kinds of relations among the criteria selecting an appropriate
weighting vector (Ahn, 2008).
By integrating different attributes using the
OWA
operators
in this research, it is intended to provide a more straight-
forward and satisfactory solution in road vectorization than
conventional approaches. In order to evaluate the perfor-
mance of the proposed approach, several simulated and real
high-resolution images were tested.
This paper consists of five sections. Following the intro-
duction, the novel road vectorization methodology using
OWA
operators and weighted graph theory are presented followed by
the experimental results on different simulated and real datas-
ets, then a discussion on the obtained results. Finally, the con-
cluding remarks and future research directions are discussed.
Methodology
The proposed automatic road centerline vectorization approach
uses a graph-based model to construct a road network. A graph-
based road model consists of a set of nodes some pairs of which
are connected by links. Each feasible link is characterized by
a cost value indicating the possibility of connection between
nodes. In this research, each road key point is represented by
a node and each line segment, which is defined between two
nodes, is symbolized by a link. The cost of each line segment
is calculated based on the aggregation of geometric criteria
according to
OWA
. In this paper, an automatic road centerline
vectorization is organized in five steps as follows (Figure 1):
1. Obtaining optimized distribution of road key points in
the binary road raster map.
2. Identifying and normalizing effective criteria of road
centerline vectorization.
3. Modeling decision strategies of
OWA
by specifying
order weights and criteria weights to establish a true
connection between road key points.
4. Aggregating the criteria values based on
OWA
by means
of different decision strategies.
5. Calculation of the cost and selection of suitable line
segments defining appropriate threshold to construct
the road network topology.
Road Key Point Determination
The procedure of road key point determination is based on the
idea of image space clustering (Doucette
et al
., 2001). In this
research, a road network is defined as continuous line seg-
ments which are covered by different road patches obtained
during clustering process. Each road patch is identified by a
point (road key point) which is the centroid of each cluster.
Based on this model, road patch determination (road key
point determination) is done by means of Dynamic Road Pix-
els Clustering using Particle Swarm Optimization (
DRPCPSO
) al-
gorithm. The
DRPCPSO
algorithm is a modification of dynamic
clustering using
PSO
algorithm (Omran
et al
., 2006) to optimize
the number and distribution of road patches in the binary road
raster map. The
DRPCPSO
algorithm is a hybrid clustering al-
gorithm composed of a binary
PSO
to automatically determine
the optimum number and distribution of road key points and
a traditional K-medians to adjust the chosen road key points.
In the
DRPCPSO
algorithm the partitioning procedures of the
spatial positions of road pixels are repeated until the best set
(the lowest cost function) of road key points is obtained. In
this way an appropriate pattern of road key points which guar-
antees the optimality of road centerline extraction is achieved.
The proposed
DRPCPSO
algorithm is summarized in Figure 2.
Road Segment Criteria Definition
Following the image space clustering method, road patches
are represented by optimized distribution of road key points.
The adjacent road key points are to be connected to construct
the road network topology. In order to make true connections,
a combination of several criteria or prior information about
the roads and their appearance should be used. The connec-
tion criteria should present fundamental attributes of roads
including radiometric and geometric characteristics. Since the
connection strategy is defined on the binary road image, no ra-
diometric criteria are used. For the road key point connection
in this paper, the compliance of each road segment with three
criteria is checked. The summarized criteria are as follows:
•
Distance (d)
This criterion is defined as the Euclidean distance of
the connecting line between two road key points. It is
identical to the line segment length.
•
Direction Difference (DD)
The direction difference is measured by the angle
between the direction vectors of the two line segments.
The direction of one line segment is defined by the direc-
tion of the vector which connects its two end points (see
Figure 1. OWA-based road centerline vectorization diagram.
108
February 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
