Content uploaded by Craig A. Knoblock
Author content
All content in this area was uploaded by Craig A. Knoblock
Content may be subject to copyright.
Automatic Extraction of Road Intersection Position,
Connectivity, and Orientations from Raster Maps
Yao-Yi Chiang and Craig A. Knoblock
University of Southern California
Department of Computer Science and Information Sciences Institute
4676 Admiralty Way, Marina del Rey, CA 90292
yaoyichi, knoblock@isi.edu
ABSTRACT
The road network is one of the most important types of in-
formation on raster maps. In particular, the set of road in-
tersection templates, which consists of the road intersection
positions, the road connectivities, and the road orientations,
represents an abstraction of the road network and is more ac-
curate and easier to extract than the extraction of the entire
road network. To extract the road intersection templates
from raster maps, the thinning operator is commonly used
to find the basic structure of the road lines (i.e., to extract
the skeletons of the lines). However, the thinning operator
produces distorted lines near line intersections, especially at
the T-shaped intersections. Therefore, the extracted posi-
tion of the road intersection and the road orientations are
not accurate. In this paper, we utilize our previous work
on automatically extracting road intersection positions to
identify the road lines that intersect at the intersections and
then trace the road orientations and refine the positions of
the road intersections. We compare the proposed approach
with the usage of the thinning operator and show that our
proposed approach extracts more accurate road intersection
positions and road orientations than the previous approach.
Categories and Subject Descriptors
H.2.8 [Database Management]: Database Applications—
Spatial Databases and GIS
General Terms
Algorithms, Design
Keywords
Raster map, road layer, road intersection template, road
orientation, thinning
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
ACM GIS ’08, November 5-7, 2008. Irvine, CA, USA
Copyright 2008 ACM 978-1-60558-323-5/08/11 ...$5.00.
1. INTRODUCTION
Raster maps are widely available and contain valuable in-
formation, such as road lines, labels, and contour lines. For
instance, a digital raster graphic (DRG), which is a georef-
erenced scanned image of a United States Geological Sur-
vey (USGS) topographic map, can be purchased from the
USGS website or accessed freely from TerraServer-USA.1
Map repositories like the University of Texas Map Library2
contain information-rich scanned maps for many areas and
even historical scanned maps. Moreover, web mapping ser-
vice providers such as Google Maps,3Microsoft Live Search
Maps,4and Yahoo Maps5provide high quality digital maps
covering many countries with rich information layers such
as business locations and traffic information.
To extract information from the raster maps, one ap-
proach is to process the vector data used to generate the
raster maps and then extract the desired information, such
as the road vectors. However, it is generally difficult to
access the original vector data for many raster maps. For
example, the web mapping services provided by Google, Mi-
crosoft, and Yahoo all provide their maps only in the raster
format and there is no public access to the original vector
data. Furthermore, many raster maps found on the Internet
are simply images without any auxiliary information of the
original vector data. Hence, a more general approach is to
utilize image processing and graphics recognition technolo-
gies to separate the information layers (e.g., the road layer
and text layer) on the raster maps and then extract and
rebuild selected layers [7].
Among the information layers on the raster maps, one of
the most important layers is the road layer. Instead of ex-
tracting the entire road layer, the set of road intersection
templates, which consists of the road intersection position,
connectivity, and orientations, provides the basic elements of
the road layout and is more accurate and easier to extract
than the extraction of the entire road network. Since the
road layers commonly exist across many different geospatial
layers (e.g., satellite imagery, vector data, etc.), by match-
ing the set of road intersection templates from a raster map
with another set of road intersection templates from a geo-
referenced data set (e.g., vector data), we can identify the
geospatial extent of the raster map and align the raster map
with other geospatial sources [3]. An example of an inte-
1http://terraserver-usa.com/
2http://www.lib.utexas.edu/maps/
3http://maps.google.com/
4http://maps.live.com/
5http://map.yahoo.com
grated and aligned view of a tourist map and satellite im-
agery is shown in Figure 1. Accurate road intersection tem-
plates (i.e., road intersection templates with accurate posi-
tions, road orientations, and connectivities) help to reduce
the search space during the matching process by selecting
road intersections with the same connectivity and similar
road orientations as initial matching candidates. The accu-
rate positions of the templates also enhance the alignment
results. Furthermore, the extracted road intersection tem-
plates with accurate positions and road orientations can be
used as seed points to extract roads from aerial images [10]
as shown in Figure 2, especially when the access to vector
data is limited.
To extract the road intersection templates from a raster
map, the first step is to detect the positions of the road inter-
sections on the raster map and then convert the road lines
Figure 1: The integration of imagery from Google
Maps and a tourist map (Tehran, Iran).
Figure 2: Use the extracted road intersection tem-
plates to extract roads from imagery
around the road intersections into vector format, which is
difficult since the raster maps are usually complicated and
contain many different layers such as characters and road
lines. In our previous work [7], we presented a technique
that works on the pixel level of the raster maps and uti-
lizes the distinctive geometric properties of the road lines
and road intersections to automatically extract road inter-
sections from raster maps. During the extraction process,
our previous technology reconnects the broken lines by first
thickening the lines and then utilizing the thinning operator
to generate one-pixel width lines to identify possible road
intersections.
The idea of the thinning operator [13] is to produce one-
pixel width lines that represent the central lines of the thick
lines (i.e., to preserve the basic structures of the thick lines).
The thinning operator first identifies every foreground pixel
that connects to one or more background pixels as candidate
to be converted to the background (i.e., to reduce the region
of foreground objects). Then a confirmation step checks if
the conversion of the candidate will cause any disappear-
ance of original line branches to ensure the basic structures
of the original objects will not be compromised. The thin-
ning operator is robust and efficient; however, the thinning
operator distorts the lines near the intersections. As shown
in Figure 3, if we apply the thinning operator directly on
the thick lines shown in Figure 3(a), the lines that are near
intersections are all distorted as shown in Figure 3(b). Our
approach in [7] to minimize the extent of the distortions is to
first erode the lines using an binary erosion operator [13] as
shown in Figure 3(c) and then apply the thinning operator.
However, there are still distortions even after the erosion
operator is applied as shown in Figure 3(d).
In the previous work, the proposed technologies focused on
efficiently extracting the positions of road intersections and
(a) Thickened road lines (b) Result of applying the
thinning operator without
erosion
(c) Thickened road lines af-
ter erosion
(d) Result of applying the
thinning operator after ero-
sion
Figure 3: Distorted road lines near road intersec-
tions with the thinning operator
extracting the road orientations using a simpler technique
that works on the one-pixel width road lines generated by
the thinning operator. Therefore, the extracted road orien-
tations suffer from the distortions produced by the thinning
operator and the positions of the road intersections are not
accurate. To overcome the distortion problem and to ex-
tract accurate road intersection templates, in this paper, we
present an approach that employs the results of our previous
work to trace the road lines for accurate road orientations
and then utilizes the road orientations to refine the positions
of the road intersections.
The remainder of this paper is organized as follows. Sec-
tion 2 explains our previous road intersection extraction
work. Section 3 describes our improved approach to au-
tomatically extract road intersection templates. Section 4
reports on our experimental results. Section 5 discusses the
related work, and Section 6 presents the conclusion and fu-
ture work.
2. BACKGROUND WORK
In our previous work [7], we proposed a technique to auto-
matically extract the road intersections from various raster
maps. There are three major steps in our previous ap-
proach. First, we remove the background pixels from the
raster maps. Next, we separate the road lines from other
objects in the foreground pixels and rebuild the road lines.
Finally, we detect road intersection candidates on the road
lines and utilize the connectivity (i.e., the number of lines
that intersect at a given road intersection candidate) to find
actual road intersections. The road orientations are also
computed in the last step as a by-product. We briefly ex-
plain these three steps of the previous approach in turn in
the following subsections.
2.1 Automatic Segmentation
Since the foreground pixels of the raster maps contain the
road layers, the first step for extracting road intersections
is to extract the foreground pixels. We utilize a technique
called segmentation with automatically generated thresholds
to separate the foreground pixels from the background pixels
of the raster maps. Because the background colors of raster
maps have a dominant number of pixels and the foreground
colors have high contrast against the background colors, we
generate the segmentation thresholds by analyzing the shape
of the grayscale histogram of the raster map [15]. We first
identify the largest luminosity cluster in the histogram as the
background cluster and then classify other clusters as either
background clusters or foreground clusters by comparing the
number of pixels in the clusters. After we remove the back-
ground clusters from the original map shown in Figure 4(a),
we produce the binary map shown in Figure 4(b)
2.2 Pre-Processing – Extract and Rebuild Road
Layer
With the extracted foreground pixels, we first utilize our
parallel-pattern tracing algorithm [7] to detect the road width
of the majority of roads on the raster map. For a given fore-
ground line, the parallel-pattern tracing algorithm employs
two convolution masks that work on the horizontal and ver-
tical directions to search for corresponding parallel lines to
determine the road width. The detected road width is shown
with gray dashed lines in Figure 4(c). The road width is im-
portant in the latter steps to rebuild the road layer.
(a) Original raster map (b) Binary map
(c) Road width (d) Extracted road lines
(e) Thickened road lines (f) Eroded road lines
(g) Thinned road lines (h) Road intersection candi-
dates
Figure 4: Automatic extraction of road intersections
After we obtain the road width, we use the text/graphic
separation technique [1] to remove the foreground pixels that
do not hold the properties that constitute roads (e.g., pixels
for labels, contour lines, etc). The extracted road layer is
shown in Figure 4(d). Finally, to extract the structure of
the road layer, we utilize the binary dilation operator with
the number of iterations determined by the detected road
width to first thicken the road line as shown in Figure 4(e).
Then we apply the binary erosion operator and the thinning
operator to generate one-pixel width road lines as shown in
Figure 4(f) and Figure 4(g).
2.3 Determine Road Intersections and Extract
Connectivity with Road Orientation
With the one-pixel-width road lines, we utilize the corner
detector [14] to detect road intersection candidates as shown
with the cross marks on the road lines in Figure 4(h). For
each road intersection candidate, we draw a box around it
and then use the number of foreground pixels that intersects
with this rectangle as the connectivity of the intersection
candidate as shown in Figure 5. If the connectivity is less
than three, we discard the point; otherwise it is identified
as a road intersection point. Subsequently, without tracing
the line pixels, we link the road intersection candidate to the
intersected foreground pixels on the rectangle boundaries to
compute the orientations of the road lines.
The final road intersection extraction results are shown in
Figure 6. Although we successfully extract most of the road
intersections, the positions of some extracted road intersec-
tions are not at the center points of the intersected lines and
the road orientations are not accurate, especially the ones on
the intersection of a T-shape roads. This is because the ex-
tracted road lines near the intersections are distorted by the
morphological operators (i.e., the binary dilation operator,
the binary erosion operator, and the thinning operator) as
shown in Figure 4(g); and the method we use to compute the
road orientations does not take into account the distortions.
In the next section, we present our improved approach to
overcome the distortion problem and build an accurate road
intersection template by utilizing the extracted road inter-
section position, the extracted one-pixel-width lines, and the
detected road width.
3. AUTOMATIC EXTRACTION OF ROAD
INTERSECTION TEMPLATES
The overall approach to extract road intersection tem-
plates is shown in Figure 7, where the gray boxes indicate
the results we use from our previous approach in [7]. Based
on our previous work, we recognize that there is distortion
near each road intersection, and the extent of the distortion
is determined by the thickness of the line, which is deter-
mined by the number of iterations of the morphological op-
erators. And the number of iterations of the morphological
operators is based on the road width that is detected using
the parallel-pattern tracing algorithm.
Therefore, with the results of the road intersection posi-
tions and the road width, we first generate a binary large
object (i.e., blob, a connected foreground object on an im-
Figure 5: Construct lines to compute the road ori-
entations
Figure 6: Extracted road intersections
Figure 7: Overall approach to extract the road in-
tersection templates from raster maps
age) for each intersection to create a blob image. We then
intersect the blob image with the thinned line image to label
the location of the potential distorted lines. Finally, we trace
the unlabeled lines to compute the road orientations and use
the road orientations to update the positions of the road in-
tersections to generate accurate road intersection templates.
3.1 Generating Road Intersection Blobs
Since the thinning operator produces distorted lines near
the road intersections and the extent of the distortion is de-
termined by the road width, we can utilize the extracted
road intersections and the road width from our previous ap-
proach [7] to label the locations of the potential distorted
lines.
For each extracted road intersection shown in Figure 6, we
generate a rectangular blob using the binary dilation oper-
ator with twice the road width as the number of iterations.
The result blob image is shown in Figure 8(a). We then in-
tersect the blob image with the thinned line image shown in
Figure 4(g) to label the locations of the potential distorted
lines. As show in Figure 8(b), the line segments within the
gray boxes are labeled by the blob image as the potential
distorted lines.
Subsequently, we use the labeled image shown in Fig-
ure 8(b) to detect road line candidates for each intersection.
We identify the lines linked to each blob and their contact
points by detecting the line pixels that have any neighbor-
ing pixel labeled by the gray boxes. These contact points
indicate the starting points of a road line candidates for the
blob. In the example shown in Figure 8(c), the road inter-
section associated with the left blob has three contact points
that are on the top, right, and bottom of the blob, and these
points will be used as starting points to trace the road line
candidates for the road intersections of the blob.
Instead of drawing a rectangle around each intersection to
find the contact points as in our previous approach [7], we
generate the blob image to take advantage of using image
processing operations, which are more robust and easier to
implement. For example, when we generate the blob image,
if two intersections are very close to each other, their blobs
automatically merge into one big blob by the binary dila-
tion operator without additional implementations to com-
pute the intersections of every rectangle box. This is impor-
tant because when two intersections are very close to each
other, the thinned line between them is totally distorted as
shown in Figure 9. With the merged blob shown in Fig-
ure 9(d), we can associate the four lines linked to this blob
with the two intersections as the road line candidates of the
intersections. In the last step of updating road intersection
templates, a filtering step will decide which candidates need
to be discarded so that the two intersections will still have
three road lines instead of four.
3.2 Tracing Road Lines
For each road intersection blob, we start to trace the road
lines from its contact points using the flood-fill algorithm
shown in Table 1. The flood-fill algorithm first labels the
contact point as visited and then checks the eight neighbor-
ing pixels of the contact point to find unvisited points. If
any of the eight neighboring pixels are not labeled as visited
and are not within a blob, the neighboring pixel will be set
as the next visit point for the flood-fill algorithm to consider.
When the flood-fill algorithm considers a new pixel, we
also record the position of the pixel to latter compute the
road orientation. The number of pixels that the flood-fill
algorithm can trace from each contact point is controlled by
the variable MaxLinePixel shown in Table 1. The flood-fill
algorithm counts the number of pixels that it has visited and
makes sure the count is less than the MaxLinePixel variable
every time before it visits a new pixel or it stops. As shown
in Figure 10, instead of tracing the whole curve starting
from the two contact points (i.e., the one on the right and
the one on the bottom), we assume that roads near the con-
tact points are straight within a small distance (i.e., several
meters within a street block on the raster map) and uti-
lize the MaxLinePixel to ensure that the flood-fill algorithm
traces only a small portion of the road lines near the contact
(a) Road intersection blob image
(b) Blob image intersected with the
thinned line image
(c) Contact points
Figure 8: Generate the blob image to label the road
lines
points.
After the flood-fill algorithm walks through the lines from
each contact point and records the position of the line seg-
ment pixels, we utilize the Least-Squares Fitting algorithm
to find the linear functions of the lines. Assuming a linear
function Lfor a set of line pixels traced by the flood-fill algo-
rithm, by minimizing the sum of the squares of the vertical
offsets between the line pixels and the line L, the Least-
Squares Fitting algorithm finds the straight line Lthat best
represents the traced line pixels. The algorithm works as
follows:6
Given a set of pixel locations {(xi, yi)},i= 1,...,n, in the
map, find the target line function L, where
L:Y=m×X+b(1)
6The proof of the Least-Squares Fitting algorithm can
be found in most statistics textbooks or on the web at
http://mathworld.wolfram.com/LeastSquaresFitting.html
m=nPxy −PxPy
nPx2−(Px)2(2)
b=Py−mPx
n(3)
The computed line functions are then used in the next step
of updating road intersection templates to identify actual
road lines and refine the positions of the road intersections.
3.3 Updating Road Intersection Templates
In the previous steps, we associated a set of road line
candidates with a road intersection and traced the road line
(a) Original roads and the
extracted road intersections
(b) Thickened road lines
(c) Distorted thinned lines (d) Merged blobs
Figure 9: Merge nearby blobs to trace roads
Table 1: Flood-fill algorithm
Flood-fill(Imag e, x, y)
1if Imag e.isALineP ixel(x, y) = T RU E
2 AND Image.isV isited(x, y)6=T RU E
3 AND Imag e.isW ithInABl ob(x, y)6=T RU E
4 AND PixelCount <MaxLinePixel
5then
6RecordP ixelP osition(x, y )
7P ixelC ount + +
8Image.SetV isited(I mage, x, y)
9Flood-fill(Imag e, x + 1, y)
10 Flood-fill(Imag e, x −1, y)
11 Flood-fill(Imag e, x, y + 1)
12 Flood-fill(Imag e, x, y −1)
13 Flood-fill(Imag e, x + 1, y + 1)
14 Flood-fill(Imag e, x −1, y −1)
15 Flood-fill(Imag e, x + 1, y −1)
16 Flood-fill(Imag e, x −1, y + 1)
Figure 10: Trace only a small portion of the road
lines near the contact points
functions. In this step, we first compute the intersections of
these road line candidates and then filter out outliers of the
road line candidates.
There are three possible intersecting cases for the road
line candidates of one intersection. The original maps of
the three possible intersection cases are shown on the left in
Figure 11. The middle images shown in Figure 11 are the
thinned lines with the locations of the potential distorted
lines labeled by the blob images, and the right images are
the traced line functions (i.e., the line functions that are
computed using the Least-Squares Fitting algorithm) drawn
on a two-dimension plane.
The first case is where all of the road line candidates in-
tersect at one point as shown in Figure 11(a). The second
case is where the road line candidates intersect at multiple
points and the intersection points are all near the road inter-
section position, as shown in Figure 11(b). The third case is
where the road line candidates intersect at multiple points
and some of the points are not near the road intersection
(i.e., the outliers), as shown in Figure 11(c).
For the first case shown in Figure 11(a), the position of the
updated road intersection template is the intersection point
of the road line candidates, and the road orientations are the
orientations of the intersecting roads, which are 0 degrees, 90
degrees, and 270 degrees, respectively. For the second case
shown in Figure 11(b), the position of the road intersection
template is the centroid of the intersection points of all road
line candidates and the road orientations are the orientations
of the intersecting roads, which are 80 degrees, 172 degrees,
265 degrees, and 355 degrees, respectively.
Since the extent of the distortion depends on the road
width, the positional offset between any intersection of the
road line candidates and the road intersection should not be
larger than the road width. We first consider the upper road
intersection in the third case shown in Figure 11(c). The in-
tersections where the dashed line intersects with the other
two lines are more than a road width away from the upper
road intersection, so we discard the dashed line and use the
(a) Case one: the three lines intersect at one point
(b) Case two: the three lines intersect at multiple points
(c) Case three (the upper road intersection): the four lines
intersect at multiple points with one outlier (the dashed line)
Figure 11: Filter out the outliers
centroid of the intersection points of the other three road
line candidates to update the position of the upper road in-
tersection. The road orientations of this road templates are
60 degrees, 150 degrees, and 240 degrees, respectively. Sim-
ilarly, for the lower road intersection shown in Figure 11(c),
the road line candidate that is almost parallel to the dashed
line will be discarded, and the dashed line is kept as a road
line for the road intersection. The road orientations of the
lower road templates are 60 degrees, 240 degrees, and 330
degrees, respectively.
The last example shows how the merged blob helps to
extract correct road orientations even when one of the lines
is totally distorted during the thinning operation. This case
holds when the distorted line is very short and hence it is
very likely to have the same orientation as the other lines.
For example, in Figure 11(c), the distorted line is part of a
straight line that goes through the intersection, so it has the
same orientation as the 240 degree line.
4. EXPERIMENTS
In this section, we evaluate our approach by conducting
experiments on raster maps from various sources. We first
explain the test data sets and our evaluation methodology
and then analyze the experimental results and provide a
comparison to our previous work [7].
4.1 Experimental Setup
We evaluate 10 raster maps from five different sources,
MapQuest Maps,7, OpenStreet Maps.8, Google Maps, Mi-
7http://www.mapquest.com/
8http://www.openstreetmap.org/
crosoft Live Search Maps, and Yahoo Maps, The 10 maps
cover two areas in the United State, one in Los Angeles, Cal-
ifornia and the other one in St. Louis, Missouri. Figure 12
shows two example maps for these areas.
We first apply our road intersection extraction application
from our previous approach to detect the road width and
extract a set of road intersection positions and the thinned
road lines for each map. Then we utilize the technique de-
scribed in this paper to extract the accurate road intersec-
tion templates from the raster maps. We report the accuracy
of the extraction results using the positional offset, orienta-
tion offset, and connectivity offset. The positional offset is
the average number of pixels between the extracted road
intersection templates and the actual road intersections on
the raster maps. The actual road intersections on the raster
maps are defined as the centroids of the intersection areas of
two or more intersecting road lines. The orientation offset is
the average number in degrees between the extract road ori-
entations and the actual road orientations. The connectivity
offset is the total number of missed road lines. We manually
examine each road intersection on the raster maps to ob-
tain the ground truth of the positions of the road intersec-
tions, the connectivities and the road orientations. Figure 13
shows the ground truth of two road intersection templates.
For the road intersection template in Figure 13(a), the road
orientations are 0 degrees, 90 degrees, and, 270 degrees, re-
spectively.
4.2 Experimental Results
From the 10 raster maps, we extracted a total of 139 road
intersection templates with 438 lines and the results are
(a) Live Search Maps, Los Angeles, Cali-
fornia
(b) Yahoo Maps, St. Louis, Missouri
Figure 12: Examples of the test maps
(a) The ground truth on a
map from Google Maps (b) The ground truth on a
map from OpenStreet Maps
Figure 13: The ground truth
shown in Table 2. The average positional offset is 0.4 pix-
els and the average orientation offset is 0.24 degrees, which
shows our extracted road intersection templates are very
close to the ground truth on these test maps. In order to
achieve higher accuracy in the positional and orientation off-
sets, we need to discard the lines that do not have accurate
orientations in the filtering step (i.e., the outliers). So the
connectivity is lower than our previous work; we missed 13
lines from a total of 451 lines in order to trace the lines to
extract the road intersection templates. These 13 lines all
belong to the road intersections that are near the boundaries
of the map and hence we cannot find road lines long enough
to compute the correct orientations.
We also compare the extracted template with the ex-
tracted template using the methods of our previous approach
in [7], which uses only the thinning operator. The compari-
son results of the positional offset and orientation offset are
shown in Figure 14. Our proposed approach in this paper
has a large improvement on the results of every map source
for both positional offset and orientation offset. Figure 16
shows example extraction results using the approach in this
paper and our previous approach in [7]. This figure shows
the distorted road lines from our previous work are corrected
using the approach in this paper.9
4.3 Computation Time
The system is built using Microsoft Visual Studio 2005
on a Windows 2003 Server with Intel Xeon 1.8 GHZ Dual
Processors and 1 GB of memory. The largest test map is 809
pixels by 580 pixels and it took 11 seconds to first extract
the road intersection positions using the approach in [7] and
another 5 seconds to extract the road intersection templates
using the techniques proposed in this paper. The dominant
factors of the computation time are the map size, the number
of foreground pixels of the raster map, and the number of
road intersections on the raster map. The average time is
10.5 seconds to extract the position of road intersections
and 4.7 seconds to extract the road intersection templates,
which is sufficient for many applications that need real-time
intersection extraction results.
5. RELATED WORK
Much research work has been performed in the field of
extracting and recognizing graphic objects from raster maps,
9The comparison results for this map using only the thinning
operator is shown in Figure 6
(a) Positional offset comparison (in pixels)
(b) Orientation offset comparison (in degrees)
Figure 14: Experimental results using the approach
from this paper and our previous work using the
thinning operator
such as extracting road intersections [6, 7, 8], separating
lines from text [1, 5, 11, 12], and recognizing contour lines [4,
9] from raster maps.
In our previous work [6], we combine a variety of im-
age processing and graphics recognition methods such as
the morphological operators to automatically separate the
road layers from the raster maps and then extract the road
intersection points. Since the thinning operator is used in
the work, the extracted road intersections have high posi-
tional and orientation offsets when the road width is wide.
To overcome this problem, in [7] we utilize the Localized
Template Matching (LTM) [2] to improve the positional off-
set. However, since the road orientations are not accurate,
it is difficult for LTM to find the correct matches to improve
the positions and further correct the road orientations. In
this paper, we utilize the detected road width, the positions
of the road intersections, and the thinned lines to trace the
road lines and then update the road intersection positions to
achieve the best results on the extraction of road intersection
templates compared to our previous work.
For the other related work specifically working on extract-
ing road intersections, Habib et al. [8] utilize several image
processing algorithms to automatically extract road inter-
sections from raster maps. In [8], they detect the corner
points on the extracted road edges and then classify the
Table 2: The positional offset, orientation offset, and the connectivity offset
Map Number of Number of Positional Orientation Connectivity
Source Extracted Extracted Offset Offset Offset
Road Intersections Road Lines (in pixels) (in degrees) (number of lines)
Google Maps 28 87 0.12 0 3
Live Search Maps 28 91 0.52 0.37 2
Yahoo Maps 27 82 0.35 0.24 5
MapQuest Maps 25 83 0.69 0.55 0
OpenStreet Maps 31 95 0.37 0.29 3
(a) MapQuest Maps (thinning operator)
(b) MapQuest Maps (this paper)
Figure 15: Experimental results using the approach
from this paper and our previous work using the
thinning operator
corner points into groups. The centroids of the classified
groups are the road intersections. Without tracing the road
lines, false-positive corner points or intersections of T-shape
roads significantly shift the centroids away from the correct
locations. In comparison, our approach explicitly traces the
road lines for accurate positions of the road intersection tem-
plates and accurate orientations of the intersecting roads.
To extract text and lines from documents or maps, the
research on text/graphics separation [1, 5, 11, 12] performs
geometrical analyses of the foreground pixels or exploits the
differences between line and character textures to achieve
their goals. Cao et al. [1] detect characters from maps where
characters overlap lines using the differences in the length
of line segments of characters and lines. Nagy et al. [11, 12]
first separate the characters from the lines using connected
component analysis and then focus on local areas to rebuild
the lines and characters using various methods. Our other
work in [5] detects the line and character pixels by exploiting
the differences between line and character textures in the
frequency domain.
Furthermore, Khotanzad et al. [9] utilize a color segmenta-
tion method and exploit the geometric properties of contour
lines such as that contour lines never intersect each other to
extract the contour lines from scanned USGS topographic
maps. Chen et al. [4] latter extend the color segmenta-
tion method from Khotanzad et al. [9] to handle common
topographic maps (i.e., not limited to USGS topographic
maps) using local segmentation techniques. However, these
text/graphics separation and line extraction approaches do
not perform further analysis of the lines to determine the lo-
cations of the line intersections and orientations; hence they
provide only partial solutions to the problem of extracting
the road intersection templates.
6. CONCLUSION AND FUTURE WORK
In this paper, we proposed a technique to automatically
extract road intersection templates from raster maps. Our
experiment shows efficient and accurate results with the po-
sitional offset as 0.4 pixels and orientation offset as 0.24 de-
grees on average. The result is a set of road intersection
templates, which consists of the positions and connectivities
of the road intersections and the orientations of the roads
that intersect at the intersection for each raster map. With
the accurate road intersection templates, conflation applica-
tions can significantly reduce the possible matching candi-
dates during the point matching process by using the road
orientations as one feature to select possible matches. More-
over, applications that work on imagery to extract roads can
also benefit from the accurate road intersection templates to
enhance the quality of their extraction results.
We plan to extend our work to fully vectorize the road
lines from the raster maps. For example, we plan to discover
the relationship between the road intersections as well as to
extract feature points from the road lines other than road
intersections, such as corner points on curve roads. The
fully vectorized road network then can be used to match
with other vector data to fuse the raster map with other
geospatial data even more efficiently.
7. ACKNOWLEDGMENTS
This research is based upon work supported in part by the
University of Southern California under the Viterbi School
Doctoral Fellowship, in part by the National Science Foun-
dation under award number IIS-0324955, in part by the
United States Air Force under contract number FA9550-08-
C-0010, and in part by a gift from Microsoft.
The U.S. Government is authorized to reproduce and dis-
(a) Google Maps (this paper)
(b) Thinning operator (c) This paper
Figure 16: Detail view of various road intersection
templates
tribute reports for Governmental purposes notwithstanding
any copyright annotation thereon. The views and conclu-
sions contained herein are those of the authors and should
not be interpreted as necessarily representing the official
policies or endorsements, either expressed or implied, of any
of the above organizations or any person connected with
them.
8. REFERENCES
[1] R. Cao and C. L. Tan. Text/graphics separation in
maps. In Proceedings of the Fourth International
Workshop on Graphics Recognition Algorithms and
Applications, pages 167–177, Springer-Verlag, London,
UK, 2002.
[2] C.-C. Chen, C. A. Knoblock, and C. Shahabi.
Automatically conflating road vector data with
orthoimagery. GeoInformatica, 10(4):495–530, 2006.
[3] C.-C. Chen, C. A. Knoblock, and C. Shahabi.
Automatically and accurately conflating raster maps
with orthoimagery. GeoInformatica, 12(3):377–410,
2008.
[4] Y. Chen, R. Wang, and J. Qian. Extracting contour
lines from common-conditioned topographic maps.
IEEE Transactions on Geoscience and Remote
Sensing, 44(4):1048–1057, 2006.
[5] Y.-Y. Chiang and C. A. Knoblock. Classification of
line and character pixels on raster maps using discrete
cosine transformation coefficients and support vector
machine. In Proceedings of the 18th International
Conference on Pattern Recognition, pages 1034–1037,
2006.
[6] Y.-Y. Chiang, C. A. Knoblock, and C.-C. Chen.
Automatic extraction of road intersections from raster
maps. In Proceedings of the 13th Annual ACM
International Symposium on Geographic Information
Systems, pages 267–276, 2005.
[7] Y.-Y. Chiang, C. A. Knoblock, C. Shahabi, and C.-C.
Chen. Automatic and accurate extraction of road
intersections from raster maps. Geoinformatica, 2008.
[8] A. F. Habib and R. E. Uebbing. Automatic extraction
of primitives for conflation of raster maps. Technical
report, The Center for Mapping, 1999.
[9] A. Khotanzad and E. Zink. Contour line and
geographic feature extraction from USGS color
topographical paper maps. IEEE Transactions on
Pattern Analysis and Machine Intelligence,
25(1):18–31, 2003.
[10] G. Koutaki and K. Uchimura. Automatic road
extraction based on cross detection in suburb. In
Proceedings of the SPIE, Image Processing:
Algorithms and Systems III, volume 5299, pages
337–344, 2004.
[11] L. Li, G. Nagy, A. Samal, S. Seth, and Y. Xu.
Cooperative text and line-art extraction from a
topographic map. In Proceedings of the Fifth
International Conference on Document Analysis and
Recognition, pages 467–470, 1999.
[12] L. Li, G. Nagy, A. Samal, S. C. Seth, and Y. Xu.
Integrated text and line-art extraction from a
topographic map. International Journal of Document
Analysis and Recognition, 2(4):177–185, 2000.
[13] W. K. Pratt. Digital Image Processing: PIKS
Scientific Inside. Wiley-Interscience, 3rd edition, 2001.
[14] J. Shi and C. Tomasi. Good features to track. In
Proceedings of the Computer Vision and Pattern
Recognition, pages 593–600, 1994.
[15] G. W. Zack, W. E. Rogers, and S. A. Latt. Automatic
measurement of sister chromatid exchange frequency.
Journal of Histochemistry and Cytochemistry,
25(7):741–753, 1977.
