ADAPTIVE SNAKES FOR URBAN ROAD EXTRACTION 
* 
Junhee Youn 
James S. Bethel 
Geomatics Area, School of Civil Engineering, Purdue University, West Lafayette, IN 47907-2051, USA 
(vounj,bethel)ecn.purduc.edu 
Commission III, WG III/4 
Keywords: GIS, Photogrammetry, Urban, Feature, Extraction, Digital, Image 
Abstract: 
For quickly populating GIS database, it is important to derive accurate and truly road information from imagery. In this paper, we 
describe the problem of urban road extraction from digital imagery using adaptive active contour models (Snakes). Our road 
extraction processing has three steps. First, we segment the image based on the dominant road directions. Second, we detect the road 
lines with the so called *acupuncture method '. Finally, we refine the road edges by applying adaptive snakes to the corner desired 
approximation to extract the city block. During the process, we assume that the road network and block pattern in the city have a 
semi-regular grid pattern. For detecting the road lines, we exploit the distribution of edges in an urban area. Linear associated with 
roads are detected and these become the basis for initial approximations to road grid pattern for snakes based refinement. In order to 
accommodate variable line characteristics, we have developed an adaptive algorithm which locally modifies the weight of the energy 
terms. These ideas are applied to same actual urban imagery and the results are displayed and evaluated. 
1. Introduction 
Rapid and accurate generation of road data from imagery is an 
important issue in modern digital photogrammetry. Besides 
normal methods, there are two categories for road extraction. 
One is semi-automatic extraction and the other is fully 
automatic extraction. 
In semi-automatic extraction, approximations are given 
manually and an automatic algorithm extracts the road. In 
Vosselman and de Knecht (1995), the operator provides the 
initial information (starting point and direction) and the 
subsequent road segments' positions are determined after 
profile matching in aerial imagery. Manually given data can be 
used as initial information for snake approaches; Gruen and Li 
(1997) proposed the method that the human operator select a 
few seed points and the linear feature is automatically extracted 
based on dynamic programming and least squares B-spline 
snakes. For the fully automatic approach, initial information 
can be given by other sources. Agouris ef a/.(2001) extended 
the traditional snakes to function in a differential mode and 
propose a framework to differentiate change detection from the 
older record of GIS. There are several approaches to use multi 
resolution aerial imagery for automatic methodologies. That 
approach usually extracts the lines in a reduced resolution 
image, and edges are extracted with the original high-resolution 
image. Both outputs are merged by using a rule based system 
(Heipke et al. 1995). And Baumgartner ef al.(1999) composed 
road segment by using both resolution level with explicit 
knowledge about the road and grouped these segments 
iteratively, considering the local context. 
So far, not so many research groups concentrate on road 
extraction in urban areas. For rural case, the geometry, 
radiometry and topology characteristics of roads are simple to 
describe, compared to urban areas. The most difficulties for 
urban road extraction are coming from the fact that material 
and shape (linear pattern) for non-road structures are very 
similar to the road. However, the road network in an urban area 
is usually semi-grid pattern. Price (1999) assumed that road 
network model is a regular grid form and the road has visible 
465 
edges. He used two manually given intersecting road segments, 
which gave nominal information for the size and orientation of 
the grid. These segments are expanded with feature-based 
hypothesis and verification, and local context and a digital 
elevation model is used for refinement and verification of road 
segment. The grid characteristic for urban road networks is 
important assumption for our approach. Hinz er al. (2003) 
proposed a context-based decision making model for urban road 
detection, using complex relationships among roads buildings 
and vehicles. Here, context information about the urban road is 
mainly obtained from high-resolution imagery and a digital 
surface model. 
In this paper, we present automatic road extraction in urban 
areas with adaptive snakes, for which initial seed points are 
calculated from a new approach for detecting road ways and 
intersections. Assuming urban networks to be a semi-grid 
pattern, there will be evident roads has several dominant 
directions. Dominant directions for roads can be determined 
based on the fact that road edges and building edges are usually 
parallel in the urban area. With dominant road directions, 
preliminary road lines are extracted for the snake refinement. 
These road lines are used an initial approximations. An 
adoptive strategy is derived wherein weighting factor vary with 
context of the road grid. Results are presented for some actual 
urban imagery. 
2. Region-Based Image Segmentation 
The Basic concept of this approach is to split the image into 
regions based on dominant road directions. In urban areas, most 
of the roads are straight and the road network often has a grid 
form, so the roads will exhibit dominant directions. Meanwhile, 
having the same grid pattern for the whole area is an ideal case. 
Certain parts of the area may have north-south direction roads 
and another part may be northeast-southwest direction roads. 
Considering such various grid patterns, a partitioning scheme is 
proposed. In our approach, the parent image is subdivided into 
four child image blocks (regions) when the parent image which 
covers an area with more than two dominant road directions.