Road Sign Detection and Recognition by using Local Energy Based Shape Histogram (LESH)

Usman Zakir, Iffat Zafar & Eran A. Edirisinghe
International Journal of Image Processing, (IJIP), Volume (4): Issue (6) 567
Road Sign Detection and Recognition by Using Local Energy based Shape
Histogram (LESH)
Usman Zakir u.zakir@lboro.ac.uk
Department of Computer Science,
Loughborough University
Loughborough, LE113TU, UK
Iffat Zafar i.zafar@lboro.ac.uk
Eran A. Edirisinghe e.a.edirisinghe@lboro.ac.uk
Abstract
This paper describes an efficient approach towards road sign detection and
recognition. The proposed system is divided into three sections namely; Colour
Segmentation of the road traffic signs using the HSV colour space considering
varying lighting conditions, Shape Classification using the Contourlet Transform
considering occlusion and rotation of the candidate signs and the Recognition of
the road traffic signs using features of a Local Energy based Shape Histogram
(LESH). We have provided experimental results and a detailed analysis to justify
that the algorithm described in this paper is robust enough to detect and
recognize road signs under varying weather, occlusion, rotation and scaling
conditions.
Keywords: Road Signs, HSV, Contourlet Transform, LESH, Autonomous Vehicles
1. INTRODUCTION
Road sign detection and recognition has drawn considerable research attention in recent years
due to its challenging nature as a computer vision problem. Road signs have a direct impact on
ones daily life be it as a driver, cyclist or as a pedestrian as possible life threats can be easily
formed due to their ignorance. In recent years a number of Driver Assistance Systems have been
implemented including vision based algorithms claiming to be efficient towards road sign
detection and recognition. Generally road signs consist of three properties; firstly they are
represented by colours such as Red, Green, Blue, and Brown etc. Secondly they consist of a
particular shape such as Circular, Triangular, Octagonal, Square etc. The inner contents of the
road signs represent the third property, which may vary depending on the application of the road
sign. In this paper we have highlighted the importance of using these properties separately by
considering different problems including various lighting conditions, scaling, angular rotation and
occlusion.
The perceptual colour of a road sign appears to be different due to varying lighting conditions.
This makes it difficult to extract the accurate colour information of a sign. The weather defined by

rain, fog, snow etc., and time of the day defined by day, dusk, night etc., play an important role in
creating the above mentioned variations of illumination.
The size of a road sign as appearing in the scene has an impact on its detection and identification
accuracies. Signs that appear small will not be even detected as picking up the colour or shape
will be a challenging issue for even the best computer vision algorithms. Therefore it is important
to include a system functionality, which keeps track of the candidate sign from the point it first
become visible in the scene until a reasonable size that will enable its recognition.
Further the detection and recognition in the presence of likely angular rotation of a road sign is
also a further challenging computer vision problem which needs to be addressed and resolved.
The detection and recognition of a road sign can also be affected by occlusion, i.e., due to the
presence of objects in the field of view. In the proposed approach the verification of the road sign
is completed at multiple stages. This helps to overcome partial occlusion at a high success rate.
To accommodate the recognition ability of all categories of road signs it is important to distinguish
road signs not only by their content but also by their colour and shape information. Table 1 below
shows different categorical divisions of road signs according to their colour and shape
information.
Triangle Red Circle Hexagon Blue Circle Square
TABLE 1: Different road signs categorized according to colour and shape information
For clarity of presentation the paper is organized into the following sections: Section 2 focuses on
a literature review in the area of road sign detection and recognition. Section 3 presents the
proposed road sign detection and recognition system. It is further categorized into three important
sections A: Colour based segmentation of a road sign using the HSV colour space, B: Shape
Classification using Contourlet Transform (CT) and C: Recognition using features from the Local
Energy based Shape Histogram (LESH) and Support Vector Machine (SVM) based classification.
Section 4 illustrates the experimental setup and results. The conclusion and future work are
presented in Section 5.
2. STATE OF THE ART
Generally road sign detection and recognition literature can be divided into two groups that
concentrate on 1: Road Sign Detection and 2: Road Sign Recognition. The Road Sign Detection
consists of the procedure of extracting the candidate road sign from a scene. Majority of the work
in detection is initiated using colour information of the road sign. The segmentation of the
candidate sign from the scene is carried out by employing a colour space. The most popular
colour spaces used for this purpose are HSI, RGB, CIElab, YCbCr, CYMK and HSV. In the colour
based segmentation approaches images are first converted to a designated colour space and
then a segmentation algorithm e.g. thresholding, is applied [2]. In [1], [3], [4], colour based
segmentation of road signs have been achieved by first transforming the original image to the HSI
colour space and subsequently marking the desired colour pixels (such as Red, Green, Blue etc.)
by a white pixel. Pixels that are outside the threshold values are treated as background or noise.
Thus a binary image is formed in which white pixels represents the desired coloured area and
black pixels represent noise or background. In [5], [6], segmentation has been achieved using
RGB (Red, Green, Blue) colour space and the desired pixels are extracted by using threshold
values for each colour. These threshold values are obtained on the basis of changing illumination
conditions during different times of the day [3]. In [2], [8], [9], segmentation is performed by

transforming the images to HSV (Hue, Saturation and Value) colour space to obtain thresholds
using Otsu’s algorithm [10]. In [11], though the segmentation is performed in the HSV colour
space, pixels of interest are obtained by employing a set of fuzzy rules. In [12], CIElab colour
space is employed with a Gaussian model to target the colour information and in [13] chromatic
channels are transformed to binary images using Otsu’s algorithm. YCbCr colour space is
employed in [14] and adaptive thresholding has been performed to obtain the pixel of interest. In
[15] CIECAM97 colour space is employed to segment out the road signs and results are
compared with HSI, RGB and CIELUV colour space segmentation. Recently a number of
attempts have been made to use combined colour spaces and combined colour space models in
road sign colour detection. In [16] a joint colour space segmentation approach has been adopted.
The Hue of the HSV colour space and image chrominance (U, V) values from YUV space have
been jointly used. The results of the two colour spaces are combined by a logical AND operation.
256 RGB and HSL transforms are used in [17] to construct colour distinctions of the image by
following simple thresholding. RGB and HSI colour spaces are jointly used to threshold the image
in [18]. HSI threshold values for Blue and Red colours are tabulated. Joint colour space has been
employed [19] in which HSI is used to extract the chromatic information and RGB is used to
extract the achromatic information. Adaptive chrominance and luminance thresholding has been
achieved in [20] by employing joint colour space i.e. CIElab and HSI. In [21] a four colour space
based colour classifier has been introduced to segment road signs captured under various
weathers and lighting conditions.
In conjunction with colour based approaches, shape based approaches are also seldom used in
the segmentation of road signs. In this case the colour information is either used as a pre
processor or never used at all. Two shape measures are used for circular and triangular road
signs by using a fuzzy shape recognizer in [22]. Distance to Border (DtB) vectors are obtained of
the segmented road sign from four different views and shape has been classified by using a
linear Support Vector Machine (SVM) [4].For shape determination Hough Transform [18] and
median filter are applied to detect the triangle and circular road sign shapes [23].The circular
shapes are identified through Fast Radial symmetry detection method (FRS) and the triangular
and square shapes are identified by using Harris corner detection algorithm [10].Difference of
candidate background histogram and template histogram is used to obtain the shape information
of the candidate sign [24].Fast Fourier Transform (FFT) is employed to retrieve shape signatures
by comparing it with shape references in [25] and [26]. The 2D correlation coefficient is derived to
represent the shape similarity by correlating the candidate ROI with 2D binary template in [27]. A
coarse-to-fine scheme is proposed for candidate shape verification in [28]. Hopfield neural
networks are used to determine triangular and rectangular shapes by obtaining information
related to angle, gradient and distance between edges [29]. The RANdom Sampling and
Consensus (RANSAC) algorithm is used to examine and determine the circular shapes in [30].
The shape context and its invariance to rotation and scale are determined by employing corners
as salient points in [31]. Geometric properties of different road sign shapes are computed by
using the Affine Moment Invariants in [11].
Both the recognition and classification of road signs has been carried out by employing series of
approaches which include Multi Layer Perceptrons (MLP), Neural Networks [32] with feed forward
topology in [33]. To recognize road signs a template matching technique was proposed in [22].
Different types of Neural Networks are employed in the recognition and classification of road
signs such as LWN++ in [34], 3-layered [35], Hopfield [36], back propagation [37] and ART1 [38].
Scale Invariant Feature Transform (SIFT) is used in the recognition of road signs [39] and
similarity measures among features are obtained through objective functions [40]. SVM
Gaussian kernel is used for content recognition in [41]. A two class SVM classifier is used in [42],
one versus all SVM classifier in conjunction with RBF kernel is proposed in [43] and SVM
polynomial kernel is used with regards to AdaBoost naive Bayes in [44]. The classification of road
signs has been performed by Normalized Correlations in [45] and Normalized Cross Correlation
is introduced in [46]. Other recognition methods such as, Fuzzy Shape recognizers [11], Adaptive
Hausdorff distance based on similarity weights [47] and Gabour Wavelets [18] filters have also
been used. The next section aims to address the above research gaps in designing a robust road

sign detection and recognition system capable of performing under wide variations of illumination
and environmental conditions.
3. SYSTEM OVERVIEW
The proposed road sign detection and recognition system comprises of three stages, which are
detailed in this section. Figure 1 gives an overview of the framework.
FIGURE 1: Overview of proposed road sign Detection and Recognition framework
A. Colour Segmentation: The road sign detection and recognition system is initiated by the
colour segmentation process. This process involve the use of a colour constancy
algorithm followed by a HSV colour space based segmentation of road signs using
threshold values described in [2].
B. Shape Classification: The segmented Region of Interest (ROI) is further processed to
remove the false detections and to identify those geometric features which truly represent
Image
Database
Yes
Recognition
Shape Index
Colour Segmentation Shape Classification
Pre-Processing
Contourlet
Features
SVM
Kernel
Pre-Processing
Convert to HSV Colour Space
Aspect Ratio?
Generate a binary Image with POI
in white and background in Black
Find the location
of connected
components and
crop the sign
from original
image
Find the location
of connected
components and
crop the white
object from the
binary image.
No
Yes
No
Yes
Valid
Shape
Index?
Achromatic
Segmentation
LESH
Features
SVM
Kernel
Train
features
based on
shape
index/Class
Shape
Database
No
Match with the
threshold range of
the candidate road
sign colour

road signs. Contourlet Transforms with Support Vector Machine (SVM) kernel are applied
to measure the shape properties at this stage.
C. Recognition: The road signs have different meanings according to their contents. The
recognition stage comprises of the use of Local Energy based Shape Histogram features
with SVM kernel to recognize various contents of road signs at this stage.
3.1 Colour Segmentation
As mentioned in section 2 the segmentation process of the road sign is initiated by employing a
colour space. The possible challenges such as variation in illumination and exposure to rain/sun
light may affect the colour definition of the road signs. We have employed HSV colour space to
segment the colour of the road sign due to its ability to enhance segmentation results [2] as
compared to other available colour spaces, specifically in the outdoor environments. RGB images
taken from the camera are first passed through a pre-processing stage which helps to maintain
the colour constancy in different illumination conditions [49]. It is then transformed to HSV colour
space using the following equation [48].
( ) ( )
( ) ( )( )







−−+−
−+−
= −
BGBRGR
BRGR
H
2
1 5.0
cos (1)
( )BGR
BGR
S ,,min
3
1 





++
−= (2)
( )BGRV ,,max= (3)
Where R,G and B are representing the three channels of a RGB image. Pixels of interest (POI)
can be segmented by using all three of the above components as shown in equation 1-3 i.e. H, S
and V. The Hue (H) component contains the colour information while Saturation(S) generates
different shades of a particular colour and Value (V) indicates the brightness or darkness of the
colour components respectively. Every colour in H component of HSV colour gamut has an
angular value which varies according to change appears in S and V components. If S and V
components remain constant i.e. S = 100% and I = 100% then Red, Green and Blue colours can
be found at
o
0 ,
o
120 and
o
240 anti clock wise respectively. These values are determined by
taking the histogram of H, S and V components. In the segmentation process a desired pixel
which belongs to Red, Blue or Green colours are represented by 1 or white colour and rest of the
pixels are treated as background or 0 as shown in the Figures 2(b) and 3(b). At this stage we
have to utilize a rule based approach to keep the segmentation process as fine as possible. At
first instance median filtering and thinning are utilized for improving the segmentation
consistency. Further to this, objects are selected as road sign candidate and discarded as noise
according to their aspect ratio. A priority definition of the colours is embedded with the aspect
ratio of the candidate sign to help in selection and rejection process which is shown in Table 2.
Red colour has high priority and it can appear to be Triangle, Hexagon or Circular shape in the
scene. Blue colour can appear as rectangle and circle where as Green coloured signs can appear
as Rectangular shape in the image. The corresponding bounding box of the segmented object is
analyzed according to its centre and corner points as illustrated in the Figure 4. The segmented
objects within the aspect ratio should meet this sign criteria definition. The Figure 2(c) and 3(c)
are the results after removing objects with unsuitable aspect ratio and sign criteria.

(a) (b) (c)
FIGURE 2: Red Colour Segmentation
(a) (b) (c)
FIGURE 3: Blue Colour Segmentation
Colour Priority Shape
Red 1 Triangle, Hexagon, Circle
Blue 2 Rectangle, Circle
Green 3 Rectangle
TABLE 2: Colours with their shapes and priority
FIGURE 4: The Sign Criteria
3.2 Shape Classification
Road Signs are represented by basic geometric shapes i.e. Triangular, Circular, Hexagonal and
Square/Rectangle. This analysis of the road sign, consisting of one of the shapes mentioned
earlier is carried out after Colour Segmentation of the road signs. It not only helps to remove non-
sign objects but also lessen the burden at recognition stage.
The segmented objects resulting from Colour Segmentation stage are pre-processed at the first
stage of the Shape Classification. The pre-processing involves edge sharpening and filling the
holes inside the objects. Figure 5 shows the procedure of pre-processing stage in the Shape
Classification where as Figure 5(b) shows processed objects which is the output of the pre-
processing stage. The Contourlet Transform [50] that is successfully used in image compression
and enhancement domains has been employed here to extract the shape features of the pre-
processed objects. The shape features comprises of contour edges of an object along 2
dimensional (2D) spaces. These contours are stored for training or testing purposes in the later

Shape Classification stage. The 3 level frequency decomposition is used due to its capability of
producing diagonal direction edges. It is a very helpful technique for distinguishing circular
shapes with hexagonal shapes. The comparison of 3-level frequency decomposition with 2-level
frequency decomposition is presented in Section 4. The proposed frequency decomposition
contains ‘Haar’ attributes of Laplacian pyramid filter and a Directional filter bank, as illustrated in
Figure 6.
(a)
(b)
FIGURE 5: Shape Classification Pre-processing procedure (a) Binary Objects (b) Processed Objects
(a) (b)
FIGURE 6: Contourlet Transform 3-level frequency decomposition (a) Hexagonal Shape (b) Triangular
Shape
Pre-Processing

The extracted Contourlet features of different shapes are fed in to Support Vector Machine (SVM)
which helps to classify the exact shape of the candidate sign. The SVM classifier first proposed in
[51] is an effective and powerful method for general purpose pattern recognition and
classification. SVM based kernels are used to map the input data to the high dimensional feature
space and process it in linear form. In our experiments we have tested few popular kernels
functions, ( )vuk , defined as follows:
( ) vuvuk ,, = (4)
( ) ( )d
vuvuk 1., += (5)
( ) 2
2
, σ
vu
evuk
−
=
(6)
In most applications it has been generally concluded that, a low degree polynomial kernel shown
in equation (5) or a Radial Basis Function (RBF) as kernel shown in equation (6) works quite well.
In our case, a polynomial kernel with degree 1 and 2 provides the best results for shape
classification. The method described here for shape classification is invariant to scale, rotation
and partial occlusion problems. The method has the capability to classify any geometric shape
due to its distinguishing multi channels feature analysis.
3.3 Recognition
Once the candidate shape is classified, it initiates the process of Recognition and classification of
the road sign contents. The Recognition process comprises of the LESH features extraction of
the road sign contents and training/testing of these features is carried out by employing SVM
kernel classifier. The candidate road signs which are extracted through Colour Segmentation and
hence validated as the road sign shape by Shape Classification are further processed to obtain
the valid road sign contents for feature extraction.
The road sign contents are normally represented with black and white colours. In this context
achromatic segmentation is introduced to extract the black and white areas of the candidate road
sign from rest of the image. The objects which are meeting the criteria of aspect ratio are
considered as valid road sign contents and vice versa. The black and white objects after
achromatic segmentation and in accordance to aspect ratio are shown in Figure 7.
(a) (b) (c) (d) (e)
FIGURE 7: Road signs after achromatic segmentation
The Local Energy Model was first introduced in [52] proving that features can be extracted at
those points from an image where local frequency components represent maximum uniformity.
The extended framework of local energy model is given in Equation (8) and is normalized by the
summation of noise cancellation factor T , Sine of phase deviation and factor W , which is the
weighting of the frequency spread. Further details of this extended framework can be found in
[53].

( ) ( ) ( ) ( )( ) ( ) ( )( )( ) 
( ) ε
φφ φφ
+
−−−−
=
∑
∑
n n
n nnn
xA
TxxxxxAxW
E
sincos (8)
The Local Energy gives reliable information to extract the interest points from an image in an
invariant manner to illumination and noise. This raw energy indicates the corners, contours or
edges of underlying shape in an image. LESH [54] features are obtained firstly by dividing the
candidate image into 16 sub-regions and then the local energy information is calculated for each
sub-region along 8 different orientations with the help of Gabor Wavelets kernels [55].
( ) ( ) ( )zzIzG vuvu ,, Ψ∗= (9)
Where ( )yxz ,= represents the image position, the symbol “∗ ” is convolution operator and
( )zvu,Ψ can be calculated as equation (10).
( ) 





−=Ψ
−









−
2
2
,
22
22
,
2
2
,
,
σσ
σ
eee
k
z
zvu
ik
zvuk
vu
vu
(10)
The orientation label map is produced representing labels of orientation of pixels containing
largest energy across all scales in an image. The local histogram h is calculated in equation (12)
as follows:
( ) ( )
2
22
2
1 σ
πσ



 −+−
=
yoryxorx
r ew
(11)
Lbrbr Ewh δ××= ∑,
(12)
Where w is the Gaussian weighting function of region r calculated as in equation (11), E
represents the local energy computed as equation (8), Lbδ represents Kronecker’s delta, L
orientation label map and b current bin. From the above description it can be seen that the LESH
descriptor of a shape is 8x16=128 dimensional feature vector.
Each object shown in Figure 7(a-e) is representing a separate class of objects. These objects are
normalized into fixed size dimensions and converted to gray level images. The next stage is to
extract the LESH features of these normalized images. LESH features are obtained by computing
the local energy along each filter orientation of image sub-region and its histogram is generated.
The overall histogram represents the concatenated histograms computed along each sub-region
of the image. Figure 8 shows the LESH features representation of 30 speed limit and round about
signs respectively. These features from different classes are classified with the help of multiclass
SVM polynomial kernel which is explained earlier in section 3.2.
4. EXPERIMENTAL SETUP AND RESULTS
Using a standard photographic camera, mounted of a car, we have obtained 1200 images of
different road signs. These signs were captured during various ambient and lit-up levels of
illumination and weather conditions. The road signs collected were limited to ones with basic
colours i.e. Red, Green and Blue. Figure 9 illustrates some samples of images taken from a
Canon IXUS80IS digital camera at a resolution of 2592 x 1944. The images are pre-processed
before converting them to the HSV colour space. This involves the stabilization process of colour

FIGURE 8: LESH feature representation for (a) 30 speed limit sign (b) round about road sign
FIGURE 9: Original coloured images contain red, blue and green colour road signs captured under different
environmental conditions
constancy at a particular illumination level. As shown in Figure 1 from section 3.1 the processed
images are then transform to HSV colour space and POI are obtained using the threshold values
[2] for Red, Blue and Green coloured road signs. The output image is then later converted to
binary image in which the white pixels represents the POI and vice versa. The bounding boxes of
the white region are analyzed according to their aspect ratios and the sign criteria.
The objects which meet the criteria of being a candidate sign are cropped from both the original
image and the binary image. The location information of these cropped signs is also noted for
later usage. Figure 10 shows some results after Colour Segmentation. The objects segmented as
candidate sign are passed through the Shape Classification module where experiments are
carried out on 5 different shape classes. The inputs of the Shape Classification module are the
cropped objects from the binary image as shown in the Figure 10(c). These objects are pre
processed at first instance in this module which consists of normalization of the object in terms of
size and removal of the contents.
The objects are resized to 100x100 fixed size dimensions then algorithms for edge sharpening
and to fill the necessary holes are applied as shown in Figure 5. The shape feature selection is
done by using Contourlet Transform, which represents the object shape with respect to its
(a)
(b)

FIGURE 10: Colour Segmentation using HSV colour space (a) Road Signs in the natural environment (b)
Cropped Road Signs from the original Image (c) Cropped Road Signs from the Binary Image
Index Shape Level* Pyramidal and
Directional Filters
Trained
Instances
1 3 Haar 50
2 3 Haar 50
3 3 Haar 50
4 3 Haar 50
5 3 Haar 50
TABLE 3: Shapes considered for obtaining Contourlet features, Level* = Sub band Decomposition
Shape
Index
Classified Shapes with Level
Level Success Level Success
1 2 75% 3 100%
2 2 0% 3 100%
3 2 75% 3 100%
4 2 100% 3 100%
5 2 100% 3 100%
TABLE 4: Success percentage at different decomposition Levels
(a) (b) (c)

relevant class. Table-3 shows the details about the shape classes we have considered for our
training and testing purposes. Table-4 shows the success percentage while using different
decomposition levels on each class. It is observed that as Hexagonal shapes require more sides
to represent the shape, when using the Level 2 decomposition which only uses the horizontal and
vertical edges of the shape, the shape classifier gives the wrong interpretation of Hexagons, i.e.
they are classified as of having circular shape. The Level 3 decomposition gives directional edges
at angle π2 by considering all sides of hexagon shape. The classification of shape involves the
comparison between testing object instances having unknown class with the offline trained object
instances named by defined classes. We have performed offline training of 50 instances as per
shape class which was proven experimentally to classify any geometrical shape at significantly
high accuracy. Figure 11 and Table-4 are showing the results of Shape Classification in which it
can be observed that by using the Level 3 decomposition of Contourlet Transform the classifier
produces 100% success rate.
FIGURE 11: SVM classification results, CT = Contourlet Transform
The recognition module takes the input which is the combination of shape index and the cropped
road sign from the original image as shown in figure 1. At first stage of recognition the contents of
the road sign are captured by applying achromatic segmentation. This segmentation uses the V
component of the HSV colour space to obtain the intensities of the pixels in the image. The image
is then converted to binary image by using suitable threshold parameters and in accordance to its
pixel intensities results. The images used in our experiments are resized to 128x128 fixed sized
dimensions and later converted as gray level image. Figure 12 and 13 show LESH feature
representation of ‘Give Way’ and ‘15 Speed Limit’ road sign contents respectively. Figure 12(a-b)
and 13(a-b) are representing the original images with colour information and their corresponding
pre processed images respectively. The extracted LESH features of these images are presented
in Figure 12(c) and 13(c) respectively. The SVM classifier is trained offline on the LESH features
of 25 classes containing 25 instances of each class.
The shape index information is fed in to the trained feature database of 25 classes. This helps to
retrieve only the specific road sign content which matches with the shape index. The SVM kernel
testing set is compared with the retrieved query results from feature database. This makes the
functionally of the recognition task more robust, accurate and efficient.
Table 5 shows the recognition accuracy results during various illumination conditions. The results
show that the best accuracy figures are indicated during daylight conditions and the worst during
fogy situations. Rainy conditions do not appear to have affected the performance of the algorithm
significantly.

(a) (b) (c)
FIGURE 12: LESH feature representation for ‘Give Way’ road sign
(a) (b) (c)
FIGURE 13: LESH feature representation for ‘15 speed limit’ road sign

Ambient Illumination
Day light Raining Fog Night
Traffic Signs 560 430 70 110
Correct Recognition 552 426 61 103
False Recognition 8 4 9 7
TABLE 5: Recognition results during various levels of ambient illumination
5. CONSLUSION & FUTURE WORK
In this paper we have presented a novel approach towards road sign detection and recognition.
The system utilizes a robust method of Colour Segmentation by employing the HSV colour space
and using empirically determined threshold values suitable for various illumination conditions. A
novel Shape Classification methodology is proposed in which all possible road sign shapes are
classified by introducing the use of Contourlet Transform with SVM classifier. The Recognition
stage introduces the SVM classifier with the Local Energy based Shape Histogram (LESH)
features. We have provided experimental results to prove the effectiveness of this approach
under varying levels of illumination and environmental conditions. Overall accuracy figures of 96-
98% have been reported. We are currently working on real time application of the algorithm within
an in-car navigation system.
6. REFERENCES
1. S. M. Bascón, S. L. Arroyo, P. G. Jiménez, H. G. Moreno and F. L. Ferreras. “Road-Sign
Detection and Recognition Based on Support Vector Machines”. IEEE Transactions on
Intelligent Transportation Systems, 8(2):264-278, 2007
2. U. Zakir, A. N. J. Leonce and E. A. Edirisinghe. “Road sign segmentation based on colour
spaces: A Comparative Study”.In Proceedings of the 11th Iasted International Conference on
Computer Graphics and Imgaing, Innsbruck, Austria, 2010
3. Y.Y. Nguwi and S.Y. Cho. “Two-tier Self-Organizing Visual Model for Road Sign Recognition”.
International Joint Conference on Neural Networks, IJCNN, 2008
4. Y.Y. Nguwi and A.Z. Kouzani. “Detection and classification of road signs in natural
environments”. Neural Computing & Applications, 265–289, 2008
5. V.Andrey and K.H.Jo. “Automatic Detection and Recognition of Traffic Signs using Geometric
Structure Analysis”. SICE-ICASE International Joint Conference, Korea, 2006
6. J. Wu and Y. Tsai. “Enhanced Roadway Inventory Using a 2-D Sign Video Image
Recognition Algorithm”. Computer-Aided Civil and Infrastructure Engineering, 369–382, 2006
7. M. Benallal and J. Meunier. “Real-time colour segmentation of road signs". Canadian
Conference on Electrical and Computer Engineering, 2003.
8. P. Medici, C. Caraffi, E. Cardarelli and P.P. Porta. “Real Time Road Signs Classification”.
Proceedings of the IEEE International Conference on Vehicular Electronics and Safety
Columbus, USA., 2008
9. W.Y. Wu, T.C. Hsieh and C.S. Lai. “Extracting Road Signs using the Color Information”, World
Academy of Science, Engineering and Technology, 32, 2007.

10.C.F. Paulo and P.L. Correia. “Automatic Detection and Classification of Traffic Signs”. Eighth
International Workshop on Image Analysis for Multimedia Int, Services, Greece, 2007
11.H. Fleyeh. “Traffic Sign Recognition by Fuzzy Sets”. IEEE Intelligent Vehicles Symposium
Eindhoven University of Technology Eindhoven, Netherlands, 2008
12.L. D. Lopez and O. Fuentes. “Colour-Based Road Sign Detection and Tracking”. Springer,
Lecture Notes in Computer Science, 2007
13.G.K. Siogkas and E.S. Dermatas, Detection. “Tracking and Classification of Road Signs in
Adverse Conditions”. IEEE Mediterranean Electrotechnical Conference, MELECON, 2006
14.M.L. Eichner and T.P. Breckon. “Integrated Speed Limit Detection and Recognition from Real-
Time Video”. In Proceedings of IEEE Intelligent Vehicles Symposium ,2008
15.X. Gao, K. Hong, P. Passmore, L. Podladchikova and D.Shaposhnikov. “Colour Vision Model-
Based Approach for Segmentation of Traffic Signs”. EURASIP Journal on Image and Video
Processing, 2008.
16.W. G. Shadeed, D. I. Abu-Al-Nadi, and M. J. Mismar. “Road traffic sign detection in color
images”. In Proceedings of the 10th IEEE International Conference on Electronics, Circuits
and Systems, 2003
17.D. Matsuura, H. Yamauchi and H. Takahashi. “Extracting Circular Road Signs Using Specific
Color Distinction and Region Limitation”. Systems and Computers in Japan, 2007
18.H. Huang, C. Chen, Y. Jia and S. Tang. “Automatic Detection and Recognition of Circular
Road Sign”. IEEE/ASME International Conference on Mechtronic and Embedded Systems
and Applications, MESA, 2008
19.N. Bose, M. Shirvaikar and R. Pieper. “A Real Time Automatic Sign Interpretation System for
Operator Assistance”. Proceedings of the 38th Southeastern Symp on System Theory
Tennessee Technological Univ. Cookeville, TN, USA, 2006
20.A. V. Reina, R. Sastre, S. Arroyo and P. Jimenez. “Adaptive traffic road sign panels text
extraction”. Proceedings of the 5th WSEAS International conference on Signal Processing,
Robotics and Automation, Madrid, Spain, 2006
21.U. Zakir, I. Zafar, E. A. Edirisinghe. “A Novel Approach to In-Car Road Sign Recognition”. The
14th International Conference on Machine Design and Production, Cyprus 2010
22.R. Malik,J. Khurshid and S.N. Ahmad. “Road Sign Detection and Recognition using colour
segmentation, Shape analysis and Template Matching”. Proceedings of the Sixth International
Conference on Machine Learning and Cybernetics, Hong Kong, 2007
23.W.J. Kuo and C.C. Lin. “Two- Stage Road Sign Detection and Recognition”. IEEE
International Conference on Multimedia and Expo, 2007
24. V.Andrey and K.H.Jo. “Automatic Detection and Recognition of Traffic Signs using
Geometric Structure Analysis”. SICE-ICASE International Joint Conference, Korea, 2006
25. A. V. Reina, R. J. L. Sastre, S. L. Arroyo and P.G. Jimenez. “Adaptive traffic road sign
panels text extraction”. Proceedings of the 5th WSEAS International Conference on Signal
Processing, Robotics and Automation, Madrid, Spain, 2006

26. P. Gil-Jiménez, S. Lafuente-Arroyo, H. Gómez-Moreno, F. López-Ferreras and S.
Maldonado-Bascón. “Traffic sign shape classification evaluation II: FFT applied to the
signature of Blobs”. 607- 612 ,2005
27.J. Wu and Y. Tsai. “Enhanced Roadway Inventory Using a 2-D Sign Video Image Recognition
Algorithm”. Computer-Aided Civil and Infrastructure Engineering, 21: 2006, 369–382
28.L.W. Tsai, J.W. Hsieh, C.H. Chuang, Y.J. Tseng, K.C. Fan, C.C. Lee. “Road sign detection
using eigen colour”. IET Computer Vision 2(3): 164–177, 2008
29. G.H. Kim, H.G. Sohn and Y.S. Song. ”Road Infrastructure Data Acquisition Using a Vehicle-
Based Mobile Mapping System”. Computer-Aided Civil and Infrastructure Engineering, 2(1):
346–356, 2006
30.M.L. Eichner and T.P. Breckon. “Integrated Speed Limit Detection and Recognition from Real-
Time Video”. In Proceedings of IEEE Intelligent Vehicles Symposium, 2008
31.P. Silapachote and A. H. R. Weiss. “A Hierarchical Approach to Sign Recognition”.
Proceedings of the Seventh IEEE Workshop on Applications of Computer Vision,2005
32.Y. Y. Nguwi and A. Z. Kouzani. “Detection and classification of road signs in natural
environments”. Neural Computing & Applications, 17: 265–289, 2008
33.P. Medici, C. Caraffi., E. Cardarelli and P. P. Porta. “Real Time Road Signs Classification”.
Proceedings of the IEEE International Conference on Vehicular Electronics and Safety
Columbus, USA, 2008
34.A. Broggi, P. Cerri, P. Medici, P. P. Porta. “Real Time Road Signs Recognition”. Proceedings
of the IEEE Intelligent Vehicles Symposium Istanbul, Turkey, 2007
35.H. Ohara, I. Nishikawa, S. Miki, N. Yabuki. “Detection and Recognition of Road Signs using
simple layered Neural Networks”. Proceedings of the 9th International Conference on Neural
Information Processing, 2002
36.G. H. Kim, H. G. Sohn and Y. S. Song. “Road Infrastructure Data Acquisition Using a Vehicle-
Based Mobile Mapping System”. Computer-Aided Civil and Infrastructure Engineering, 21:
346–356, 2006
37.N. Kehtarnavaz, A. Ahmad. “Traffic Sign Recognition in Noisy Outdoor Scenes”. IEEE,
Intelligent Vehicles. 1995
38.A. D. L. Escalera., J. M. Armingol and M. Mata. “Traffic sign recognition and analysis for
intelligent vehicles”. Image and Vision Computing, 21: 247–258, 2003
39.A. A. Farag and A. E. Abdel-Hakim. “Detection, Categorization and Recognition of Road Signs
for Autonomous Navigation”. Proceedings of Acivs (Advanced Concepts for Intelligent Vision
Systems), Brussels, Belgium, 2004
40.W. G. Shadeed, D. I. Abu-Al-Nadi, and M. J. Mismar. “Road traffic sign detection in color
images”. In Proceedings of the 10th IEEE International Conference on Electronics, Circuits
and Systems, 2: 890–893,2003
41.S. Lafuente-Arroyo, S. Maldonado-Bascon, P. Gil-Jimenez, J. Acevedo- Rodriguez, and R.
Lopez-Sastre. ”A tracking system for automated inventory of road signs”. IEEE Intelligent
Vehicles Symposium,166-171, 2007

42.L. Simon, P. J. Tarel and R. Bremond. “Alerting the Drivers about Road Signs with Poor Visual
Saliency”. Proceedings of the IEEE Intelligent Vehicle Symposium, China, 2009
43.C. Kiran, L. Prabhu, R. Abdu, and K. Rajeev. “Traffic sign detection and pattern recognition
using support vector machine”. Seventh International Conference on Advances in Pattern
Recognition,2009
44.A. Z. Kouzani. ”Road-Sign Identification Using Ensemble Learning”. Proceedings of the IEEE
Intelligent Vehicles Symposium Istanbul, Turkey,2007
45.A. D. L. Escalera, J. M. Armingol, J. M. Pastor and F. J. Rodríguez. “Visual sign information
extraction and identification by deformable models for intelligent vehicles”. IEEE Transactions
on Intelligent Transportation Systems, 5(2):,2004
46.G. K. Siogkas and E. S. Dermatas. “Detection, Tracking and Classification of Road Signs in
Adverse Conditions”. IEEE Mediterranean Electrotechnical Conference, MELECON, 2006
47.Y. Y. Nguwi and S. Y. Cho. ”Two-tier Self-Organizing Visual Model for Road Sign
Recognition”. International Joint Conference on Neural Networks, IJCNN, 2008
48.M. Kalčič and J.F.Tasič. “Colour Spaces – perceptual, historical and applicational
background”. In Eurocon Computer as a Tool Conference Proceedings, 2003
49.J. van de Weijer, T. Gevers, A. Gijsenij. “Edge-Based Color Constancy”. IEEE Transactions
on Image Processing, 16(9):2207 – 2214, 2007
50.M. N. Do and M. Vetterli. “The Contourlet Transform: An efficient directional multiresolution
image representation”. 14(12): 2091 – 2106, 2005
51.C. Cortes and V. Vapnik. “Support-Vector Networks”. Machine Learning, 20(3):273-297,1995
52.M. C. Morrone and R. A. Owens. “Feature Detection from Local Energy”. PR Letters 6, 303–
313, 1987
53.P. D. Kovesi. “Phase Congruency: A Low-Level Image Invariant”. Psychological Research 64,
136–148, 2000
54.M. S. Sarfraz and O. Hellwich.” An Efficient Front-end Facial Pose Estimation System for Face
Recognition”. In International Journal of Pattern Recognition and Image Analysis, distributed
by Springer, 18(3):434–441,2008
55.S. Shiguang, C. Xilin C, and G. Wen. “Histogram of Gabor Phase Patterns (HGPP): A Novel
Object Representation Approach for Face Recognition”. IEEE Transactions on Image
Processing 16 (1): 57–68, 2007

Road Sign Detection and Recognition by using Local Energy Based Shape Histogram (LESH)

Recommended

More Related Content

What's hot (18)

Similar to Road Sign Detection and Recognition by using Local Energy Based Shape Histogram (LESH) (20)

Recently uploaded (20)

Road Sign Detection and Recognition by using Local Energy Based Shape Histogram (LESH)