ROAD EXTRACTION FOR THE UPDATE OF ROAD DATABASES IN SUBURBAN 
AREAS 
A. Grote*, C. Heipke 
Leibniz Universität Hannover, Institute of Photogrammetry and Geoinformation, 30167 Hannover, Germany - 
(grote, heipke)@ipi.uni-hannover.de 
KEY WORDS: Image understanding, Road extraction, Computer Vision, Feature extraction, Updating 
ABSTRACT: 
This paper deals with road extraction in suburban areas from high resolution aerial images. The extraction results are intended to be 
used for updating a road database. The road extraction algorithm follows a region-based approach in which the image is first 
segmented using the normalized cuts algorithm with colour and edge criteria. Then, the initial segments are grouped to larger 
segments in order to overcome the oversegmentation which is a result of the first step. The segments are subsequently evaluated by 
shape criteria in order to extract road parts. Large segments that contain several roads are shaped irregularly; therefore, large 
segments are split prior to the road part extraction. The splitting is based on the skeleton of the segment. After the road part 
extraction, most roads in the image are covered by one extracted road part. However, some roads are covered by several road parts 
with gaps between them. In order to combine these road parts to one road, neighbouring road parts are connected if they have a 
similar main direction and a relatively high continuation smoothness. Results for some test images show that the approach is suitable 
for the extraction of roads in suburban images. 
1. INTRODUCTION 
Up-to-date road databases are important for many applications, 
for example map production, traffic management and spatial 
planning. As the need for up-to-date road data increases, so 
does the need for automatic road extraction methods, because 
manual acquisition of road data is quite time-consuming and 
expensive. Therefore, road extraction has been extensively 
researched in recent years. 
For rural areas, many approaches already exist, for example 
(Bacher and Mayer, 2005; Geraud and Mouret, 2004). For 
urban and suburban areas, on the other hand, there are only few 
approaches. The main problems with road extraction in urban 
areas are the more complex scene content and the different 
structure of the road network compared to rural areas. Urban or 
suburban scenes consist of many different objects like houses, 
trees and vehicles, which leads to a scene that is composed of 
many small regions. Urban roads typically do not have the 
distinct line-shaped appearance that they have in rural areas, 
and the main network function is not a short connection 
between two distant places but the connection to the major road 
network for every building. Roads in urban areas often are laid 
out in a regular grid, which can be exploited by road extraction 
algorithms (Price, 1999; Youn and Bethel, 2004). But 
especially in Europe, urban road networks can be quite irregular. 
Effective strategies for road extraction in urban areas often 
work from small entities to bigger entities (for example Hinz, 
2004), where lines are grouped to lanes, carriageways and road 
networks, according to a detailed road model. It is also helpful 
to include both local and global features in an extraction 
strategy (see for example Doucette et al., 2001; Doucette et al., 
2004). Road extraction only from grey value images seems to 
be impossible in urban areas. Most approaches for urban areas 
take some additional information into account: Hinz (2004) 
uses a DSM and multiple overlapping images; colour images 
are widely used (for example Zhang and Couloigner, 2006; 
Doucette et al., 2004), also information from geographic 
databases, often combined with a verification or update of the 
database (Gautama et al., 2006; Sims and Mesev, 2007; Zhang, 
2004). Hu and Tao (2007) use hierarchical grouping of line 
segments to extract the main roads in an urban area from 
satellite images. Another promising approach is described by 
Hu et al. (2007); it is a road tracking approach where very little 
user input is necessary. They use a region-based model to 
extract reliable road parts from which the tracking starts. 
In this paper, we describe an approach for road extraction in 
suburban areas based on previous work about the extraction of 
road parts for the verification of database roads in suburban 
areas (Grote et al., 2007). Road parts are extracted as described 
there, but without the database information, because the focus 
of this paper is the extraction of new roads for database update. 
We employ a region-based approach on high resolution aerial 
images working from small local regions to roads as groups of 
road parts. The image is first segmented using the normalized 
cuts algorithm (Shi and Malik, 2000). The resulting relatively 
small segments are grouped to form larger segments, and from 
these grouped segments road parts are extracted. To cover cases 
where the road is fragmented (due to different road surfaces or 
context objects) road parts with similar main directions are 
assembled into strings of road parts or subgraphs. 
2. APPROACH 
2.1 Overview 
The goal of the work described in this paper is to extract roads 
in suburban areas for the updating of a road database. The focus 
of this paper lies on the extraction of road parts and their 
assembling to road strings or subgraphs. 
Corresponding author. 
563