In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
CIRCULAR ROAD SIGN EXTRACTION FROM STREET LEVEL IMAGES USING COLOUR,
SHAPE AND TEXTURE DATABASE MAPS
A. Arlicot *, B. Soheilian and N. Paparoditis
Institut Géographique National, Laboratoire Matis, 73, avenue de Paris, 94165 Saint-Mandé cedex, France
aurore.arlicot@polytech.univ-nantes.fr, bahman.soheilian@ign.fr, nicolas.paparoditis@ign.fr
http ://recherche.ign.fr/labos/matis/
KEY WORDS: mobile mapping system, road sign recognition, color detection, ellipse detection, pattern matching
ABSTRACT
Detection and recognition of road signs can constitute useful tools in driving assistance and autonomous navigation
systems. We aim at generating a road sign database that can be used for both georeferencing in autonomous vehicle
navigation systems and also in high scale 3D city modelling. This paper proposes a robust algorithm that can detect road
signs shape and recognizes their types.
1 INTRODUCTION
Road signs are very important features for providing rules
of navigation. Indeed, they are key landmarks when navi
gating on the roads. Their visual properties are very strong
because they have been designed to be remarkable and un-
missible objects. Road signs are thus key objects to en
rich road model databases to generate roadbooks, short
est paths, etc. The automatic detection and recognition of
road signs from images (together with objects such as road
marks) is thus a key topic and issue for road model updat
ing but also for tomorrow's applications of these databases,
i.e. driving assistance, and accurate localisation functions
for autonomous navigation. Most of the previous work in
image based road sign extraction deal with three following
issues:
• Color detection : road signs are often red or blue
with some black and white. Many authors used this
property to detect them. Often, color base rules are
defined in a color space and used for segmentation,
(de la Escalera, 1997) use RGB color space and work
with relations between the red , green and blue. Other
authors works with color spaces that are less sensitive
to lighting changes. Although the HSI (Hue, Satu
ration, Intensity) space is the most common (Piccioli
et al., 1996). More complicated color space such as
LCH (Lightness, Chroma, Hue) (Shaposhnikov et al.,
2002) and CIELAB (Reina et al., 2006) are also used.
• Shape detection: road signs forms are often rect
angular, triangular or circular. In order to strengthen
the detection, some authors propose to detect these
geometric forms within ROIs * 1 provided by color de
tection. (Ishizuka and Hirai, 2004) present an algo
rithm for circular road sign detection. (Habib and Jha,
2007) propose an algorithm for road sign forms de
tection by line fitting. An interesting measure of el-
lipticity, rectangularity, and triangularity is proposed
by (Rosin, 2003).
• T^pe recognition: It consists in recognising road
sign type using its pictorial information. It is often
*A. Arlicot is currently at Polytech’Nantes, IRCCyN lab France.
1 Region of Interest
performed by comparing the inside texture of a de
tected road sign with the textures in a database. For
this purpose different kind of algorithms are used in
the state of the art. (Priese et al., 1995) propose an
algorithm that is based on neural networks. SIFT de
scriptors are used by (Aly and Alaa, 2004). (de la Es
calera et al., 2004) used intensity correlation score as
a measure of similarity to compare the detected road
sign with a set of standard signs.
2 OUR STRATEGY
We propose an algorithm consisting in three main steps.
Diagram of Figure 1 shows the pipeline of our algorithm.
First step uses color properties of signs and perform a pre
detection (Section 3). It provides a set of ROIs in image
space. Then, an ellipse detection algorithm is applied to
detect circular shape signs within the ROIs (Section 4).
The detected shapes are considered as road sign hypothe
ses. Final step consists in validation or rejection of hy
potheses. This is performed by matching detected hypothe
ses with a set of standard circular signs of the same color
(Section 5). Results and evaluations are presented in Sec
tion 6.
3 COLOR DETECTION
A large number of road signs are blue or red. It can sim
plify their detection by looking for red and blue pixels.
However their RGB values depend on illumination condi
tions. We use HSV (Hue, Saturation, Value, see Equation
1) color space because it is robust against variable condi
tions of luminosity. In order to choose the adapted thresh
old of saturation and hue, we learn these parameters from
a set of road sign sample in different illumination condi
tions. Figure 2(a) shows our running example image and
result of blue color detection is shown in Figure 2(b). In
order to provide ROIs, connected pixels are labeled (see
Figure 2(c)). Each label defines a window in image space.
The following form detection and validation steps are per
formed within these windows.
