Professional Documents
Culture Documents
=
=
N
i
c
i
t
i
c t
h h H H E
1
2
) (
,
, (6)
where N is number of bins in histogram,
( ) c
i
h and
( ) t
i
h
are i-th bins in candidates and templates histogram
respectively.
If value of E was less than some predefined
threshold, candidate was decided to be circle or triangle
(depending on template); otherwise, candidates were
discarded.
For the second group candidates that step was not
applied.
Table 1. Background shape histograms
Sign
Vertical
histogram
Horizontal
histogram
Extracting black regions
Calculating centroid and area
Template matching
S
h
a
p
e
a
n
a
l
y
s
i
s
Candidate
mask
Sign description
T
e
m
p
l
a
t
e
d
a
t
a
b
a
s
e
Fig.7. Recognition of first-group signs.
(a)
(b)
(c)
(d)
(e)
(f)
Fig.6. Sign detection example:
(a,b) input images; (c,d) candidate regions for first and second group signs; (e,f) candidates selected after labeling
and applying constraint rules.
4.3 Extracting meaning part
Every sign of the first group can be divided into the
following parts: border (red), inner part background
(white or yellow) and meaningful part (black region
inside the sign). Color properties of border were used to
detect sign candidate, shape properties - to classify
candidate as restricted or warning. But to understand the
meaning of sign shape of black region has to be
analyzed [5]. As far as color has no meaning in that
analysis each candidate can be represented as a binary
mask. To extract black regions the following method
was applied:
1. Mark all red pixels as 0.
2. If sign is warning mark all pixels in which
G > B*1.1 and R > G*1.1 as 0; otherwise mark
all pixels for which min(R,G,B) > (average
intensity)*1.3 as 0. (Threshold values were
selected experimentally)
3. Mark all other pixels as 1.
4. Recursively look through all neighbors of each
pixel marked as 1. If distance in color space
is less than some predefined threshold mark
pixels as 1;
Result: binary mask
( ) i
x
C
50 50
,
{ } 0 , 1
) (
,
=
i
y x
C
4.4 Template matching
Information from shape analyses was used to select
template group for matching. Each template mask
(
) (
50 50
j
x
T , j-template index) was formed by the following
way value of each pixel was weighted due to number
of its nonzero neighbors.
8
,
nonzero j
y x
N
T =
, where
nonzero
N
is number of nonzero
neighbors of pixel at x,y.
For the more robust results, matching was processed
relative to centroid coordinates of template and
candidate (Fig.5.).
Fig.8. Overlaying using centroid coordinates
Matching problem for candidate
i
C was formalized as
follows (see Fig.9.):
( ) ( )
(
(
|
|
.
|
\
|
+ =
i j
j
i j
i
C T error C T error matchID , 2 , 1 min arg
,
where
( )
( )
=
=
j j
j j
T y x
i j
T y x
i i j
T y x
i j
T y x
j i j
C T C C T error
C T T C T error
, ,
, ,
, 2
, 1
(7)
Result of this step is identifying number of template
which is best matching the candidate.
Fig.9. Calculating matching error.
4.5 Recognition of second-group signs
Some approaches for recognition of information signs
shown at [1,2,4,6]. Our algorithm is shown in Fig.10.
To understand the meaning of information sign we
separate it into basic components: arrows and text, and
then analyze positional relationship between those
segments. Main part of algorithm is candidate
segmentation. On this step structure of sign was
extracted and analyzed in order to distinguish symbolic
information (arrows, road numbers, etc) from text.
5. EXPERIMENTAL RESULTS
Real images were used for experiments. Images were
taken in different lighting and weather conditions in
order to analyze robustness of proposed algorithm.
Results of detection step are presented in Tables 2~3.
Table 2. Average detection results.
Warning and
restricting signs
Information
signs
Number of
images
124 95
Number of
signs
172 119
Not
detected
4
(2.3%)
1
(0.8%)
Fig.10. Recognition of information signs
Labeled symbol mask
Detect arrow region
Detect arrowheads
E
x
t
r
a
c
t
d
i
s
t
a
n
c
e
D
e
t
e
c
t
r
o
a
d
n
u
m
b
e
r
s
Separate sign candidate
into clusters
(Number of clusters equal
to number of arrowheads)
Text clusters Distance to
crossroads
Road
numbers
Arrow region
Arrow
region
Number and
positions of
arrowheads
Table 3. Recognition results for first-group signs.
Number of images 124
Number of signs 172
Not recognized 11 (6.4%)
False-positive recognition 4 (2.3%)
All experiments were done on Pentium-IV 2.6 with
512MB RAM under Borland Builder 6.0 environment.
Images of size 640x480 were used. Computation time
was measured by means of C++ to hundreds of second.
Results are presented in Table 4.
Table 4. Average computation time.
Stage
Time for stage /
accumulated time
(sec)
Detection 0.01
First-group signs recognition 0.03 / 0.04
Information signs recognition 0.05 / 0.09
6. CONCLUSIONS
Experimental results indicate that the proposed
method can be used for real-time detection and
recognition of restricting, warning and information
signs. The proposed system shows good recognition rate
for various lighting and weather conditions. It provides
robust results in dark and rainy environment.
Nevertheless our system can be improved to decrease
number of false-positive detections for reducing total
computation time.
ACKNOWLEDGMENTS
The authors would like to express both financial
supports from NARC(Network-based Automation
Research Center) and post-BK(Brain Korea) 21
program which are sponsored by both of Ulsan
Metropolitan City and MOCIE(Ministry of Commerce,
Industry and Energy), and MOE(Ministry of Education
and Human Resources Development), respectively.
REFERENCES
[1] S. Azami, S. Katahara and M. Aoki, Route
Guidance Sign Identification Using 2-D Structural
Description, Proceedings of the IEEE
Intelligent Vehicles Symposium, pp. 153158,
1996.
[2] S. Azami and M. Aoki, Route Guidance Sign
Recognition, Proceedings of the IEEE Intelligent
Vehicles Symposium, pp. 338343, 1995.
[3] P. Gil-Jimenez, S. Lafuente-Arroyo, H.
Gomez-Moreno, F. Lopez-Ferreras and S.
Maldonado-Bascon, Traffic sign shape
classification evaluation. Part II. FFT applied to the
signature of blobs, Proceedings of the IEEE
Intelligent Vehicles Symposium, pp. 607612, 2005.
[4] T. Kato, A. Kobayasi, H. Hase and M. Yoneda, An
Experimental Consideration for Road Guide Sign
Understanding in ITS, Intelligent Transportation
System,. IEEE, pp. 268273, Singapore, 2002.
[5] S. Lafuente-Arroyo, P. Gil-Jimenez, R.
Maldonado-Bascon, F. Lopez-Ferreras and S.
Maldonado-Bascon, Traffic sign shape
classification evaluation I: SVM using distance to
borders, Proceedings of the IEEE Intelligent
Vehicles Symposium, pp. 557 562, 2005.
[6] J-H. Lee and K-H. Jo, Traffic sign recognition by
division of characters and symbols, Science and
Technology, Proceedings KORUS 2003. The 7th
Korea-Russia International Symposium, vol.2, pp.
324-328, 2003.
[7] P. Paclik and J. Novicova, Road sign classification
without color information, 6
th
Annual Conference
of the Advanced School for Computing and
Imaging, Lommel, Belgium, 2000.
[8] X-W. Wang, Z. Wang, J-T. Sun and H-M. Zhang,
The correlation template matching algorithm
based TD filter and ESO filter, Proceedings of
International Conference on Machine Learning
and Cybernetics, vol.9, pp. 5361-5365, 2005.