See discussions, stats, and author profiles for this publication at: [Link]

net/publication/259481902

BENCHMARK EVALUATION OF HOG DESCRIPTORS AS FEATURES FOR

CLASSIFICATION OF TRAFFIC SIGNS

Article in International Journal for Traffic and Transport Engineering (IJTTE) · December 2013
DOI: 10.7708/ijtte.2013.3(4).08

CITATIONS READS

16 2,132

2 authors, including:

Hasan Fleyeh
Dalarna University
81 PUBLICATIONS 1,113 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Smart Parking View project

Retroreflection material of traffic signs View project

All content following this page was uploaded by Hasan Fleyeh on 30 December 2013.

The user has requested enhancement of the downloaded file.

 Fleyeh H. et al. Benchmark Evaluation of HOG Descriptors as Features for Classification of Traffic Signs
UDC: 656.1.054/.057 DOI: [Link]

BENCHMARK EVALUATION OF HOG DESCRIPTORS AS FEATURES

FOR CLASSIFICATION OF TRAFFIC SIGNS
Hasan Fleyeh1, Janina Roch2
1
Department of Computer Engineering, School of Technology and Business Studies, Dalarna University,
Sweden
2
Department of Business Studies and Economic, TU Kaiserslautern, Kaiserslautern, Germany
Received 20 June 2013; accepted 23 September 2013

Abstract: The purpose of this paper is to analyze the performance of the Histograms of Oriented
Gradients (HOG) as descriptors for traffic signs recognition. The test dataset consists of speed
limit traffic signs because of their high inter-class similarities. HOG features of speed limit signs,
which were extracted from different traffic scenes, were computed and a Gentle AdaBoost classifier
was invoked to evaluate the different features. The performance of HOG was tested with a dataset
consisting of 1727 Swedish speed signs images. Different numbers of HOG features per descriptor,
ranging from 36 features up 396 features, were computed for each traffic sign in the benchmark
testing. The results show that HOG features perform high classification rate as the Gentle AdaBoost
classification rate was 99.42%, and they are suitable to real time traffic sign recognition. However,
it is found that changing the number of orientation bins has insignificant effect on the classification
rate. In addition to this, HOG descriptors are not robust with respect to sign orientation.

Keywords: traffic signs, HOG descriptors, Gentle AdaBoost, classification.

1. Introduction Swedish speed signs, as one class of traffic

signs, have identifying features which make
Traffic sign recognition is an important them distinguishable from other objects
component of Advanced Driver Assistant and increase the intra-class variability
Systems (ADAS). Recognizing traffic signs with other traffic sign classes. The colors
gives important information to the driver as well as the shape and the digits give
and helps to identify changing conditions important information (Fig. 1). Therefore,
and understand complex traffic situations. they represent the basis to recognize the
Early recognition of speed limit signs can different types of speed signs. However,
help a driver to adjust the speed of the all speed limit signs have a round surface
vehicle in time which represents a potential with a red border, a yellow fill and black
to not only reduce pollution, but also the digits. This inter-class similarity makes the
number of accidents which will save lives classification of this class of traffic signs a
(Fleyeh, 2004). challenging issue.

1
Corresponding author: hfl@[Link]

448
International Journal for Traffic and Transport Engineering, 2013, 3(4): 448 - 464

Fig. 1.
Speed Limit Traffic Signs
Source: Wikipedia (2010)
Although speed signs are located in good change the perception of the color of parts or
visible spots on the side of the road where they the entire speed sign. Bad weather conditions
can be expected by the drivers (Fleyeh, 2006), such as snow, rain or fog can affect the color
they can be found in different conditions. or sharpness of the sign. Images of traffic sign
They could be damaged or mounted on a can be blurry when they are collected from a
damaged pole. Sign colors fade by aging of moving vehicle. The signs can be rotated or
the traffic sign because of environmental occluded by paint, trees or other obstacles.
effects. In the dusk or dawn the colors are A number of examples of problematic speed
perceived to be much darker. Shadows can sign images are shown in Fig. 2.

Fig. 2.
Examples for Critical Images of Speed Signs Included in the Dataset
Furthermore, speed signs are not the only which are very similar to speed limit signs.
signs with a red border, yellow fill and black Some examples of similar looking signs are
pictogram. There are a number of signs shown in Fig. 3.

Fig. 3.
Examples of Critical Images which Look Similar to Speed Sign Images

449
 Fleyeh H. et al. Benchmark Evaluation of HOG Descriptors as Features for Classification of Traffic Signs

Classifying this group of traffic signs, i.e. the HOG descriptors were invoked in different
speed limit traffic signs, which is characterized applications other than human detection
by high inter-class similarity and low intra- which it was originally designed for.
class variability needs highly sophisticated
feature descriptors. Among these descriptors Dong et al. (2010) developed a HOG-based
is the Histograms of Oriented Gradients multistage approach for object detection and
(HOG) which was introduced by Dalal and object pose recognition for service robots.
Triggs (2005) as features for pedestrian This approach made use of the merits of
recognition. However, it has shown that it both multi-class and bi-class HOG-based
is capable of describing other objects than detectors to form a three-stage algorithm
pedestrians (Dong et al., 2010; Zhang et al., at low computing cost.
2007; Zhang, 2010; Bertozzi et al., 2007;
Zaharescu et al., 2009; Rybski et al., 2010). Zhang et al. (2007) proposed a multi-
resolution framework which was based
The main contribution of this paper is to on HOG descriptors for general object
investigate the suitability of this descriptor to detection. Different resolutions were
achieve traffic sign recognition. Performance represented using a coarse-to-fine feature
manifestation conducted by this paper hierarchy. During detection, the lower
include high recognition rates, real-time resolution features were initially used to
implementation, many traffic sign categories reject the majority of negative windows at
as recognition objects, robustness for variant relatively low cost, leaving a relatively small
environments, and feasibility under poor number of windows to be processed in higher
visibility conditions. A benchmark test was resolutions. The lowest resolution classifier
conducted by training a Gentle AdaBoost was able to reject almost 80% of the windows
(Friedman et al., 2000) as a multiclass with a detection rate of 98.6%. W hen
classifier with a set of HOG features computed combined, the first three low resolutions
for this set of traffic signs. The classifier was rejected more than 99.9% of the windows,
trained and tested with different numbers of with a detection rate of almost 94%.
HOG features and the behavior of the classifier
should reflect the ability of HOG descriptors Zhang (2010) proposed an off-line signature
to represent this group of traffic signs. verification and identification method by
invoking a local shape descriptors pyramid of
The remainder of the paper is constructed so HOG. The signature image was divided into
that in Section 2 the related work is presented. a number of non-overlapped regions in each
Section 3 demonstrates the system setup and level of the pyramid. It starts by the whole
in Section 4 the experiments conducted on the signature image as one region, then 4 regions,
HOG descriptors are illustrated and discussed. and 16 regions and so on. HOG features are
Finally, the conclusions are drawn in Section 5. computed for each cell in the level and merged
together to represent the whole signature. A
2. Related Work binary SVM classifier was invoked for the
purpose of verification and multi-class SVM
In recent years, research in object recognition was invoked for identification. The system
has grown rapidly because of the real need achieved FRR of 4.0% and FAR 3.25% for
for such systems in future applications. skilled forged signatures.

450
International Journal for Traffic and Transport Engineering, 2013, 3(4): 448 - 464

Bertozzi et al. (2007) invoked HOG and proposed system was trained with a set of
SVM classifier to filter the region of interest (HOG) descriptors to recognize different
(ROI) gathered by visible and far infrared orientations of vehicles detected in imagery.
cameras for the purposes of pedestrain The HOG orientation-specific classifiers
detection. The system was able to classify achieved 88% classification accuracy on a
91% of the pedestrains in the ROI areas. test database of 284 images.

Zaharescu et al. (2009) proposed a 3D 3. Experiment Setup

feature detector (MeshDOG) and a 3D
feature descriptor (MeshHOG) for uniformly The block diagram of the proposed approach
triangulated meshes, invariant to changes in is illustrated in Fig. 4. The candidate
rotation, translation, and scale. The descriptor speed limit traffic sign is extracted by the
was able to capture the local geometric and/or procedure described in subsection 3.1.
photometric properties in a succinct fashion. The HOG descriptors of this object are
computed and invoked either for training
Rybski et al. (2010) developed vision- of the classifier or in the prediction of the
based algorithms for determining vehicle speed limit traffic sign when the trained
orientation of vehicles in images. The classifier is used for this purpose.

Fig. 4.
Block Diagram of the System Setup

451
 Fleyeh H. et al. Benchmark Evaluation of HOG Descriptors as Features for Classification of Traffic Signs

3.1. Object Extraction 1. Fill the red region which represents a

candidate sign (subimage A) with white
Extracting speed limit traffic signs starts by pixels in order to produce (subimage B).
exploiting a shadow and highlight invariant
color segmentation algorithm (Fleyeh, 2006) 2. Extract the area corresponding to the
to extract all red objects in the image. The pictogram of the sign by XOR operator
second column of Fig. 5 depicts the results between subimage A and subimage B,
of the segmentation of a number of traffic i.e. (subimage C = subimage A XOR
scenes. Binary objects in the segmented image subimage B). The resulting area will
were labeled using connected components correspond to the pictogram of the
labeling. All objects with red rims, yellow traffic sign.
interiors, and appropriate dimensions were
selected for further investigation. 3. Extract the corresponding pictogram and
convert it into grey level (subimage D).
The process of extracting the object under
consideration f rom the image can be 4. Compute the features of this pictogram
described in the following steps: for training and testing.

A B C D
Fig. 5.
Steps of Object Extraction

452
International Journal for Traffic and Transport Engineering, 2013, 3(4): 448 - 464

3.2. Histogram of Oriented Gradients shadows, the local histogram is accumulated

(HOG) over a larger area called blocks. To improve
the contribution of the cells in the final
Edges play a very important role in computer image descriptor, overlapping between these
vision and their orientations describe an cells is invoked.
important feature for the differentiation and
recognition of objects in the scene. HOG- Edge orientations are divided into a number
descriptors, which were introduced by Dalal of bins. These bins are equally spaced over
and Triggs (2005), make use of objects’ edges the interval 0-180 o for unsigned gradient
to create a set of features to describe the and 0-360 o for signed gradients. Edge
object under consideration. orientations should fit into one of these
bins. The histograms collected for the
To compute the HOG descriptors of any different cells in the bins in the same block
image containing the extracted object, this are concatenated to make the final set of
image is divided into a number of cells and features of the object under consideration.
a number of orientation bins as depicted in
Fig. 6. For each cell a local 1-D histogram of The result of the HOG algorithm is a discrete
the gradient directions of edge orientations amount of features which describe the input
over the pixels of the cell is collected. For image. The number of features depends on
better invariance to illumination such as the number of cells and orientation bins.

Fig. 6.
Computing the HOG Descriptors

453
 Fleyeh H. et al. Benchmark Evaluation of HOG Descriptors as Features for Classification of Traffic Signs

3.3. Gentle AdaBoost Classifier noisy data. The algorithm uses adaptive
Newton steps rather than exact optimization
Gentle AdaBoost is a boosting algorithm at each step to minimize the exponential
(Schapire, 2003) which was introduced by criterion in order to stabilize the learning
Friedman et al. (2000). It is a robust and processing.
stable version of the Real AdaBoost and
performs slightly better than the latter on Let X be a finite training set which is denoted
regular data and considerably better on by Eq. (1):

(1)

where x i is it h t ra i n i ng data, y i is its the number of training samples. The Gentle

corresponding target label, N is the space AdaBoost algorithm can be described by the
of the data set (number of features) and M is following steps:

1. Start the weights

2. Repeat for

a. Fit the regression function by weighted least-squares of yi and xi with the

weights wi .

b. Update

c. Update and renormalize.

3. Output the classifier sign

T he genera l izat ion of bi na r y Gent le available from Dalarna University (Fleyeh,

Ada Boost to its multiclass version is 2009).
straightforward. The multiclass problem
can be transformed into several binary A total of 1710 speed limit signs and 1025
classi f icat ion problems. T h is ca n be non-traffic sign objects were extracted by
accomplished by using either the One- the objects extractor described in Subsection
Against-One or the One-Against-All model. 3.1.

3.4. The Dataset The number of speed limit traffic signs in

each class is depicted in Table 1. Fig. 7 shows
The dataset which was invoked in this part of the database of the pictograms used
work comprises 1727 images and it is for recognition of the speed limit signs.

454
International Journal for Traffic and Transport Engineering, 2013, 3(4): 448 - 464

Table 1
The Dataset used to Test the HOG
Class Speed km/h Abbreviation Number of Extracted signs
0 - NOSL 1025
1 30 SL30 98
2 50 SL50 213
3 60 SL60 114
4 70 SL70 380
5 80 SL80 104
6 90 SL90 415
7 100 SL100 198
8 110 SL110 116
9 120 SL120 72
Total 2735

Fig. 7.
The Training Dataset

4. Experiments and Results cross validation experiment was employed.

In this experiment a Gentle AdaBoost
To analyze the performance of HOG features, classifier was trained and tested with
region of interests (ROI) containing the different numbers of features. The number
traffic sign candidates were extracted from of features was based on different numbers
the traffic scenes. The HOG descriptors of of cells per block and a constant number of
all 2735 objects which were detected by color orientation bins. The number of orientation
segmentation were computed without size bins was 9 while the number of cells varied
normalization or any other transformation from 2x2 to 6x6 with an increment of 1 in
of the ROI. The HOG of the candidates each direction, i.e. 2x2, 3x3… 6x6. This
were then assigned to one of the 10 classes gives 4, 9, 16, 25, and 36 cells in each block.
shown in Table 1. Table 2 depicts the classification accuracy
for the different numbers of features in the
4.1. Effect of Number of Features different folds.

In order to evaluate the effect of number of A plot of the classification error versus the
features deduced from the ROI, a 10-fold number of features indicates clearly that

455
 Fleyeh H. et al. Benchmark Evaluation of HOG Descriptors as Features for Classification of Traffic Signs

increasing the number of features deduced is essential to traffic sign recognition

per block decreases the classification error. because images or footage may contain
Fig. 8 depicts the relationship between traffic signs with different sizes depending
the classification error and the number of on the distance between the vehicle and
features. Since the extracted traffic signs the traffic sign. To have a set of descriptors
which were exploited in this experiment were which performs with scale invariance means
of different scales, it is obvious that HOG that time required for normalization can be
descriptors are scale invariant. This property saved for real time applications.
Table 2
Classification Accuracy versus Number of HOG Features
Classification Accuracy %

Features = 36 Features = 81 Features = 144 Features = 225 Features = 324

Fold 2x2 cells 3x3 cells 4x4 cells 5x5 cells 6x6 cells
9 bins 9 bins 9 bins 9 bins 9 bins
1 96.7 98.5 99.3 99.3 99.3
2 96.0 98.2 98.5 98.9 98.9
3 97.1 99.3 99.6 99.6 100
4 95.2 98.2 98.5 98.5 98.9
5 95.2 97.4 98.5 99.3 98.9
6 96.7 98.5 98.5 98.9 99.3
7 96.0 98.5 98.9 98.9 99.6
8 97.4 100 99.6 100 100
9 96.0 98.9 98.5 99.3 99.6
10 96.0 98.9 100 99.3 99.3
Ave. 96.31 98.6 99.0 99.2 99.4

Fig. 8.
Classification Error versus Number of Features

456
International Journal for Traffic and Transport Engineering, 2013, 3(4): 448 - 464

4.2. Effect of Number of Orientation Bins bins. In contrast to the variation of cells,
the var iation of the bin size did not
The effect of number of orientation bins on show major differences in the accuracy.
the classification rate was also tested. The However, a slight improvement can be
Gentle AdaBoost classifier was trained observed with 10 bins with a size of 18 o
with different numbers of HOG features each (Table 3). Fig. 9 depicts the effect
which were derived from the training set of the number of orientation bins on the
based on different numbers of orientation classification rate.
Table 3
Effect of Orientation Bins on the Classification Rate
Number of Orientation Bins Number of HOG Features Average Accuracy %
7 (6x6x7) = 252 99.34
8 (6x6x8) = 288 99.27
9 (6x6x9) = 324 99.38
10 (6x6x10) = 360 99.42
11 (6x6x11) = 396 99.23

Fig. 9.
Classification Accuracy versus Number of Orientation Bins
Deep analysis of the objects which were classifier did not generate any false positives.
misclassified showed that the classifier Common to all false negatives was the fact
committed 16 misclassifications. Among that they were blurry because of bad weather
them there was one 90 km/h speed limit sign conditions or because they were far away
which was classified as 80 km/h and all other in the background. A number of examples
errors were false negatives. However, the which were misclassified are depicted Fig. 10.

457
 Fleyeh H. et al. Benchmark Evaluation of HOG Descriptors as Features for Classification of Traffic Signs

Fig. 10.
Signs Misclassified by the Classifier

4.3. Classifying Disoriented Traffic Signs The classifier was tested with HOG features
derived from speed limit images which were
Traffic signs are usually installed on poles rotated by different angles in clockwise and
which are always vertical with respect counter clockwise directions. Fig. 11 depicts
to the ground level. However, for many a plot of the classification rate versus angle
reasons such as the nature of the ground of rotation.
or environmental effects, these traffic
signs depart from the vertical situation. T he plot show s a l a rge d rop i n t he
Two of the speed limit traffic signs shown classification rate when disoriented traffic
in Fig. 2 belong to this category. In this signs were classified with the trained
experiment, the effect of disorientation classifier, which means that HOG descriptors
(rotation) of traffic signs is studied. The are rotation variant and cannot be utilized
Gentle AdaBoost classifier was trained in these situations. As this is a crucial issue
with 360 HOG features (6x6 cells and 10 as far as traffic signs are concerned, a proper
bins). All of these features were derived solution is essential in this case to avoid this
from vertically oriented speed limit signs. kind of invariance.

458
International Journal for Traffic and Transport Engineering, 2013, 3(4): 448 - 464

Fig. 11.
Classification Rate of the Gentle AdaBoost Trained with Vertically Aligned Signs versus Angle of
Rotation of Traffic Signs
This experiment was repeated by training was the same. The classification rate did not
the Gentle AdaBoost with a set of HOG drop as traffic signs rotate. Although there was
descriptors which was derived from speed limit a slight variation in the classification rate, the
sign rotated by different angles between -90o average classification rate was 92%. The plot
to 90o. These HOG descriptors were derived of the classification rate versus the angle of
in the same manner described in the first part rotation when Gentle AdaBoost was trained
of this experiment and the number of features with rotated signs is depicted in Fig. 12.

Fig. 12.
Classification Rate of the Gentle AdaBoost Trained with Rotated Signs versus Angle of Rotation of Traffic Signs

459
 Fleyeh H. et al. Benchmark Evaluation of HOG Descriptors as Features for Classification of Traffic Signs

4.4. The Overall Performance of Gentle AdaBoost was trained with HOG
Classification descriptors of the training subset and
tested with the remainder of the images.
To test the effect of HOG descriptors This experiment was repeated for different
as discrimination features among speed numbers of HOG descriptors and each time
limit signs and non-speed limit signs, the rejection ratio was computed. Table 4
the set of images illustrated in Table 1 is depicts the ratio of rejection for the non-
invoked. The set was divided by 66.6% speed limit signs for different values of
for training and 33.3% for testing. The HOG descriptors.

Table 4
Performance Analysis of the Gentle AdaBoost based on Different Numbers of HOG Descriptors
HOG Features TP TN FP FN Misclassified
(2x2x9) = 36 543 328 1 16 24
(3x3x9) = 81 570 328 1 7 6
(4x4x9) = 144 574 328 1 7 2
(5x5x9) = 225 574 329 0 6 3
(6x6x7) = 252 574 329 0 6 3
(6x6x8) = 288 577 329 0 5 1
(6x6x9) = 324 575 329 0 6 2
(6x6x10) = 360 574 329 0 8 1
(6x6x11) = 396 575 329 0 5 3

I n order to eva luate t he overa l l number of true negatives (non-speed limit

performance of the classifier versus the signs classified as non-speed limit signs),
number of features involved in each and N is the number of non-speed limits
case, the geometric mean was calculated signs. Fig. 13 depicts a plot of the geometric
for each number of features as follows means versus the number of features. The
(Eq. (2)): plot indicates that increasing the number
of features beyond 225 HOG features is
not so beneficial because it burdens the
classifier and increases the computational
(2) time without obtaining better classification
performance. Each of the best classifications
Where TP is the number of the true positives generated the same misclassification. Two
(speed limit signs classified as speed limit examples of the false positives and their
signs), P is the number of speed limits predicted classes are shown in Fig. 14.
sign involved in the experiment, TN is the

460
International Journal for Traffic and Transport Engineering, 2013, 3(4): 448 - 464

Fig. 13.
Geometric Mean versus Number of HOG Descriptors

Fig. 14.
False Positives Generated by Mixed up HOG Descriptors

In order to validate this issue, the receiver is clear that the optimum features which
operating characteristic (ROC) diagram of are 225 descriptors perform as well as the
the lowest, the optimum, and the highest maximum number of features. This result
number of features was generated and is in full agreement with that obtained from
illustrated in Fig. 15. From this plot, it the geometric mean plot.

461
 Fleyeh H. et al. Benchmark Evaluation of HOG Descriptors as Features for Classification of Traffic Signs

Fig. 15.
ROC Plot of the Gentle AdaBoost Trained with the Lowest Number of Features (36 HOGs ), Highest
Number of Features (396 HOGs ), and Best Number of Features (225 HOGs )

4.5. Training and Testing Time speed limit sign using different numbers of
HOG descriptors are depicted in Table 5.
This experiment is accomplished using While training time increases almost linearly,
Macbook 6.1 w it h I ntel Core 2 Duo, the testing time is almost constant regardless
2.26 GHz processor and the code was of the number of features. The increment
implemented using Matlab R2012a (maci64) in the testing time between a low number
version. Timings of training the Gentle of features and a high number of features is
AdaBoost classifier with 2735 descriptors not crucial. Fig. 16 depicts a plot of training
together with the classification time of a and testing times versus number of features.

Table 5
Time Required for Training and Testing of Different Numbers of HOG Descriptors
Number of HOG Features Time of Training (sec.) Time of classification (msec.)
36 2.67 13.8
81 5.83 15.2
144 10.16 15.4
225 21.02 16.2
324 22.77 16.8

462
International Journal for Traffic and Transport Engineering, 2013, 3(4): 448 - 464

Fig. 16.
Training and Testing Time of the Classifier with HOG Descriptors

5. Conclusion According to the experiments a slight

change in the number of orientation bins
Histograms of Oriented Gradients (HOG) is not effective and does not improve the
are descriptors which were developed for classification rate. Finally it is important to
human detection. However, they have mention that HOG descriptors are variant
been exploited in other fields of computer against angle of rotation. Disoriented traffic
vision and object detection. This paper signs might not be correctly classified.
invest igated t he su itabi l it y of t hese Unless the classifier is trained with rotated
descriptors for traffic sign recognition and signs, the classifier trained with HOG fails
classification. For this purpose, a dataset to classify those disoriented traffic signs.
of speed limit traffic signs were invoked.
The major outcomes of this benchmark
test are that HOG descriptors are suitable
for traffic sign recognition because they are
easy to implement, characterized with high
robustness against intra-class variability,
and fast enough for real time detection and
classification. The descriptors are scale
invariant which means that traffic signs
can be extracted and classified without any
need for size normalization. In addition,
the number of descriptors needed for
classification is not more than 225 which
a lso suppor t rea l t i me appl icat ions.

463
 Fleyeh H. et al. Benchmark Evaluation of HOG Descriptors as Features for Classification of Traffic Signs

References Wikipedia. 2010. Road signs in Sweden. Available from

Internet: <[Link]
Bertozzi, M.; Broggi, A.; Rose, M.D.; Felisa, M.; in_Sweden>.
Rakotomamonjy, A.; Suard, F. 2007. A pedestrian
detector using histograms of oriented gradients and a Zaharescu, A.; Boyer, E.; Varanasi, K.; Horaud, R.
support vector machine classifier. In Proceedings of the 2009. Surface feature detection and description with
Intelligent Transportation Systems Conference, ITSC 2007. applications to mesh matching. In Proceedings of the IEEE
IEEE. 143-148. Conference on Computer Vision and Pattern Recognition,
CVPR 2009. 373-380.
Dalal, N.; Triggs, B. 2005. Histograms of oriented
gradients for human detection. In Proceedings of the IEEE Zhang, B. 2010. Off-line signature verification
Computer Society Conference on Computer Vision and Pattern a nd ident i f ic at ion b y p y r a m id h i s tog r a m of
Recognition, CVPR 2005. 886-893. oriented gradients, International Journal of Intelligent
Computing and Cybernetics. DOI: [Link] x .doi.
Dong, L.; Yu, X.; Li, L.; Hoe, J. 2010. HOG based multi- org/10.1108/17563781011094197, 3(4): 611-630.
stage object detection and pose recognition for service
robot. In Proceedings of the 11th International Conference on Zhang, W.; Zelinsky, G.; Samaras, D. 2007. Real-time
Control Automation Robotics & Vision (ICARCV). 2495-2500. accurate object detection using multiple resolutions.
In Proceedings of the IEEE 11th International Conference on
Fleyeh, H. 2009. Traffic Sign Database. Available from Computer Vision, ICCV 2007. 1-8.
Internet: <[Link]

Fleyeh, H. 2004. Color detection and segmentation for

road and traffic signs. In Proceedings of the IEEE Conference
on Cybernetics and Intelligent Systems, Singapore. 808-813.

Fleyeh, H. 2006. Shadow and Highlight Invariant

Colour Segmentation Algorithm for Traffic Signs.
In Proceedings of the IEEE Conference on Cybernetics and
Intelligent Systems, Bangkok, Thailand. 108-114.

Friedman, J.; Hastie, T.; Tibshirani, R. 2000. Additive

logistic regression: a statistical view of boosting,
Annuals of Statistics. DOI: [Link]
aos/1016218223, 28(2): 337-407.

Rybski, P.; Huber, D.; Morris, D.; Hoffman, R. 2010.

Visual classification of coarse vehicle orientation using
histogram of oriented gradients features. Presented at
the IEEE Intelligent Vehicles Symposium (IV). 921-928.

Schapire, R. 2003. The boosting approach to machine

learning: An overview. Lecture Notes in Statistics-New
York-Springer Verlag. 149-172.

464

View publication stats