IMAGERY AND LIDAR DATA
Xiangyun Hu
C. Vincent Tao
Yong Hu
Geospatial Information and Communication Technology Lab (GeoICT),
Department of Earth and Space Science and Engineering,
York University, 4700 Keele Street, Toronto ON,Canada, M3J 1P3
xyhu@yorku.ca, tao@yorku.ca, yhu@yorku.ca
KEY WORDS: Extraction, Road, Multisensor, High Resolution, LIDAR, Urban, Contextual
ABSTRACT:
Automated and reliable 3D city model acquisition is an increasing demand. Automatic road extraction from dense urban areas is a
challenging issue due to the high complex image scene. From imagery, the obstacles of the extraction stem mainly from the difficulty
of finding clues of the roads and complexity of the contextual environments. One of the promising methods to deal with this is to use
data sources from multi-sensors, by which the multiple clues and constraints can be obtained so that the uncertainty can be
minimized significantly. This paper focuses on the integrated processing of high resolution imagery and LIDAR (LIght Detection
And Ranging) data for automatic extraction of grid structured urban road network. Under the guidance of an explicit model of the
urban roads in a grid structure, the method firstly detects the primitives or clues of the roads and the contextual targets (i.c., parking
lots, grasslands) both from the color image and lidar data by segmentation and image analysis. Evidences of road existing are
contained in the primitives. The candidate road stripes are detected by an iterative Hough transform algorithm. This is followed by an
procedure of evidence finding and validation by taking advantage of high resolution imagery and direct height information of the
scene derived from lidar data. Finally the road network is formed by topology analysis. In this paper, the strategy and corresponding
algorithms are described. The test data set is color ortho-imagery with 0.5 m resolution and lidar data of Toronto downtown area. The
experimental results in the typical dense urban scene indicate it is able to extract the roads much more reliable and accurate by the
integrated processing than by using imagery or lidar separately. It saliently exhibits advantages of the integrated processing of the
multiple data sources for the road extraction from the complicated scenes.
1. INTRODUCTION
Automatic road extraction from remotely sensed imagery has
attracted much attention for the last few decades. In this issue, a
great number of research papers were published both in
geospatial and computer vision communities. In general, road
extraction consists of four steps (Gruen and Li, 1995): road
sharpening, road finding, road tracking, and road linking. In the
earlier research (Bajesy and Tavakoli, 1976; Nevatia and Babu,
1980), some line detection algorithms were presented for
extracting the roads from remotely sensed imagery. There is not
much high-level knowledge involved in the methods for road
finding. To process gaps bridging, road tracing and handle the
complicated image scenes, more sophisticated strategies should
be used for more reliable extraction. Knowledge or rule based
methods or similar methods based on hypothesis-verification
(Mckeown and Delinger, 1988; Tonjes, R., and S. Growe, 1998;
Trinder and Wang, 1998) had been used for handling the issue
of linear feature alignment and fragmentation. Optimal route
search algorithms were frequently employed as semiautomatic
road extraction. The optimization can be realized by dynamic
programming (Gruen and Li, 1995; Bazohar and Cooper, 1998),
snakes (Trinder and Li, 1995; Gruen and Li, 1997; Tao et. al,
1998; Agouris ef. al. 2001) and Kalman filtering (Vosselman
and de Knecht , 1995). Furthermore, contextual information
supported methods (Stilla, 1995; Baumgartner et.al. 1999) were
applied to extract road more reliably. Actually we can also find
many strategies (Bazohar and Cooper, 1998; Couloigner and
Ranchin 2000; Laptev et. al 2000; Katartzis, et.al., 2001, Hinz
and Baumgartner; Hu and Tao, 2003; Hu and Tao, 2004) which
attempt to combine the methods or use the specific techniques
in order to deal with different scenarios in image scale,
complexity and road type etc. However, automating road
extraction is still challenging as the involved problems of
intelligent image understanding are too complicated to be
solved straightforward. Most of the methods applied to extract
roads from open or rural areas were successful to some extent
due to the relative simple image scene and road model. For the
extraction of roads in dense urban areas, especially from high
resolution imagery, there are primary obstacles which lead to
unreliable extraction results: complicated image scene and road
model, furthermore, occlusion caused by high buildings and
their shadows. In other words, the lack of information,
especially three-dimensional information is the principle
difficulty in obtaining the road information with high reliability
and accuracy in the urban scenes.
Airborne lidar (Light Detection And Ranging) is a relatively
new data acquisition system complementary to traditional
remote sensing technologies. Lidar data contains plenty of scene
information, from which most ground features such as roads and
buildings are discernible. Roads have homogeneous reflectivity
in lidar intensity and the same height as bare surface in
elevation. Lidar range data is able to improve the analysis of
optical images for detecting roads in urban areas (Hofmann,
2001). But the use of range data requires that the urban areas be
relatively flat. Some researchers (Zhang et al., 2001; Alharthy
and Bethel, 2003; Hu, 2003) used the height information
derived by subtracting the DTM from the DSM to reason if a
region is on the ground and to compensate the missing
information in classification of aerial images. In cases when
shadows or buildings occlude road segments, their shape can be
well detected due to the height information. Lidar intensity data
has good separability if the wavelength of the laser is suitable
Interna
for gro
feature:
compar
are ver
2002).
structui
of pare
main oi
detecte:
imager’
lack of
of the :
and toj
improv
As mer
can be
while i
verifica
existen:
motivat
method
resoluti
informé
section
strategy
the roa
informé
detectic
sources
remark:
2.0
LID
2.1 Dat
In early
mission
its ATL
region.
pebbles
intensit
million
generat
also ob
range
informe
bare Ea
plane. I
and the
from oi
resoluti
sample:
lidar da
area. Fi
imager
comput
area, w
contain:
kinds o
vehicle:
