Faster R-CNN based Traffic Sign Detection and Classification
MONIRA ISLAM, MD. SALAH UDDDIN YUSUF
Department of Electrical and Electronic Engineering
Khulna University of Engineering & Technology
Khulna-9203, Khulna
BANGLADESH
Abstract: - Traffic sign is the key aspect in road and also for the autonomous car. Detection and classification of
these sign plays a vital role for the invention of driverless vehicles. Convolutional neural network (CNN) has the
ability to learn local features using series of convolutional and pooling layer observing the image sequences. In
this work, traffic sign detection and classification has been performed based on deep learning approach. The
experiment conducted on Germen Traffic Sign Detection Benchmark (GTSDB) and Recognition Benchmark
(GTSRB) for detection and recognition. For traffic sign detection a two-stage detector, Faster R-CNN with
ResNet 50 backbone structure is used where the CNN layers extracted the features of traffic signs from the images
and the region proposal network (RPN) filter the object from the image to create bounding box based on the
extracted feature map. The classification network classifies the traffic signs and predict the proposal confidence
score. A general deep learning model is transferred into a specific output with weights with transfer learning by
tuning the pretrained model based on COCO image dataset. The performance is compared with ResNet 152,
MobileNet v3 and RetinaNet based on the confidence score and mean average precision (mAP). Faster R-CNN
with ResNet-50 shows better detection performance comparing with other backbone structure. In addition, a
series of convolution layer with batch normalization followed by max pooling layer is used to build a classifier
and softmax is used in the output for 43 class classification and 97.89% test accuracy has been obtained.
Key-Words: - Traffic sign detection, faster R-CNN, ResNet 50 FPN, MobileNet v3 large FPN, traffic sign
classification, residual network
Received: February 17, 2021. Revised: December 13, 2021. Accepted: January 20, 2022. Published: February 10, 2022.
1 Introduction
As technology advances, traffic sign detection and
recognition become very important in case of
automatic driving system, and it is an active research
topic in real world scenario. Automatic detection of
traffic signs is essential for autonomous vehicles,
driverless car, or robotic car. Additionally, for
general traffic system it may reduce driver’s effort as
well as mitigate the risk of accident at foggy weather
or adverse environmental condition. To build a traffic
sign recognition system, traffic sign detection and
recognition are the two important stages. Traffic sign
detection and classification from the captured image
is a real world problem that has been dealt with in this
paper. For many years, the local feature descriptor
dominates all the research area of computer vision. A
series of traditional classifiers were used for object
detection that were trained with various manually
designed features obtained from histogram of
oriented gradient (HOG) or HAAR wavelet features
[1]. HOG and HAAR features with support vector
machine (SVM) are the traditional approach to deal
with recognition problems in terms of computer
vision application. SVM classifier was used to
perform 3D object detection on the image regarding
as points of a high dimensional space in [2]. These
features lack in diversity and may change with time
which limits the performance in terms of accuracy
due to motion blur, light, or other environmental
factors. With the advancement of deep learning
technology, object detection has become active field
of research that aims end-to-end optimization and
requires more computational power and cost.
Convolution neural network (CNN) based object
detection methods achieved better performance.
Currently, research on traffic sign classification
has drawn much attention in autonomous vehicles. In
literature, different deep learning methods were
applied for this task. The deep learning networks
extract the high-level features that can achieve better
performance in terms of detection and classification.
CNN-SVM based method is proposed to classify
traffic signs where the convolution layer extracts the
temporal features and SVM classifies the signs based
on the deep learning features [3]. In [1], a new
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2022.18.1
Monira Islam, Md. Salah Udddin Yusuf
E-ISSN: 2224-3488
1
Volume 18, 2022
method was proposed using HOG-SURF features and
CNN was used for classification. The HOG features
are global descriptors and SURF are rotational
invariant features. These hybrid feature built higher
dimensional feature vectors and the convolution
layers effectively extracted deep learning features
yielding 98.48% accuracy. A deep learning approach
was adopted for mobile robotics detection where
scale invariant features (SIFT) and SURF features
were used to reduce the feature dimension as PCA
algorithm [4]. They applied CNN and attained great
advantage over traditional machine learning model.
Before the adoption of CNN, various methods were
adapted for traffic-sign classification, e.g., based on
SVM [5] and sparse representations [6]. To detect
traffic signs at different environmental settings and
blur images, different complicated algorithms were
addressed which required long training time. Multi-
scale deconvolution network (MDN) with multi-
scale convolutional neural network with
deconvolution sub-network was proposed in [7] that
efficiently localized the features for detecting traffic
signs and enhanced reliability to train the model. In
[8], a fast and precise traffic sign recognition system
was proposed based on two CNN networks using
Swedish traffic sign database. One of that was for
region proposals of traffic signs and other for
classification of each region. In [9], low-cost
embedded system was developed using Raspberry Pi
module based on HOG features to detect traffic signs.
Authors in [10] used different pre-trained CNN
model to show the effectiveness of deep learning
model for medical image classification. Then they
applied fine-tuning to evaluate the importance of
transfer learning for this task. To detect objects in real
world scenario, CNN based two stage object
detection methods i.e., Region-based Convolutional
Neural Network (R-CNN), Fast R-CNN [11], Faster
R-CNN [12] were proposed that used the feature
maps to build region proposal first and then detected
the object by locating the bounding regression box.
In [13], proposal free SSD model was used based on
VGG-16 architecture. YOLO is also popular single
stage object detector that was used in [14]. R-CNN
shows better accuracy but limits the performance for
the slow processing speed. Faster R-CNN algorithm
is effective over R-CNN and Fast R-CNN in terms of
faster training and speed up the detection
performance [15].
CNN outperforms other classifiers for detecting
objects while tested on the GTSRB benchmark. A
committee of neural network [16], multi-scale CNNs
[17] with hinge loss [18] was adopted for this task
that achieved better performance than handcrafted
feature-based approach in [19]. It is observed that
CNN are inherently efficient when used in a sliding
window fashion, as many computations can be reused
in overlapping regions [20]-[21]. CNN was
demonstrated to determine an object’s bounding box
together with its class label. Another widely used
strategy for object detection using CNNs is to first
calculate some generic object proposals and perform
classification only on these candidates. Selective
search takes about 3s to generate 1000 proposals for
the Pascal VOC 2007 images [22] whereas the more
efficient Edge Boxes approach still takes about 0.3 s
[23]. In addition, it applies a deep CNN to every
candidate proposal, which is very inefficient and
time-consuming. In this paper, traffic signs are
detected based on the two-stage detector, Faster R-
CNN object detection algorithm using GTSDB
database that overcome the limitations of R-CNN and
Fast R-CNN. This two-stage detector show their
effectiveness over one-shot detector in terms of mean
average precision (mAP) and losses. Faster R-CNN
with ResNet-50 Feature Pyramid Network structure
depends on the feature activation output of each
stage’s last residual block. The focus of this study is
to track and recognize the traffic signs from the
image. Then, 2D CNN architecture is used to
recognize the traffic signs from GTSRB database to
classify 43 classes of traffic signs.
The rest of the paper is organized as follows.
Proposed methodology is described in section II and
section III shows the experimental results with
analysis. Finally, conclusion is stated in section IV.
2 Proposed Methodology
2.1 Traffic Sign Detection and Classification
Dataset
In this work, traffic sign has been localized and
classified using object detection techniques based on
deep learning. For this purpose, Germen Traffic Sign
Detection Benchmark (GTSDB) and Germen Traffic
Sign Recognition Benchmark (GTSRB) two freely
available dataset have been used. GTSDB is used to
detect the traffic sign and there are 900 images
consisting of street objects and traffic signs and the
non-target images are eliminated for further
processing. Faster R-CNN is used to detect the traffic
sign with regression box analysis and classify them
with the classifier. The other dataset GTSRB is used
containing 43997 images of traffic signs. These
images are categorized into 43 classes.
2.2 Faster R-CNN Algorithm for Object
Detection
The R-CNN algorithms are two stage object
detection methods that are all region based. In the
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2022.18.1
Monira Islam, Md. Salah Udddin Yusuf
E-ISSN: 2224-3488
2
Volume 18, 2022
first stage the model proposes a set of regions of
interests (ROI) with the region proposal network
(RPN) and in the second stage a classifier processes
the region candidates to classify the targets. In R-
CNN, category independent region proposals are
obtained via selective search. For each image region,
one forward propagation through CNN generates a
feature vector. The selective search algorithm is a
fixed algorithm that is unable to learn the features
which might generate the bad candidate region of
proposals. Additionally, the localization of candidate
box might be inaccurate for detecting the objects.
Fast R-CNN algorithm resolved some issues related
to R-CNN by processing the input image and
enhancing the processing speed. However, ROI is
generated by selective search and the region
proposals are not learnt from the training data for Fast
R-CNN. Faster R-CNN algorithm overcome the
limitations of Fast R-CNN by designing the RPN that
generates the region proposals after training. After
RPN training, the ROI pools features from the
proposals for training Fast R-CNN model with a
classification head and a regression head. RPN
provides different sized feature maps and ROI
pooling reduce the feature map by generating the
same size feature map and splits into a fixed number
(k) or in equal number and apply max-pooling on
every region. So, the output of ROI pooling is a fixed
number regardless the size of input. This algorithm
shares parameters for both the RPN and Fast R-CNN
model and trains jointly. The training speed enhances
with Faster R-CNN because time cost is less for
generating region proposals with RPN than selective
search. The RPN ranks the region boxes which are
called anchor boxes and proposes the most likely
ones that contain objects. The anchor boxes predict
the anchors to be background and foreground and
refines the anchors accordingly.
For Faster R-CNN, anchors play a vital role at a
position of an image and work well for both Pascal
VOC and COCO dataset. The ground truth box labels
the anchors according to the overlapping region. The
higher Intersection over Union (IoU) is termed as
foreground or positive class whereas the lower IoU is
termed as background or negative class. For each
anchor there exists two possible labels called logit
(foreground and background) which is feed into
softmax to predict the labels. The training dataset is
processed with feature map and labels in this way for
training a classifier. The receptive fields of every
positions on the feature map needs to cover all
anchors to provide sufficient information for
prediction. Figure 1 shows the basic architectural
framework for Faster R-CNN algorithm.
Fig.1 The framework of Faster R-CNN
architecture for object detection [4].
2.3 Feature Pyramid Network (FPN)
Faster R-CNN ROI pools features from only the
topmost feature maps. Their spatial size might be too
small for localizing small-scale objects. To detect
object, ROIs of different scales are needed to be
assigned to the pyramid levels. Feature pyramid is a
series of features with different scales, so that large
and small objects can be detected from low-
resolution and high-resolution feature maps
respectively. FPN up-samples feature maps with bi-
linear interpolation and add skip connections to
additively fuse encoder-decoder feature maps. The
feature maps are selected as following based on the
width w and height h of the ground truth box,
(1)
where, denotes the scale of topmost feature maps
and k denotes the layer of FPN to generate the
feature patch.
2.4 Loss Function
For training RPN, training objective need to
determine, and positive and negative label need to
determine. The anchors are assigned as positive label
if any of the two condition is satisfied. (i) The anchor
that has higher IoU overlap with the ground truth
box; (ii) The anchor that has IoU overlap higher than
0.7 is termed as positive label. If IoU ratio of any
anchor is lower than 0.3 for all ground truth box, it is
assigned as negative label. The anchor that is not
labelled as positive or negative, do not contribute to
training. The loss of the classifier comprises the
classification and regression loss. Equation 2
represents the loss function for Faster R-CNN
algorithm.
(2)
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2022.18.1
Monira Islam, Md. Salah Udddin Yusuf
E-ISSN: 2224-3488
3
Volume 18, 2022
where, i is the index of anchor in a mini-batch. is
the predicted probability of an anchor for being a
target object and
is the ground truth label. is a
vector that represents 4 parameterized co-ordinates of
predicted bounding box and
is the ground-truth
bounding box. is the log loss classification loss
and , ,
is the regression loss activated only
when the anchor contains an object (i.e., ground-truth
). is the target of regression layer consisting
of 4 variables , , , . Here, is the
coordinate of the bounding box center and h and w is
the height and width of the box, respectively. The
target of the regression box
can be obtained by
parameterization of the four coordinates.
(3)
2.5 Faster R-CNN with ResNet FPN for
Object Detection
For object detection, deeper network like ResNet 50,
ResNet 152 bottleneck architecture will be used that
contains 4 stages and performs the initial convolution
and max- pooling using and kernel sizes
respectively. For each stage, the network contains 3
residual blocks having 3 layers stacked one over
other. The 3 layers are ,
convolution. The FPN extract the feature maps and
feed to the RPN. In FPN, for each scale level,
convolution filter is applied over the feature maps
followed by convolution for object prediction
and boundary box regression called RPN head.
The convolution layer is responsible for
reducing and then restoring the dimensions and the
layer is left as a bottleneck with smaller
input/output dimensions. With the progress of each
layer the input will be reduced to half whereas
channel width will be doubled. Finally, the network
has an average pooling layer followed by a fully
connected layer for object detection. Figure 2 shows
the basic structure of FPN with RPN implemented on
ResNet architecture.
2.6 Traffic Sign Detection with Faster R-
CNN
In this work, traffic sign is detected from the image.
Faster R-CNN performs better for PASCAL VOC
dataset for detecting 20 objects. But detecting small
objects like traffic sign from images even on different
environmental condition like motion or blur. So, the
Fig.2 FPN with RPN based on ResNet architecture
parameters need to adjust to achieve for better
performance. 43 classes of traffic signs were in
GTSDB. The model was pretrained with COCO
dataset to initialize the parameters of RPN network
and Fast R-CNN. Faster R-CNN is trained by
GTSDB dataset which have fewer levels than COCO
dataset. When the number of datasets is limited then
transfer learning is required [24]. The RPN network
generates the region proposals and non-maximum
suppression strategy is applied to score the anchors
and passes the positive and negative labels to RoI
pooling layer. After training, the loss of classifier and
regressor is obtained and the confidence score shows
the object detection performance in terms of testing
image according to the amount of input.
2.7 CNN for classification
A 2D CNN model is built for classification. CNN is
better due to the convolution of 2D kernel that
effectively extracts the discriminatory temporal
features for classifying the objects [25]. In this work,
three Convolution layer with Relu activation
function, followed by Batch normalization, max-
pooling and dropout layer are used for building the
CNN model for recognition of traffic sign using
GTSRB dataset. The batch size is 32 with learning
rate 0.0001. Back propagation algorithm is used for
effective training that adjust the parameters
minimizing the cost function. These updated
parameters can be obtained from chain rule through
the gradient of each parameter with activation
function and input vector in each neuron. The 2D
convolution extracts the features in time domain that
classify the traffic signs efficiently.
3 Analysis of Experimental Result
In this work, traffic signs are detected trained on
Faster R-CNN model based on GTSDB. The non-
target images are rejected, and 506 images are used
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2022.18.1
Monira Islam, Md. Salah Udddin Yusuf
E-ISSN: 2224-3488
4
Volume 18, 2022
for analysis where the images size are
pixels. 406 images are used for training and 100 for
testing. The experiment is performed on Google
Colab dedicated GPU platform. The Resnet 50
architecture is used to detect the object and the
performance is compared with MobileNet or
RetinaNet or shallow ResNet 152 architecture. mAP
and the loss of classifier, regressor are observed as
the performance indicator of each method.
3.1 Performance with Faster R-CNN
based on Resnet FPN Architecture
The experiment is performed to detect traffic sign
detection using GTRDS database with Faster R-CNN
which overcome the limitations of R-CNN and Fast
R-CNN. The ResNet 50 FPN and ResNet 152 FPN
shallow layers are implemented as network structure.
Stochastic gradient descent (SGD) is used as
optimizer with weight decay and 0.0005 initial
learning rate. Cosine annealing warm restart learning
scheduler is used. As number of iterations increase
the learning rate are reduced. The training and testing
time per iteration are 0.754 sec and 0.157 sec
respectively whereas required time for accumulating
evaluation result is 0.06 sec.
The results of Faster R-CNN with backbone
Resnet architecture are shown as bounding box to
capture the traffic sign and the confidence score. The
bounding box (bbox) is attained from RPN which
allows to select the appropriate ROI and the
confidence score measures the fitness of the box. The
bbox indicates the localization of the predicted box
with a class label. Figure 3 depicts the distribution of
number of traffic signs with the class number. Figure
4 and Figure 5 show the confidence score of detected
traffic signs in images with ResNet 50 and 152
network architecture respectively. Figure 4 shows
that Faster R-CNN with ResNet 50 can detect the
traffic signs effectively at foggy street, different
lighting state or normal condition attaining equal or
greater than 0.96 score at all conditions even with
multiple signs at single frame. Figure 4(b) shows that
Fig. 3 Distribution of number of traffic sign with
class number of GTSDB dataset
(a)
(b)
(c)
(d)
Fig. 4 Traffic Sign detection with Faster R-CNN
based ResNet 50 FPN architecture, at (a) foggy
street; (b) different lighting condition; (c) multiple
traffic sign detection; (d) normal lighting condition.
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2022.18.1
Monira Islam, Md. Salah Udddin Yusuf
E-ISSN: 2224-3488
5
Volume 18, 2022
at different lighting condition, the detector detects the
three traffic signs with 0.97, 0.98 and 0.98 confidence
score with accurate class label termed as “yield” on
both left and right side of road and “keep right”
direction respectively. The accuracy of localization
falls with less confidence score when Faster R-CNN
was used with Resnet 152 architecture. While there
exist multiple targets in one image frame, score of
one target is found 0.53 at foggy weather as shown in
Figure 5(c) which is much lower than ResNet 50
FPN. Resnet50 backbone is found more suitable to
(a)
(b)
(c)
Fig.5 Traffic Sign detection with Faster R-CNN
based ResNet 152 FPN architecture, at (a) different
lighting condition; (b) normal lighting condition with
multiple signs; (c) foggy weather.
detect the traffic signs at any environmental
conditions. The loss of the regressor and classifier of
Faster R-CNN with Resnet50 and 152 are shown in
Figure 6 and Figure 7 respectively where the
horizontal axis indicates the number of training
epochs and vertical axis indicates the value of losses.
It is observed the neural network trained for 35
epochs and converged with progression of iteration.
(a)
(b)
(c)
Fig.6 Losses of (a) RPN box regression; (b)
Classifier; (c) objectness with ResNet 50 FPN
architecture.
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2022.18.1
Monira Islam, Md. Salah Udddin Yusuf
E-ISSN: 2224-3488
6
Volume 18, 2022
(a)
(b)
(c)
Fig.7 Losses of (a) RPN box regression; (b)
Classifier; (c) objectness with ResNet 152 FPN
architecture.
3.2 Performance with MobileNet and
RetinaNet FPN architecture
The performance of traffic sign detection is
compared with Faster R-CNN based MobileNetV3
Large FPN architecture and one-shot detector
RetinaNet ResNet 50 FPN architecture is tabulated in
Table 1. The losses of classifier, regressor as well as
total losses are illustrated in Table 1. The loss is
found small for a two-shot detector, Faster R-CNN
whereas it increases notably for the one-shot detector,
RetinaNet for both classification and regression
while both detectors considered the Resnet50 FPN.
Table 1 Performance comparison of different
network architecture in terms of losses.
Method
Losses
Lclassifier
Lregressor
Faster R-CNN +
ResNet 50 FPN
0.09
0.05
0.04
Faster R-CNN +
ResNet 152 FPN
0.11
0.07
0.04
Faster R-CNN +
MobileNet v3 large
FPN
0.181
0.085
0.096
RetinaNet + ResNet
50 FPN
0.56
0.45
0.09
The losses of classifier and regressor are found 0.05
and 0.04 respectively for Faster R-CNN with ResNet
50 FPN which are the least among all the network
architectures. Figure 8(a) and Figure 8(b) indicate the
confidence score of detected traffic sign with Faster
R-CNN+MobileNetV3 FPN and RetinaNet +ResNet
FPN structure. It is observed that Faster R-CNN with
lager FPN MobileNetV3 detects the object with
confidence score 0.79 as shown in Figure 8(a). The
performance declines with 0.55 score for RetinaNet
with ResNet 50 backbone as displayed in Figure 8(b)
which indicate the efficacy of two-shot detector,
Faster R-CNN for the intended task than the one-shot
detector RetinaNet.
(a)
(b)
Fig.8 Traffic sign detection with (a) MobileNet v3
FPN; (b) RetinaNet ResNet 50 FPN architecture.
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2022.18.1
Monira Islam, Md. Salah Udddin Yusuf
E-ISSN: 2224-3488
7
Volume 18, 2022
Table 2 and Table 3 list the performance of different
architecture in terms of mean average precision
(mAP) and mean average recall (mAR) evaluation
index for traffic sign detection. Precision measures
the correctness of prediction and recall indicates the
goodness of predicting the positives. IoU measures
overlapping between the predicted boundary and the
ground truth one whereby threshold is predefined as
0.5. Precision and Recall can be obtained as
Average precision (AP) indicates the area under the
precision-recall curve whereas the evaluation metric,
precision or recall ranges in 0 to 1. In case of COCO
dataset AP and mAP considers the same meaning to
measure the accuracy of the object detector. Table 2
indicates that, the mAP for Faster R-CNN with
ResNet 50 FPN is greater than other networks
yielding 0.29, 0.34, and 0.64 for the small, medium,
and large area respectively. The RetinaNet with
ResNet 50 FPN provides larger mAR as 0.49 and
0.77 for small and large area respectively as listed in
Table 3. Faster R-CNN with ResNet 50 architecture
also attained a notable mAR for small, medium, and
large area. The score of bbox, classification and
regression losses, mAP and mAR evaluation metric
show a clear advancement in the performance for
Faster R-CNN with ResNet50 FPN network
architecture than Faster R-CNN based other
architecture. Additionally, the two-shot detector,
Faster R-CNN outperforms one-shot detector,
RetinaNet in this task for traffic signal detection and
recognition.
Table 2 Performance comparison of different
network architecture in terms of mAP for IoU
0.5:0.95.
Method
mAP
Areas
mAP
Aream
mAP
Areal
Faster R-CNN +
ResNet 50 FPN
0.29
0.34
0.64
Faster R-CNN +
ResNet 152 FPN
0.16
0.28
0.46
Faster R-CNN +
MobileNetV3 large
FPN
0.085
0.15
0.33
RetinaNet +
ResNet 50 FPN
0.087
0.19
0.17
Table 3 Performance comparison of different
network architecture in terms of mAR for IoU
0.5:0.95.
Method
mAR
Areas
mAR
Aream
mAR
Areal
Faster R-CNN +
ResNet 50 FPN
0.42
0.59
0.64
Faster R-CNN +
ResNet 152 FPN
0.33
0.53
0.64
Faster R-CNN +
MobileNetV3 large
FPN
0.15
0.32
0.35
RetinaNet + ResNet
50 FPN
0.49
0.56
0.77
3.3 Traffic sign recognition with CNN
Figure 9 (a) and Figure 9(b) depict the accuracy and
loss curve respectively of the CNN classifier for 35
epochs. The classification of traffic signs is
performed with GTSRB dataset. A Classifier is built
(a)
(b)
Fig. 9 (a) Plot of training and validation
(a) accuracy; (b) loss of the CNN classifier.
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2022.18.1
Monira Islam, Md. Salah Udddin Yusuf
E-ISSN: 2224-3488
8
Volume 18, 2022
with convolution layer followed by batch
normalization, maxpooling and dropout layer is used
to classify the signs. The total data samples are
partitioned randomly where 70% of the total samples
were taken for training, 15% for testing and 15% for
validation. The training, testing and validation
accuracy is found 99.56%. 97.89% and 99.76%
respectively with validation loss 2.48%. The
precision, recall and F-score is 97.96%, 97.89% and
97.87% respectively on test data.
4 Conclusion
Traffic sign detection from street image is
challenging because of motion, blur, object size or
different lighting condition. Faster R-CNN based two
stage object detection method has been applied based
on different deep learning network structure. Faster
R-CNN based ResNet 50 FPN network shows the
improved result comparing with the other structure in
terms of confidence score and mAP. The losses are
higher for RetinaNet which is the one shot detector.
The two-stage detector Faster R-CNN performs
better for traffic sign detection despite motion, blur,
fog or lighting condition than the one-shot detector
which can be further implemented in real time
scenario.
References:
[1] R. Madan, D. Agrawal, S. Kowshik, H.
Maheshwari, S. Agarwal, and D. Chakravarty,
Traffic Sign Classification using Hybrid HOG-
SURF Features and Convolutional Neural
Networks. In ICPRAM, pp. 613-620, 2019.
[2] M. Pontil, and A. Verri, Support vector
machines for 3D object recognition. IEEE
transactions on pattern analysis and machine
intelligence, 20(6), pp. 637-646, 1998.
[3] Y. Lai, N. Wang, Y. Yang, and L. Lin, Traffic
signs recognition and classification based on
deep feature learning. In 7th International
Conference on Pattern Recognition Applications
and Methods (ICPRAM), Madeira, Portugal,
2018, pp. 622-629.
[4] A. Lee, Comparing Deep Neural Networks and
Traditional Vision Algorithms in Mobile
Robotics. Swarthmore College, 2015.
[5] S. Maldonado-Bascon, S. Lafuente-Arroyo, P.
Gil-Jimenez, H. Gomez-Moreno, and F. Lopez-
Ferreras. Road-sign detection and recognition
based on support vector machines. Intelligent
Transportation Systems, IEEE Transactions on,
8(2):264–278, June 2007.
[6] K. Lu, Z. Ding, and S. Ge. Sparse-
representation-based graph embedding for
traffic sign recognition. IEEE Transactions on
Intelligent Transportation Systems, 13(4):1515–
1524, 2012
[7] S. Pei, F. Tang, Y. Ji, J. Fan, and Z. Ning,
Localized Traffic Sign Detection with Multi-
scale Deconvolution Networks, 2018.
[8] Y. Yang, S. Liu, W. Ma, Q. Wang, and Z. Liu,
Efficient Traffic-Sign Recognition with Scale-
aware CNN, 2018. arXiv preprint
arXiv:1805.12289.
[9] A. Soetedjo, and I. K. Somawirata, An Efficient
Algorithm for Implementing Traffic Sign
Detection on Low Cost Embedded System.
International Journal of Innovative Computing
Information and Control, 14(1), pp. 1-14, 2018.
[10] S. Hoo-Chang, H. R. Roth, M. Gao, L. Lu, Z.
Xu, I. Nogues, and R. M. Summers, Deep
convolutional neural networks for computer-
aided detection: CNN architectures, dataset
characteristics and transfer learning. IEEE
transactions on medical imaging, 35(5), pp.
1285, 2016.
[11] R. Girshick, Fast R-CNN, IEEE International
Conference on Computer Vision (ICCV), 2015,
pp. 1440-1448.
[12] S. Ren, K. He, R. Girshick, J. Sun, Faster R-
CNN: Towards real-time object detection with
region proposal networks, Advances in neural
information processing systems, 2015, pp. 91-
99.
[13] W. Liu, D. Anguelov, D. Erhan, S. Christian, R.
Scott, F. Cheng-Yang, B. Alexander, SSD:
Single Shot MultiBox Detector, European
Conference on Computer Vision, pp. 21-37.
[14] J. Redmon, S. Divvala, R. Girshick R, et al, You
only look once: Unified, real-time object
detection, Proceedings of the IEEE conference
on computer vision and pattern recognition,
2016, pp. 779-788.
[15] W. Cai, J. Li, Z. Xie, T. Zhao and K. LU, “Street
Object Detection Based on Faster R-CNN,”
Chinese Control Conference (CCC), 2018, pp.
9500-9503.
[16] D. C. Ciresan, U. Meier, J. Masci, and J.
Schmidhuber. A committee of neural networks
for traffic sign classification. In International
Joint Conference on Neural Networks, pages
1918–1921, 2011.
[17] P. Sermanet and Y. LeCun. Traffic sign
recognition with multi-scale convolutional
networks. In Neural Networks (IJCNN), The
2011 International Joint Conference on, pages
2809–2813, July 2011.
[18] J. Jin, K. Fu, and C. Zhang. Traffic sign
recognition with hinge loss trained
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2022.18.1
Monira Islam, Md. Salah Udddin Yusuf
E-ISSN: 2224-3488
9
Volume 18, 2022
convolutional neural networks. IEEE
Transactions on Intelligent Transportation
Systems, 15(5):1991–2000, 2014.
[19] J. Stallkamp, M. Schlipsing, J. Salmen, and C.
Igel. Man vs. computer: Benchmarking machine
learning algorithms for traffic sign recognition.
Neural Networks, (0):–, 2012.
[20] A. Krizhevsky, I. Sutskever, and G. E. Hinton.
Imagenet classification with deep convolutional
neural networks. In F. Pereira, C. Burges, L.
Bottou, and K. Weinberger, editors, Advances in
Neural Information Processing Systems 25,
pages 1097–1105. Curran Associates, Inc.,
2012.
[21] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu,
R. Fergus, and Y. LeCun. Overfeat: Integrated
recognition, localization and detection using
convolutional networks. CoRR, abs/1312.6229,
2013.
[22] J. Uijlings, K. van de Sande, T. Gevers, and A.
Smeulders. Selective search for object
recognition. International Journal of Computer
Vision, 2013.
[23] L. Zitnick and P. Dollar. Edge boxes: Locating
object proposals from edges. In ECCV.
European Conference on Computer Vision,
September 2014.
[24] M. Milicevic, K. R. U. N. O. S. L. A. V.
Zubrinic, I. N. E. S. Obradovic, and T. O. M. O.
Sjekavica, Data augmentation and transfer
learning for limited dataset ship
classification. WSEAS Trans. Syst.
Control, vol. 13, no. 1, pp. 460-465, 2018.
[25] A. El-Sawy, M. Loey, and H. El-Bakry, Arabic
handwritten characters recognition using
convolutional neural network. WSEAS
Transactions on Computer Research, vol. 5,
2017, pp. 11-19.
Author Contributions:
Monira Islam: Conceptualization, methodology,
simulation and the optimization, original drafting.
Md. Salah Uddin Yusuf: Simulation, formatting,
technical review, editing and methodology.
All authors have read and agreed to the published
version of the manuscript.
Sources of funding for research presented in
a scientific article or scientific article itself
This research did not receive any grant from funding
agencies in the public, commercial, or non-profit
sectors.
Authors Biography
Monira Islam received her B.Sc.
and M.Sc. degree from Khulna
University of Engineering
Technology (KUET), Khulna,
Bangladesh and serving as an
Assistant Professor at the same
university. Currently, she is
doing her PhD at The Chinese
University of Hong Kong. Her research interest
includes Brain-computer interaction, Bio-Signal
processing, Image Processing and Computer vision.
Md. Salah Uddin Yusuf received
the Ph.D. degree in Electrical
and Electronic Engineering
from Khulna University of
Engineering and Technology
(KUET), Bangladesh, in 2016.
Since 2001, he has been with the
same department of same
University as a faculty member
and currently serving as Professor. In professional
activities, he is life fellow of Institute of Engineers
(IEB), Bangladesh and member of IEEE. His
research interests mainly focus on deep learning
based signal, image and video processing with
quality assessment, Face liveness detection and
authentication, GAN-based Human activity analysis
for modern healthcare system.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
Llll
Llll
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2022.18.1
Monira Islam, Md. Salah Udddin Yusuf
E-ISSN: 2224-3488
10
Volume 18, 2022
