3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
119
AN AUTOMATED SYSTEM FOR TRAFFIC SIGN
RECOGNITION USING CONVOLUTIONAL NEURAL
NETWORK
Sanam Narejo
Department of Computer Systems Engineering, Mehran University of Engineering & Technology,
Jamshoro, (Pakistan).
E-mail: sanam.narejo@faculty.muet.edu.pk ORCID: https://orcid.org/0000-0002-3537-8949
Shahnawaz Talpur
Department of Computer Systems Engineering, Mehran University of Engineering & Technology,
Jamshoro, (Pakistan).
E-mail: shahnawaz.talpur@faculty.muet.edu.pk ORCID: https://orcid.org/0000-0002-2660-6145
Madeha Memon
Department of Computer Systems Engineering, Mehran University of Engineering & Technology,
Jamshoro, (Pakistan).
E-mail: madeha.memon@gmail.com ORCID: https://orcid.org/0000-0003-2147-0944
Amna Rahoo
Department of Computer Systems Engineering, Mehran University of Engineering & Technology,
Jamshoro, (Pakistan).
E-mail: f16cs05@students.muet.edu.pk ORCID: https://orcid.org/0000-0002-9376-6528
Recepción:
13/11/2020
Aceptación:
13/11/2020
Publicación:
13/11/2020
Citación sugerida Suggested citation
Narejo, S., Talpur, S., Memon, M., y Rahoo, A. (2020). An automated system for trac sign recognition
using convolutional neural network. 3C Tecnología. Glosas de innovación aplicadas a la pyme. Edición Especial,
Noviembre 2020, 119-135. https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
120
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
ABSTRACT
TSR (Trac Sign Recognition) represents an important feature of advanced driver
assistance system, contributing to the safety of the drivers, autonomous vehicles as well
and to increase driving comfort. In today’s world road conditions drastically improved as
compared with past decades. Obviously, vehicle’s speed increased. So, on driver’s point
of view there might be chances of neglecting mandatory road signs while driving. This
paper explores the system to helps the driver about recognition of road signs to avoid road
accidents. TSR is challenging task, while its accuracy depends on two aspects: feature
extractor and classier. Current popular algorithms mainly deploy CNN (Convolutional
Neural Network) to execute both feature extraction and classication. In this paper, we
implement the trac sign recognition by using CNN, the CNN will be trained by using the
dataset of 43 dierent classes of trac signs along with TensorFlow library. The results will
show the 95% accuracy.
KEYWORDS
Convolutional Neural Network, Trac Sign Recognition, Autonomous Vehicles,
Exploratory Data Analysis
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
121
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
1. INTRODUCTION
Road and trac signs must be properly installed in the necessary locations and an inventory
of them is ideally needed to help ensure adequate updating and maintenance. Meetings
with the highway authorities in both Scotland and Sweden revealed the absence of but a
need for an inventory of trac signs. An automatic means of detecting and recognizing
trac signs can make a signicant contribution to this goal by providing a fast method
of detecting, classifying and logging signs. This method helps to develop the inventory
accurately and consistently. Once this is done, the detection of disgured or obscured
signs becomes easier for human operator. Road and trac sign recognition is the eld of
study that can be used to aid the development of an inventory system (for which real-time
recognition is not required) or aid the development of an in-car advisory system (when
real-time recognition is necessary). Both road sign inventory and road sign recognition
are concerned with trac signs, face similar challenges and use automatic detection and
recognition. A road and trac sign recognition system could in principle be developed as
part of an ITS (Intelligent Transport Systems) that continuously monitors the driver, the
vehicle, and the road in order, for example, to inform the driver in time about upcoming
decision points regarding navigation and potentially risky trac situations (Fleyeh, 2008).
The aim of intelligent transport systems is to increase transportation eciency, road
safety and to reduce the environmental impact with the use of advanced communication
technologies (Sermanet, & LeCun, 2011; De la Escalera, Armingol & Mata, 2003).
Automatic TSR, as an important task of Advanced Driver Assistance Systems and ITS has
been of great interest in recent years. The road signs are placed on either roadside or above
the roads. These signs provide mandatory information regarding to guiding, warning and
regulating the behaviors to drivers in order to make driving safer and easier.
There are several dierent TSR like speed limits, no entry, trac signals, turn left or right,
children crossing, no passing of heavy vehicles, etc. Trac sign classication/recognition is
the process of identifying the, which class trac sign belongs to. TSR has a direct impact on
the safety of drivers, and damages can be easily produced due to their ignorance. Automatic
systems are developed to assist the drivers, based on detection and recognition of signs
which corrects the most unsafe driving behaviors.
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
122
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
The main purpose of advanced driver assistance systems is to collect signicant information
for drivers in order to reduce their eort in safe driving. Because drivers must pay attention
to various conditions, including vehicle speed and orientation, passing cars and to many
more. So, if driver assistance systems collect such information, it will greatly reduce the
burden of drivers. Thus, road signs are designed in such a way that attracts a driver’s
attention with colors and simple geometric shapes.
The research work available on recognition of trac signs for local roads is quite sparse
and still at the preliminary stage. Mostly it is focused on recognition of trac signs for local
roads and they are also still at the preliminary stage and focused on recognition of trac
signs through static images. In this work, algorithms were developed in “Google Colab”
environment to recognize trac signs while vehicles in motion. Project is mainly focused
on automatic recognition of warning signs placed in local roads captured by image clips.
Trac signs were recognized based on its geometrical characteristics and color information
(Gunawardana 2010). In this project, we develop a deep NN (Neural Network) model
that that can classify the trac signs present in the image into dierent categories. With
this model, we can read and understand trac signs which are very important task for
autonomous vehicles.
2. LITERATURE SURVEY
There are many researches in the literature dealing with Road TSR problem. According
to Kale and Mahajan (2015), the road sign recognition system is to be divided into two
parts, the rst part is detection stage which is used to detect the signs from a whole image,
and the second part is classication stage that classies the detected sign in the rst part
into one of the reference signs which are presents in the dataset. They used PCA (Principle
Component Analysis) and ANN (Articial Neural Network) techniques for detection and
recognition with the dataset of dierent road signs from Maharashtra RTO (Regional
Transport Oce). Tool used by them was MATLAB. For TSR, mostly the preferable model
for image classication is CNN. The recognition system is implanted into autonomous
vehicle ‘Eurecar’. The system was followed by HSV (Herpes Simplex Viruses) and Hough
transform algorithms for extracting ROI (Return on Investment) of trac signs. Gaussian
blurring was also applied as canny edge detectors. Then extracted area is given to CNN
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
123
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
model. They used dataset of 6 types of (Korean-Version) trac signs to train the CNN
model. The obtained results demonstrated low accuracy; it shows overlapping results when
several proposal regions point to a same trac sign. Development of clustering algorithm
is considered as a future work for robust recognition system (Jung et. al. 2016). Most of
the research used color segmentation technique with C-CNN with GERMAN trac signs
dataset for detection. The C-CNN method consists of selecting a set of ROIs by applying
a color thresholding on the input image, thus reducing the search space. Then, a trained
CNN is used to classify the ROI (whether it contains a trac sign or not), followed by
another CNN with the same architecture, that is used to recognize the detected trac
signs. Therefore, 2 datasets are selected, one for detection and another one, to recognize
the trac sign. Therefore, CNN was trained to recognize two classes: trac sign/no trac
sign. It was concluded that C-CNN is slow and sensitive to weather conditions (Boujemaa
et al., 2017).
Researchers presented a three-stage real-time TSR system, consisting of a segmentation, a
detection and a classication phase. They combine the color enhancement with an adaptive
threshold to extract red regions in the image. The detection is performed using an ecient
linear SVM (Support Vector Machine) with HOG (Histogram of Oriented Gradients)
features. The tree classiers, K-d tree and Random Forest, identify the content of the
trac signs found. A spatial weighting approach is proposed to improve the performance
of the K-d tree. The Random Forest and Fisher’s Criterion are used to reduce the feature
space and accelerate the classication. They presented that only a subset of about one third
of the features is enough to attain a high classication accuracy on the GTSRB (German
Trac Sign Recognition Benchmark (Zaklouta & Stanciulescu 2014). The paper proposed
a method for Trac Sign Detection and Recognition using image processing for the
detection of a sign and an ensemble of CNN for the recognition of the sign. TensorFlow is
used for the implementation of the CNN. They have achieved higher than 99% recognition
accuracies for circular signs on the Belgium and German data sets (Vennelakanti et al.,
2019). Research based on TSR methods proposed mechanism for real time TSR using
CNN. The training database was established by eld sample collection, with which the
neural network model was trained. SGD (Stochastic Gradient Descent) optimizer is utilized
during training to improve the learning eciency. The test results show that the proposed
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
124
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
method achieves good performance in speed, accuracy and robustness for real time TSR
(Xu et al., 2018).
R-CNN was the rst to use this strategy, but it is very slow for two reasons (Girshick et al.,
2014). Firstly, generating category-independent object proposals is costly; it takes about 3s
to generate 1000 proposals for the Pascal VOC 2007 images (Simonyan & Zisserman 2014).
Secondly, it applies a whole deep CNN to every candidate proposal calculated, which is
obviously very inecient and time consuming. To improve eciency, the SPP-Net (Spatial
Pyramid-Pooling Network). Girshick et al. (2015) calculated a convolutional feature map for
the entire image and extracts feature vectors from the shared feature map for each proposal.
This speed up the R-CNN approach about 100 times. They have proposed the Fast R-CNN
model, which is a faster version of the R-CNN approach. He et al. (2015) proposed RPNs
(Region Proposal Networks), which generate object proposals using convolutional feature
maps. This allows the generator of the object proposal to share the convolutional features
of the whole image with the detection network. With this technique detection system can
achieve a frame rate of 5 fps on a powerful GPU (Graphics Processing Unit). Szegedy et al.
(2016) improved the network architecture, to achieve a frame rate of 50 fps in testing, with
competitive detection performance.
In current trac management systems, there is high probability that driver may miss some
of the trac signs on the road because of overcrowding due to neighboring vehicles. So, we
have introduced the TSR system with an aim of detecting and recognizing all the emerging
trac signs.
3. METHODOLOGY
Initially, we opted for the data collection. We applied some techniques on the data for
exploration and then visualized the data through EDA techniques. And nally, we build
in this study a classication model based on the CNN. Afterwards, model is trained and
validated and then based on the validated model, we attempted a test. And afterwards, we
deployed our classier. As we mentioned earlier that for classication, we have used CNN
model, it is based on convolution of lters and images or raw inputs.
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
125
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
In this section, we will describe the CNN based approach. A owchart of the main phases
is shown in Figure 1.
Figure 1. Flowchart of methodology.
In NNs, CNN is one of the main categories to do images recognition, images classications.
CNN image classications take an input image, process it and classify it under certain
categories (Eg., Dog, Cat, Tiger, Lion).
Technically, deep learning CNN models to train and test, each input image will pass it
through a series of convolution layers with lters (Kernals), Pooling, fully connected layers
(FC) and apply Softmax function to classify an object with probabilistic values between 0
and 1. The calculation of the convolutional layer can be simplied as shown in Equation
(1):
(1)
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
126
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
Where, k w
ij
is the weight value of the convolution kernel; i
j
x the input pixel value
corresponding to the weight of the convolution kernel; the oset corresponding to the rst
convolution kernel; k b is the bias; f () x is the activation function. CNN commonly used
activation functions as an RLU (Rectied Linear Unit).
The Figure 2 is a complete ow of CNN to process an input image and classies the objects
based on values.
Figure 2. Neural network with many convolutional layers.
Convolution Layer is the rst layer to extract features from an input image. Convolution
preserves the relationship between pixels by learning image features using small squares of
input data. It is a mathematical operation that takes two inputs such as image matrix and
a lter or kernel.
In Non-Linearity (ReLU): ReLU stands for RLU for a non-linear operation. Its purpose
is to introduce non-linearity in our ConvNet. Since, the real-world data would want our
ConvNet to learn would be non-negative linear values.
The Pooling Layer would reduce the number of parameters when the images are too large.
Spatial pooling also called sub-sampling or down-sampling which reduces the dimensionality
of each map but retains important information. Spatial pooling can be of dierent types:
Max Pooling
Average Pooling
Sum Pooling
Max pooling takes the largest element from the rectied feature map. Taking the largest
element could also take the average pooling. Sum of all elements in the feature map call as
sum pooling.
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
127
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
Fully Connected Layer is attached attened the matrix into vector and feed it into a fully
connected layer like a NN. The feature map matrix will be converted as vector (x
1
, x
2
, x
3
,
…). With the fully connected layers, we combined these features together to create a model.
Finally, we have an activation function such as SoftMax or sigmoid to classify the outputs
as cat, dog, car, truck etc.
In the subsequent sections, we present the detailed steps of our methodology.
4. DATA PREPROCESSING
For this project, we have selected the publicly available dataset from Kaggle. The dataset
contains more than 50,000 images of dierent trac signs. It is further classied into 43
dierent classes. Few of them are shown in Figure 3. Prior to training a NN, data is divided
in training and test sets.
Figure 3. Dataset of trafc signs.
The ‘trainfolder contains 43 folders each representing a dierent class. The range of the
folder is from 0-42. With the help of OS module, we must iterate overall the classes and
append images and their respective labels in the data and label list.
An image is made up of pixels and each pixel has 3 values to specify its color i.e. RGB
(Red, Green Blue). In order for machines to understand the image, we converted the image
into numbers. For this purpose, we use the PIL (Python Imaging Library) that can perform
many image manipulations tasks. The subsequent step was of resizing the images into some
uniform criteria. Therefore, we resize all the images to a xed size, such as,30x30.Let’s
traverse through all the classes, open the image using PIL and also resize the image to
30x30 dimensions. Then we will append the data and label in the X and Y list respectively.
After storing all the images and their labels into lists (data and labels), the list was further
transformed into NumPy arrays for feeding the model. Nevertheless, nally the shape of
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
128
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
data is (39209, 30, 30, 3) which means that there are 39,209 images of size 30x30 pixels and
the last 3 means the data contains colored images (RGB Value).
4.1. APPLYING EDA TECHNIQUES
In order to apply the EDA techniques, we have divided the process in three steps. Initially, we
have focused on in-depth understanding for the dataset. Subsequently, some data cleaning
is applied. EDA is the practice of using visual and quantitative methods to understand and
summarize a dataset without making any assumptions about its contents. It is a crucial step
to take before diving into machine learning or statistical modeling because it provides the
context needed to develop an appropriate model for the problem at hand and to correctly
interpret its results. Afterwards, we have attempted to nd some correlations available in
the dataset.
Step-1: Understanding the Data: Tables 1-2 present the details of data set by
summarizing the rst and last 5 rows respectively, its attributes and associated values. The
dataset contains 12630 rows and 8 columns. This indicates that the data is enough to be
used for any Deep learning architecture.
Table 1. Showing head function.
Data
No.
Width Height ROI.X1 ROI.Y1 ROI.X2 ROI.Y2 Class Id Path
12625
42
41
6
36
Test/12625.png
12626
50
51
5
46
Test/12626.png
12627
29
29
6
24
Test/12627.png
12628
48
49
6
44
Test/12628.png
12629
32
31
5
26
Test/12629.png
Table 2. Showing tail function.
Data
No.
Width Height ROI.X1 ROI.Y1 ROI.X2 ROI.Y2 Class Id Path
0
53
54
5
49
Test/00000.png
1
42
45
5
40
Test/00001.png
2
48
52
6
47
Test/00002.png
3
27
29
5
24
Test/00003.png
4
60
57
5
52
Test/00004.png
Step-2: Cleaning the Data: After understanding the data it was further examined null or
missing values. Data set is free from missing values as depicted in Table 3.
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
129
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
Table 3. Showing cleaning process.
Width
0
Height
0
RoiX1
0
RoiY1
0
RoiX2
0
RoiY2
0
ClassId
0
Path
0
dtype
Int64
Step-3: Analyzing the Relationship: Correlation matrix is used to summarize the data
as an input into more advanced analysis. We get visual representation of correlation matrix
by heatmap. The heatmap is a way of representing the data in a 2D (Two-Dimensional)
form. The data values are represented as colors in the graph. The goal of the heatmap is to
provide a colored visual summary of information. It is a graphical representation of data
where the individual values contained in a matrix are represented as colors. It is a bit like
looking a data table from above. It is useful to display a general view of numerical data, not
to extract specic data point. Figure 4 shows heatmap.
Figure 4. Heatmap.
A bar chart or bar graph is a chart or graph that presents categorical data with rectangular
bars with heights or lengths proportional to the values that they represent. The bars can be
plotted vertically or horizontally. A vertical bar chart is sometimes called a column chart.
We used bar to plot graphs for any column’s visual representation. Here we used width and
height column of dataset with bar plot function. Figure 5 shows bar plot for input.
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
130
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
Figure 5. Bar plot.
5. EXPERIMENTAL SETUP FOR CNN
To classify the images into their respective categories, we developed and trained. The CNN
have proved the state of the art in image classication tasks and this is what we will be using
for our model. A CNN is made up of convolutional and pooling layers. At each layer, the
features from the image are extracted that helps in classifying the image.
Apart from this, a dropout layer is also added, which is used to handle the over tting of the
model. The dropout layer drops some of the neurons while training. Finally, the model is
compiled with cross entropy measures, because the dataset has multi classes to be classied.
The Table 4 shows layers in details of CNN model along with parameters.
Table 4. Architecture of CNN model.
Layer Number
Layer Type
L1
Conv2D (32x5x5), ReLU
L2
Conv2D (32x5x5), ReLU
L3
MaxPool2Dlayer(pool_size=(2,2))
L4
Dropout layer (rate=0.25)
L5
Conv2D (64x5x5), ReLU
L6
Conv2D (64x5x5), ReLU
L7
Flatten Layer (1 Dimension)
L8
DenseFullyconnected layer (256 nodes,ReLU)
L9
Dropout layer (rate=0.5)
L10
Dense Layer (43 nodes, softmax)
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
131
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
5.1. TRAIN AND VALIDATE THE MODEL
After building the architecture of model. The model is dened, and the data is ready. To
start the training of our model we initialize the training set, validation set, batch size and
no of epochs.
So, for classier, we opted for multiple architectures of CNN model. We experimented with
various number of batch sizes and activation functions. The Table 5 shows the details of all
our implemented models.
Table 5. CNN trained model summary.
CNN Model No.
Batch Size
Epochs
Test Accuracy (%)
1
32
100
82.00
2
32
110
91.00
3
64
100
94.00
4
64
110
95.00
However, as we nd out that CNN4 has outperformed to the rest of other models. Therefore,
we used this model to further demonstrations associated with it.
Therefore, this model is trained with batch size 64. And after 110 epochs, the accuracy
was stable. Our model got 95% accuracy on the training dataset. We plot the graph for
accuracy and loss. Figures 6-7 showing accuracy and loss respectively.
Figure 6. Accuracy of trained model CNN4. Figure 7. Loss of trained model CNN 4.
Finally, we build a graphical user interface for our trac signs classier. The GUI (Graphical
User Interface) is built for uploading the image and to predict the trac sign we must
provide the same dimension we have used when building the model. A GUI will save a lot
of time in testing and seeing the results of our model prediction.
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
132
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
From the interface of the GUI application, we will ask the user for an image and extract the
le path of the image. Then we use the trained model that will take the image data as input
and provide us the class our image belongs to.
6. RESULTS
This section presents the series of experiments carried out to validate the CNN approach
for recognition of road signs. To get the road images, we use camera of 16 megapixels, to
get the good distribution of images at dierent angles and positions. The images in each
class are randomly chosen.
By observing Figures 8-9, we got an understanding of how autonomous vehicles can take
advantage of CV (Computer Vision) and Deep Learning techniques to automatically
recognize and classify from multiple classes.
Figure 8. Showing results of speed Limit(30km/hr). Figure 9. Showing results of turn right ahead.
7. CONCLUSION
In this research work, we demonstrated and developed an ecient alert trac sign detection
and recognition system. Both color information and the geometric property of the road
signs are used to classify the detected trac signs. The experiment shows that the system
can achieve a high detection rate of 95%.System is giving accurate results under dierent
illumination conditions, weather conditions, day light conditions and dierent speed levels
of the vehicle. We have successfully classied the trac signs classier with 95% accuracy
and visualized how our accuracy and loss changes with time, which is pretty good from a
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
133
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
simple CNN model. The techniques implemented in this research can be used as a basis for
developing general purpose, advanced intelligent trac surveillance systems.
ACKNOWLEDGEMENT
We are thankful to the Department of Computer Systems Engineering, Mehran University
of Engineering & Technology, Jamshoro, Pakistan, for providing facilities to conduct this
research work.
REFERENCES
Boujemaa, K.S., Berrada, I., Bouhoute, A., & Boubouh, K. (2017). Trac Sign
Recognition Using Convolutional Neural Networks. In 2017 International Conference
on Wireless Networks and Mobile Communications (WINCOM), pp. 1-6. https://doi.
org/10.1109/WINCOM.2017.8238205
De la Escalera, A., Armingol, J. M., & Mata, M. (2003). Trac Sign Recognition and
Analysis for Intelligent Vehicles. Image and Vision Computing, 21(3), 247-258. https://
e-archivo.uc3m.es/bitstream/handle/10016/7089/traffic_escalera_IVC_2003_
ps.pdf
Fleyeh, H. (2008). Trac and Road Sign Recognition (Doctoral Dissertation). Napier University.
https://www.diva-portal.org/smash/get/diva2:523372/FULLTEXT01.pdf
Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014). Rich Feature Hierarchies
for Accurate Object Detection and Semantic Segmentation. In 2014 IEEE Conference
on Computer Vision and Pattern Recognition, pp. 580-587. https://doi.org/10.1109/
CVPR.2014.81
Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2015). Region-Based Convolutional
Networks for Accurate Object Detection and Segmentation. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 38(1), 142-158. https://doi.org/10.1109/
TPAMI.2015.2437384
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
134
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
Gunawardana, P.M. (2010). Automatic Detection and Recognition of Trac Signs (Doctoral
Dissertation). University of Colombo, Sri Lanka.
He, K., Zhang, X., Ren, S., & Sun, J. (2015). Spatial Pyramid Pooling in Deep
Convolutional Networks for Visual Recognition. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 37(9), 1904-1916. https://doi.org/10.1109/
TPAMI.2015.2389824
Jung, S., Lee, U., Jung, J., & Shim, D.H. (2016). Real-Time Trac Sign Recognition
System with Deep Convolutional Neural Network. In IEEE 13th International
Conference on Ubiquitous Robots and Ambient Intelligence (URAI), pp. 31-34 https://www.
semanticscholar.org/paper/Real-time-Trac-Sign-Recognition-system-with-deep-
Jung-Lee/17a608865ba3c6b2c27fcb18973c27139560954a
Kale, A.J., & Mahajan, R.C. (2015). A Road Sign Detection and the Recognition for
Driver Assistance Systems. IEEE International Conference on Energy Systems and Applications,
pp. 69-74.
Sermanet, P., & LeCun, Y. (2011). Trac Sign Recognition with Multi-Scale
Convolutional Networks. In The 2011 International Joint Conference on Neural Networks,
pp. 2809-2813. https://doi.org/10.1109/IJCNN.2011.6033589
Simonyan, K., & Zisserman, A. (2014). Very Deep Convolutional Networks for Large-
Scale Image Recognition. arXiv preprint. https://arxiv.org/abs/1409.1556
Szegedy, C., Ioe, S., Vanhoucke, V., & Alemi, A. (2016). Inception-V4, Inception-
Resnet and the Impact of Residual Connections on Learning. arXiv Preprint. https://
arxiv.org/abs/1602.07261
Vennelakanti, A., Shreya, S., Rajendran, R., Sarkar, D., Muddegowda, D., &
Hanagal, P. (2019). Trac Sign Detection and Recognition Using a CNN Ensemble.
In 2019 IEEE International Conference on Consumer Electronics (ICCE), pp. 1-4. https://
doi.org/10.1109/ICCE.2019.8662019
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 – 4143 Edición Especial Special Issue Noviembre 2020
135
https://doi.org/10.17993/3ctecno.2020.specialissue6.119-135
Xu, S., Niu, D., Tao, B., & Li, G. (2018). Convolutional Neural Network Based Trac
Sign Recognition System. IEEE 5th International Conference on Systems and Informatics
(ICSAI), pp. 957-961.
Zaklouta, F., & Stanciulescu, B. (2014). Real-Time Trac Sign Recognition in Three
Stages. Robotics and Autonomous Systems, 62(1), 16-24. https://doi.org/10.1016/j.
robot.2012.07.019