Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
1 Available online at www.sciencedirect.com
1

ScienceDirect
Procedia Computer Science 00 (2019) 000–000
Procedia Computer Science 00 (2019) 000–000 www.elsevier.com/locate/procedia
Procedia Computer Science 165 (2019) 252–258 www.elsevier.com/locate/procedia

INTERNATIONAL CONFERENCE ON RECENT TRENDS IN ADVANCED COMPUTING

INTERNATIONAL CONFERENCE 2019,
ON RECENT
ICRTACTRENDS
2019 IN ADVANCED COMPUTING
2019, ICRTAC 2019
A
A Deep
Deep Neural
Neural Network
Network Framework
Framework for
for Road
Road Side
Side Analysis
Analysis and
and
Lane
Lane Detection
Detection
a b c d
Utkarsh
Utkarsh Shukla
Shuklaa ,, Ayush
Ayush Mishra
Mishrab ,, Graceline
Graceline Jasmine
Jasmine S
S c , Vaidehi Vd , Subramaniam
, Vaidehi V , Subramaniam
e
Ganesan
Ganesane
a ATLAN, Delhi - 110030
a ATLAN, Delhi - 110030
b DailyCutting, Bangalore - 560044
b DailyCutting, Bangalore - 560044
c Vellore Institute of Techonology, Chennai - 600127
c Vellore Institute of Techonology, Chennai - 600127
d Mother Teresa Women’s University, Kodaikanal, Tamil Nadu - 624101
d Mother Teresa Women’s University, Kodaikanal, Tamil Nadu - 624101
e Oakland University Rochester, USA
e Oakland University Rochester, USA

Abstract
Abstract
This paper presents a computer vision based framework with the aim of aiding the task of driving. The framework serves the
This paper presents a computer vision based framework with the aim of aiding the task of driving. The framework serves the
purpose of road analysis. Road analysis is further divided into two sub-tasks. The first task aims at recognition of the different
purpose of road analysis. Road analysis is further divided into two sub-tasks. The first task aims at recognition of the different
road signs, the second task aims at lane analysis. The task of automatic driving requires humans to multitask and perform many
road signs, the second task aims at lane analysis. The task of automatic driving requires humans to multitask and perform many
operations in split seconds. The framework is introduced to aid this task of driving if not completely automate it while keeping in
operations in split seconds. The framework is introduced to aid this task of driving if not completely automate it while keeping in
mind of using it with a simple hardware and software setup. The effectiveness of the framework lies in its feature of having minimal
mind of using it with a simple hardware and software setup. The effectiveness of the framework lies in its feature of having minimal
complexity which enables it to be used in real-time. The results of the pipeline are quantified by first measuring its accuracy in
complexity which enables it to be used in real-time. The results of the pipeline are quantified by first measuring its accuracy in
the classification of road signs, second measuring its ability to gather the information about the road (lane analysis and 2 vehicle
the classification of road signs, second measuring its ability to gather the information about the road (lane analysis and 2 vehicle
detection) thirdly by performing the time bench-marking.
detection) thirdly by performing the time bench-marking.
©
c 2019
2019TheTheAuthors.
Authors.Published
PublishedbybyElsevier
ElsevierB.V.
B.V.
c 2019 The Authors. Published by Elsevier B.V.
This
This is
is an
an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
This is an open access
Peer-review article under thescientific
CC BY-NC-ND license the(http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under
underresponsibility
responsibilityofofthe
the scientificcommittee ofof
committee INTERNATIONAL
the INTERNATIONAL CONFERENCE
CONFERENCE ONONRECENT
RECENT TRENDS IN
TRENDS
Peer-review under
ADVANCED responsibility
COMPUTING of the scientific committee of the INTERNATIONAL CONFERENCE ON RECENT TRENDS
2019.
IN ADVANCED COMPUTING 2019.
IN ADVANCED COMPUTING 2019.
Keywords: Computer Vision, Image Processing, Vehicle Detection, Convolution Neural Network,Transfer Learning
Keywords: Computer Vision, Image Processing, Vehicle Detection, Convolution Neural Network,Transfer Learning

1. Introduction
1. Introduction
Driving is an amalgamation of different complex tasks and requires un-deviated attention from the driver. Long
Driving is an amalgamation of different complex tasks and requires un-deviated attention from the driver. Long
stretches of driving can become really exhaustive and might cause a lack of attentiveness may lead to accidents. Due
stretches of driving can become really exhaustive and might cause a lack of attentiveness may lead to accidents. Due
to the recent advancement in image processing and deep learning techniques frameworks can be developed with the
to the recent advancement in image processing and deep learning techniques frameworks can be developed with the

1 *Corresponding Author Tel: +91-790-505-7363

1 *Corresponding Author Tel: +91-790-505-7363
utkarshshukla2912@gmail.com
utkarshshukla2912@gmail.com
1877-0509  c 2019 The Authors. Published by Elsevier B.V.
1877-0509 c 2019 The Authors. Published by Elsevier B.V.

This is an open
1877-0509 © access
2019 Thearticle under
Authors. the CC BY-NC-ND
Published license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
by Elsevier B.V.
This is an
Peer-reviewopen access
under article under
responsibility the scientific
of the CC BY-NC-ND license
committee (http://creativecommons.org/licenses/by-nc-nd/4.0/)
oflicense
the INTERNATIONAL CONFERENCE ON RECENT TRENDS IN ADVANCED
This is an open access article under the CC BY-NC-ND (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under
COMPUTING responsibility of the scientific committee of the INTERNATIONAL CONFERENCE ON RECENT TRENDS IN ADVANCED
2019.
Peer-review under responsibility of the scientific committee of the INTERNATIONAL CONFERENCE ON RECENT TRENDS IN
COMPUTING
ADVANCED 2019. COMPUTING 2019.
10.1016/j.procs.2020.01.081
2 Utkarsh Shukla et al. / Procedia Computer Science 00 (2019) 000–000

aim to aid in the process of driving. The figure 1 describes a framework capable of road signs recognition and analysis
involving lane and vehicle detection. The road sign recognition consists of a deep learning architecture while road
analysis is achieved by a combination Utkarsh Shukla etimage
of different al. / Procedia Computer
processing Science 165
pipelines. For(2019) 252–258
training our model for recognition 253
2
of road signs we incorporate aUtkarsh
huge Shukla
varietyet al. Procedia Computer Science 00 (2019) 000–000
of /traffic signs used by different road work departments. This is done
to
aimachieve a generalized
to aid in recognition
the process of model.
driving. The figureTraining models
1 describes on such acapable
a framework huge variety
of roadofsigns
data recognition
cause a reduction in the
and analysis
accuracy, so the deep learning architecture is build to encounter a variety of traffic signs like performed
involving lane and vehicle detection. The road sign recognition consists of a deep learning architecture while road in [1] [2].The
next taskisaims
analysis at theby
achieved detection of different
a combination vehiclesimage
of different that are in motionpipelines.
processing on the roadForand analyze
training ourthe lanefor
model of recognition
the moving
vehicle.
of road signs we incorporate a huge variety of traffic signs used by different road work departments. This is done
to achieve a generalized recognition model. Training models on such a huge variety of data cause a reduction in the
accuracy, so the deep learning architecture is build to encounter a variety of traffic signs like performed in [1] [2].The
next task aims at the detection of different vehicles that are in motion on the road and analyze the lane of the moving
vehicle.

2. Literature Survey

Considerable amount of work has been done towards driver assisting technologies. Some recent work include a
framework with the aim of detection, tracking, and recognition of different road signs. For detection of the road sign
a2.hear-cascade
Literature Surveybased system coupled with adaboost was used [3]. The classification of the detected road signs was
performed using a Bayesian classifier The main focus of this paper is a joint modeling of colour and shape within the
AdaBoost framework.
Considerable amount In of
thework
paperhas Robust
been method
done towards for road sign assisting
driver detectiontechnologies.
and recognition a three
Some recentfold framework
work include isa
proposed. The first part aims at detection of road signs by leanings from some prior
framework with the aim of detection, tracking, and recognition of different road signs. For detection of the road sign information of the scenario or
colour. The second part focuses on geometric analysis of the detected edges in the
a hear-cascade based system coupled with adaboost was used [3]. The classification of the detected road signs was first step, which generates candidates
to be circular
performed anda triangular
using signs. The
Bayesian classifier Thelastmain
aspect focus of of
thethis
framework
paper is aisjoint
a recognition
modeling phase thatand
of colour performs validation
shape within the
using cross-correlation methods. [4] The post processing part of the framework
AdaBoost framework. In the paper Robust method for road sign detection and recognition a three fold framework is consists of time based integration
based
proposed.on Kalman
The firstfiltering.
part aims Inat[5]detection
Like [3]offocusroadissigns heavily put on using
by leanings from features
some prior based on colourofand
information theshape
scenarioof the
or
traffic signs for detection and identification of the road signs. They used Support vector
colour. The second part focuses on geometric analysis of the detected edges in the first step, which generates candidates machines as a classifier for
identification
to be circular of and thetriangular
different road
signs.signs.
The There recognition
last aspect process canisbea divided
of the framework into three
recognition phaseparts. First discrimination
that performs validation
according to the color of the pixel second traffic-sign detection by shape
using cross-correlation methods. [4] The post processing part of the framework consists of time based classification using support vectorintegration
machines;
and
based theonthird context
Kalman understanding
filtering. In [5] Like based
[3] on Gaussian-kernel
focus is heavily putSVMs. on using Thefeatures
quoted based
resultson puts lightand
colour on high
shapesuccess
of the
rate and a very low amount of false positives in the final recognition stage. Wen-Jia
traffic signs for detection and identification of the road signs. They used Support vector machines as a classifier Kuo and Chien-Chung Lin in [6]
for
proposed a two
identification ofstep framework
the different road forsigns.
detection
Thereand recognition
recognition of thecan
process road besigns.
divided The first
into stepparts.
three is a detection step which
First discrimination
uses Hough
according to transformation,
the color of the pixelcornersecond
detection, and projection
traffic-sign detectionto byestimate the locationusing
shape classification of thesupport
road sign
vector in machines;
the image
under noisy and complicated environment. The recognition task uses tree based
and the third context understanding based on Gaussian-kernel SVMs. The quoted results puts light on high success approach where convolution, (RBF)
deep neural network and K-dtree are implimented to recognise the road signs in two
rate and a very low amount of false positives in the final recognition stage. Wen-Jia Kuo and Chien-Chung Lin in [6] stages.
In [7] aa two
proposed newstep traffic sign detection
framework systemand
for detection wasrecognition
proposed of that
thesimultaneously
road signs. Theestimates
first step the
is a location
detectionand stepprecise
which
boundary of traffic signs using convolutional neural network (CNN). It solves the
uses Hough transformation, corner detection, and projection to estimate the location of the road sign in the image problem of most of the method that
only provide bounding boxes of traffic signs as output, and hence requires
under noisy and complicated environment. The recognition task uses tree based approach where convolution, (RBF)processes such as contour estimation or
image segmentation to obtain the precise boundary of signs. In
deep neural network and K-dtree are implimented to recognise the road signs in two stages. [8] Songwen Pei1, Fuwu Tang1, Yanfei Ji1, Jing Fan1
andInZhong
[7] a Ning proposed
new traffic signthe use of Multiscale
detection system wasDeconvolution
proposed thatNetworkssimultaneouslyto solveestimates
the problem of the huge
the location and amount
precise
of time taken for preprocessing the images and applying complicated algorithms
boundary of traffic signs using convolutional neural network (CNN). It solves the problem of most of the method for improving and finding blurred
that
and
onlysubpixel
provide images
bounding of the
boxessigns. Multi-Scale
of traffic signs as Deconvolution
output, and hence Networks (MDN),
requires smoothly
processes suchconflates
as contour multi-scale
estimation conv
or
nets
image with deconvolution
segmentation network,
to obtain resulting
the precise in an efficient
boundary of signs. andInrobust localizedPei1,
[8] Songwen traffic
FuwusignTang1,
recognition
Yanfeimodel training.
Ji1, Jing Fan1
In
and[9] is presented
Zhong a road signs
Ning proposed the userecognition and classification
of Multiscale Deconvolution system focused
Networks to on a three-step
solve the problem algorithm consisting
of the huge amount of
color segmentation, shape identification, and a deep neural network architecture.
of time taken for preprocessing the images and applying complicated algorithms for improving and finding blurred The ultimate aim of the algorithm is
to recognise and distinguish varied road signs present along Italian roads. The
and subpixel images of the signs. Multi-Scale Deconvolution Networks (MDN), smoothly conflates multi-scale conv system is proposed to achieve real time
application. The shape detection
nets with deconvolution network, is achieved
resulting in by using two
an efficient anddifferent model oftraffic
robust localized pattern matching
sign recognitionand model
the other using
training.
edge detection and geometrical cues. Radu Timofte Karel Zimmermann Luc Van
In [9] is presented a road signs recognition and classification system focused on a three-step algorithm consisting ofGool in [10] proposed a multiview
traffic sign detectionshape
color segmentation, which uses a multidimentional
identification, and a deep neural algorithm
networkto augment
architecture. results
Thebeyond
ultimatetheaim state-of-the-art.
of the algorithm Theis
focus was to shift the detection method which still focuses on single view detection.
to recognise and distinguish varied road signs present along Italian roads. The system is proposed to achieve real time A speedup in the process is
achieved
application. through
The shapea novel boundedis evaluation
detection achieved by of using
ensemble two AdaBoost
different modeldetectors. The 2D
of pattern detection
matching in the
and multiple
other views
using
are combined to estimte a 3D hypotheses.
edge detection and geometrical cues. Radu Timofte Karel Zimmermann Luc Van Gool in [10] proposed a multiview
traffic sign detection which uses a multidimentional algorithm to augment results beyond the state-of-the-art. The
focus was to shift the detection method which still focuses on single view detection. A speedup in the process is
achieved through a novel bounded evaluation of ensemble AdaBoost detectors. The 2D detection in multiple views
are combined to estimte a 3D hypotheses.
254 Utkarsh Shukla et al. / Procedia Computer Science 165 (2019) 252–258
Utkarsh Shukla et al. / Procedia Computer Science 00 (2019) 000–000 3

Figure 1: Framework capable of road signs recognition and analysis involving lane and vehicle detection

3. Data Descriptions

For the purpose of detection of signals from the roadside, the model is trained over standard images of differ-
ent Traffic signals. The dataset used is the German Traffic Sign Benchmark used in the paper . The dataset has the
following features

• The dataset contains in total 43 different categories of road sign which are all used for recognition
• More than 50,000 images in total, depicting ground-truth data.
• Physical traffic sign instances are unique within the dataset ( each real-world traffic sign only occurs once).

For road analysis, KITTI Vision benchmark dataset is used. The dataset was collected by equipping a standard station
wagon with two high-resolution color and gray scale video cameras. Dataset was collectied by driving around the
mid-size city of Karlsruhe, in rural areas and on highways.

4. Proposed Method

The framework has been divided into two objectives first of recognizing road signs, the second part is the analysis
of the road. For meeting the first objective we dive into Deep Learning approach of Convolution Neural Network.
We first compute our results over different architectures that are popularly used and then finalized over using transfer
learning for this purpose. The concept of transfer learning is been implemented in the domain of deep learning where
models trained for a task is re-purposed on a second related task. We use ResNet50 architecture present on Keras as
our base model for final implementation purpose. We customize this model by hyper-parameters optimization and
adding custom layers to use it with our use-case. The model displayed is just the customized architecture attached to
 Utkarsh Shukla et al. / Procedia Computer Science 165 (2019) 252–258 255
4 Utkarsh Shukla et al. / Procedia Computer Science 00 (2019) 000–000

the pre-trained Resnet model. The second objective has to first deal with the main task of road analysis as it plays a
vital role in roadside vehicle detection while traveling. This process involves two sub-processes of:

• Analysing different lanes of roads and identifying the current lane so it can be used for determining future
maneuver
• Tracking the vehicles and objects around the car.

At first, the focus is on finding the track on which our vehicle is running or the track which is vacant that is having
no vehicles in front. Then we aim at detecting the lanes in a video frame with cars present on them.

Figure 2: Convolution Neural Network Architecture

5. Experimental Setup

This section gives a detailed information about the experimental setup involved in the

5.1. Road Sign Detection

5.1.1. Data Prepossessing

Image Data before it can be used by the neural network are pre-processed. The efficiency of a deep learning models
depends vastly on the kind of the data used which in turn determines the effectiveness of feature representation leaned
by the model. First of all the images is cleaned by denoising the image samples using the total variation filter combined
by wavelet denoising and bilateral filters. Denoising images result in posterized images with flat domains separated
by sharp edges. Degree of posterization is controlled by considering the trade off between denoising and faithfulness
to the original image. The images are then converted into grey-scale which is then resized into [50 X 50] pixel values.
Even though the image size is reduced it contains good representation of the data, the resizing is necessary as lower
the size less computationally expense also by converting RGB representation of the images to grey scale reduces the
depth of image which again reduces the expense.
256 UtkarshUtkarsh Shukla
Shukla et et al. / Procedia
al. / Procedia ComputerComputer Science
Science 00 (2019)165 (2019) 252–258
000–000 5

(a) Learning Curve (b) Loss Curve

Figure 3: Model Training Report

5.1.2. Model Construction

The neural network is a derivative of the Resnet50, initialized with the pertained weights from Image-Net. The last
fully connected layer is removed and combine it with three convolution block which is defined by a combination of
single convolution layer, batch normalization max pooling and then a dropout layer as shown in figure 4. The figure
2 shows the proposed architecture and the arrangement of the convolution blocks with the aim of learning the proper
representation of the training images. Batch normalization reduces the amount by which the hidden unit values shift
during training. Maxpooling layer is simply defined as trans-versing of a moving window on a 2D input space, where
the max magnitude within that window is the output. It reduces the dimension of the 2D array space, the reduction is
dependent on the size of the window that we choose to convolve over the 2D input matrix. Dropout is a technique used
to reduce over-fit on neural networks. While training a certain amount of neurons on a particular layers are deactivated.
This is done to introduce generalization because it forces the layer to learn with different neurons.

5.1.3. Training and Testing of the Model

The dataset is been divided into training testing and validation with the ratio of 70:20:10 respectively. Validation
data is used to alter the hyperparameters used by the convolution network. While our final results are cited over the
testing data. Each split of the data of training, testing and validation have equal representation of the different classes
this is done by a custom splitter. The model is trained for 500 epochs with a batch size of 128 images. The metric for
which the neural network optimizes the weights is categorical accuracy which checks if the index of the maximal true
value is equal to the index of the maximal predicted value. The optimizer used is Adagrad with the default parameters
provided with Keras it vriest its weight based on the variation of the model parameter, The learning rate changes
based on how frequently the weights gets changed during training. More the variation and frequency of updates of a
parameter, the smaller the updates by the optimizer. In Figure 3 we show the learning curve and the error rate curve
for our model. The blue line denotes the performance while training while the orange shows for the validation set,
from the graph it’s clearly evident that the model doesn’t overfit on the data and the learning is pretty generalized.

5.2. Lane and Vehicle Detection

5.2.1. Camera Calibration

The feed from the cameras mounted over the vehicle is passed to the second pipeline where geometrical analysis
computes the parameters of an image sensor and lenses of the video camera. The parameters are then used to estimate
lens distortion, calibrate the object size, or narrow down the location of the camera in the scene.

5.2.2. Edge Detection

Canny Edge detection is the most frequently used edge detection algorithm where we determine the intensity
gradient of the Image and also determine the magnitude and direction of the gradient, a compete analysis of an data is
6 Utkarsh Shukla et al. / Procedia Computer Science 00 (2019) 000–000

Utkarsh Shukla et al. / Procedia Computer Science 165 (2019) 252–258 257

(a) Convolution Block (b) Combined Convolution Block

Figure 4: Neural Architecture with Convolution Block

performed to eliminate any noisy pixels which does not constitute the edge. This is used to highlight the edges of the
different lane of roads

5.2.3. Perspective Transformation

Is used to perform varied geometrical transformations on two dimensional space. The aim is to prevent any alter-
ation of the image content but to ac hive deformation of the pixel grid and map this deformed grid to the deformed
image. An approximate road curvature equation is defined below which A and B are constants and y is a variable
F(y) = Ay2 + By + C. (1)
The above equation is for finding the radius of curvature from any arbitrary point on road. The equation for finding
the radius of curvature used is.
3
(1 + ( ddyx )2 ) 2
Rcurve = (2)
2A
This helps in determining the angle at which the car has to turn and at what distance. Therefore if there is no turn then
the radius of curvature will have a very large value or if having a very steep turn, this value will be very small. The
equations given above helps in determining the turn in the road and the distance of vehicle from that turn.

5.2.4. Vehicles Detection

For this purpose we use (HOG) which is a descriptor used in image processing in order to achieve object local-
ization. The algorithm enumerates occurrences of gradient orientation in localized portions of an image. It is a scale
Invariant feature. The algorithm is provided with a huge set of training data which contains images of vehicles and non
vehicles. The idea behind doing this is to determine a set of feature that distinctively represents the training sample
and can be used by machine learning algorithms for distinguishing the two classes. As a feature set we combine HOG
features with Spatial Binning features and Histograms of Color features. These extracted features are then passed to
a SVM with a RBF kernel for training so that it can now distinguish between vehicles and non-vehicle frames. This
process is a subset of the lane identification task as it helps the algorithm to get aware of the vehicles surroundings.

6. Results and Conclusions

After training the proposed deep learning architecture for road sign detection for 500 epochs as shown in 3 we get
an accuracy of 98.21% with a precision of 93.94%. The lane detection workflow was able to detect lanes, objects and
vehicles and in both still images and moving video frame figure 5 with total time computation of 64 frames per second
with a complexity of
O(n2 ) (3)
The model after it has been trained is predicts a batch of 128 images in 3secs. The model size ranges from 56.8 MB to
57.6 MB based on the amount of data used to train it. Since the size of the model is pretty less it is easily deployable
 258 Utkarsh Shukla et al. / Procedia Computer Science 165 (2019) 252–258
Utkarsh Shukla et al. / Procedia Computer Science 00 (2019) 000–000 7

(a) Road Sign Detection (b) Lane Detection

Figure 5: Results

and the lane detection uses only matrix operations which are heavily optimized to perform on low level hardware. If
GPUs can be used on the vehicles algorithms like MAGMA can hugely reduce the time and computational expenses
for lane and object detection.

7. Future Work

The future of this work consists of adding modularity to the implementation of the different objectives, setting of
better hardware compatible coding standards. Also we have only proposed the algorithm for multi-lane roads which
suggests the driver to be on the same lane. There are many roads which are single lanes and the traffic is from both the
sides on the same lane, specially in India, therefore an improvement on single lane roads is required. This task may
involve various new complexities. Also we can extend this work to bad weather conditions where its hard to predict
the lane while driving.

References

[1] B. Huval, T. Wang, S. Tandon, J. Kiske, W. Song, J. Pazhayampallil, M. Andriluka, P. Rajpurkar, T. Migimatsu, R. Cheng-Yue, et al., “An
empirical evaluation of deep learning on highway driving,” arXiv preprint arXiv:1504.01716, 2015.
[2] C. Chen, A. Seff, A. Kornhauser, and J. Xiao, “Deepdriving: Learning affordance for direct perception in autonomous driving,” in Proceedings
of the IEEE International Conference on Computer Vision, pp. 2722–2730, 2015.
[3] C. Bahlmann, Y. Zhu, V. Ramesh, M. Pellkofer, and T. Koehler, “A system for traffic sign detection, tracking, and recognition using color,
shape, and motion information,” in IEEE Proceedings. Intelligent Vehicles Symposium, 2005., pp. 255–260, IEEE, 2005.
[4] G. Piccioli, E. De Micheli, P. Parodi, and M. Campani, “Robust method for road sign detection and recognition,” Image and Vision Computing,
vol. 14, no. 3, pp. 209–223, 1996.
[5] S. Maldonado-Bascón, S. Lafuente-Arroyo, P. Gil-Jimenez, H. Gómez-Moreno, and F. López-Ferreras, “Road-sign detection and recognition
based on support vector machines,” IEEE transactions on intelligent transportation systems, vol. 8, no. 2, pp. 264–278, 2007.
[6] W.-J. Kuo and C.-C. Lin, “Two-stage road sign detection and recognition,” in 2007 IEEE international conference on multimedia and expo,
pp. 1427–1430, IEEE, 2007.
[7] H. S. Lee and K. Kim, “Simultaneous traffic sign detection and boundary estimation using convolutional neural network,” IEEE Transactions
on Intelligent Transportation Systems, vol. 19, no. 5, pp. 1652–1663, 2018.
[8] S. Pei, F. Tang, Y. Ji, J. Fan, and Z. Ning, “Localized traffic sign detection with multi-scale deconvolution networks,” in 2018 IEEE 42nd
Annual Computer Software and Applications Conference (COMPSAC), vol. 1, pp. 355–360, IEEE, 2018.
[9] A. Broggi, P. Cerri, P. Medici, P. P. Porta, and G. Ghisio, “Real time road signs recognition,” in 2007 IEEE Intelligent Vehicles Symposium,
pp. 981–986, IEEE, 2007.
[10] R. Timofte, K. Zimmermann, and L. Van Gool, “Multi-view traffic sign detection, recognition, and 3d localisation,” Machine vision and
applications, vol. 25, no. 3, pp. 633–647, 2014.