Anda di halaman 1dari 5

International Journal of Information Technology (IJIT) – Volume 4 Issue 4, Jul - Aug 2018

RESEARCH ARTICLE OPEN ACCESS

RGB Color based Myanmar Traffic Sign Recognition System from


Real-time Video
Kay Thinzar Phu [1], Lwin Lwin Oo [2]
Ph.D Student [1]
Faculty of Natural Science [2]
University of Computer Studies
Mandalay

ABSTRACT
In this paper color based image segmentation method and an effective structural feature points selection algorithms are
proposed in four-stage architecture including localization, detection, feature extraction and recognition. In the driving
environment, the traffic sign colors are seem as same but it is varied due to light conditions which make difficult problem for
traffic sign detection. The system presents using RGB color based thresholding method for traffic signs detection and shape
verification followed by proposed features point extraction. Lastly, adaptive neuro fuzzy inference system (ANFIS) is deployed
for classification to achieving results in real-time. This system provides good results in sunny, drizzle rain and cloudy weather
conditions at day time. The main objective of this system is to develop an effective Myanmar traffic sign detection and
recognition (TSDR) system.
Keywords — Adaptive Thresholding based on RGB color, ANFIS, Myanmar Traffic Signs, Structural Feature Points..
the same objects appear differently when the lighting changes.
I. INTRODUCTION However, the human visual system perceives a constant color
The advent of cars around us, this has result in an increased under the changes of illumination. It has been shown to be
traffic accidents, traffic crashes and traffic jams. There is effective due to changes in lighting to detect traffic signs
nowadays technology development for advanced driver coded by colors. Some researchers have proposed a method to
assistance system (ADAS). Advanced Driver Assistance overcome these difficulties for the TSDR system in real time.
Systems (ADAS) are becoming more and more common in This is less motivated than many researchers have presented
modern vehicles to provide assistance in driving situations. In all types of road signs in the TSDR system in real time.
particular, the methods of the computer field have been Therefore, this ongoing system offers data sets of Myanmar
emerging as important techniques for Intelligent Vehicle traffic signs for the recognition of traffic signs in real time.
Systems including pedestrian detection, lane detection, vehicle There are three types of Myanmar traffic signs: red, blue,
recognition and recognition of traffic signs. Since ADAS are and yellow color, and shapes are circle and diamond. First,
directly related to the safety of drivers and pedestrians. In prohibitory signs are placed to prohibit drivers on the road
ADAS, traffic sign detection and recognition (TSDR) system restrictions such as speed limit signs. Moreover, warning signs
is a key technology to become safe driving environment. The provide the dangers on the road situations to drivers so drivers
main objective of TSDR is to enable drivers to understand all can predict them, such as junctions, slow down, and ordered
the visual information of traffic signs they receive while signs are presented drivers what it must do. In this paper, we
driving. In general, the most important information is focus on red, blue and yellow color traffic signs detection step.
provided by signs posted in the driver's field of view. Traffic There have been a few publicly available traffic sign
signs are designed to assist the driver during his travel in an datasets. Many researchers proposed advances in traffic sign
efficient and secure way. The accuracy and short processing detection for German Traffic Sign Detection Benchmark.
time are extremely important for traffic sign recognition [ 16]. TSDR (Traffic Signal Detection and Identification System)
TSDR system needs achieving robust recognition: there is no consists of two steps: detection and identification. In detection,
time for complex transformations or sophisticated image the visual information is extracted from the image of the
processing techniques. They need a short processing time and sequence. By using the optical information specific to road
real-time analysis of a situation. signs, this section allows to extract the parts of the image that
Traffic signs have unique properties that distinguish them may contain the signs of points (Interest region). After a
between different countries that have signed a treaty. The traffic sign is detected within an image, the features of image
variations apparently are not important for a human, but could are calculated. The extracted features are inputted into
pose significant challenges for a computer vision algorithm learning machine that produces a traffic sign indication.
[8]. The main problem with the detection of traffic signs is So far, research on recognition systems has focused on
that the images captured by a camera cannot keep the color European labeling, but signs may look very different in
constant when the lighting changes. These problems occur different parts of the world. For the success of the detection
when it rains, cloudy or at night. Because the color of an process, methods based on colors and shapes are important.
image depends on the reflectance of the object and the lighting, Win and Myint are converted from the RGB color model to

ISSN: 2454-5414 www.ijitjournal.org Page 12


International Journal of Information Technology (IJIT) – Volume 4 Issue 4, Jul - Aug 2018
the HSL color model, and the color filter is used to extract Radial Basis Function Support Vector Machine Classifier
traffic signs from the candidate image. For shape analysis, the provided better results than other classifiers [14].
Harris corner detector is used on Myanmar traffic signs. And The researchers of [2] detected only the red circular frames
then the Sobel edge detector is used to separate the traffic and Romdhane et al. 25 red traffic signs from the GTSDB data
signs. Finally, the sorted road signs correspond to the images set were also detected. The authors of [5] proposed GTSDB
stored in the database using the template mapping method [1]. datasets for circular traffic images, and researchers [6] and [9]
The gray image is converted into a binary image. They used can also detect only red traffic signs and speed limit signs.
the cycle search algorithm to detect only circular objects. And This article is presented as follows: In the first section,
six attributes that will be given to ANFIS to be recognized. describes the most recent work of the TSDR system. After
They received a satisfactory result that led to the use of the presenting the second section which is the detail method of
ANFIS system for identifying road signs [2]. The authors [3] propose system, shows the techniques we have chosen. Next,
proposed methods of detecting occluded Korean traffic signs presents the some experimental results of our system and
and recognizing speed-limit signs in real-time: RGB finally includes the conclusion..
Normalized with fixed threshold and ROI Detection (size,
aspect ratio and color compositions). Neural Network is used II. PAGE LAYOUT
for speed limit signs recognition. Researchers [4] have applied The proposed system composes of three main steps:
the proposed algorithm to detect bandwidth as a new area of detection step, feature extraction step, and recognition step.
interest and have used a threshold in each of the HSV Detection step has two components: color detection process
components. Then they used a closed morphology and a and shape identification process. In the first process, the
characteristic frame (height, width, area) for all possible areas. interest region that is traffic sign localize from the image
The HOG features are calculated and a SVM classifier is used frame. RGB color thresholding method is used to each frame
to identify traffic signs. They only proposed red traffic signal from video sequences in order to classify the candidate traffic
images in the GTSDB [4] database. The authors [5] proposed sign. The output of color detection process includes a binary
an RGB-based color-defining technique with a developed image. The second process applies shape validation process
threshold and circle detection algorithm. They combined the on a binary image to decide whether traffic sign or not. The
features of Gabor, local bit patterns and a graded gradient candidate signs are validated which are input of the feature
histogram in a vector carrier classification. In [6], the extraction step of the system. Lastly, the system recognizes
researchers improved the Weiner filter and the equilibrium the extracted features points and gives the results.
histogram. The area of interest is detected using the enhanced
linked tagging algorithm in conjunction with a vector to A. Detection
reduce the calculation time and use the histogram to match the The detection process is used to segment the image into
detected image to other images stored in the image database. two parts, the part that contains the ROIs and the part that
Aparna et al. the total threshold used in RGB color. The contains the rest of the objects in the road scene.Traffic signs
linked element algorithm is used to form the return on are always with different background colors and figures.
investment and compares the initial return on investment with Therefore, the region that does not include traffic signs are
the new ROI for the form's analysis. The attributes used here need to be removed from the original image. The following
are the correlation, the mean value, and the red and black approach is used to localize the traffic signs region. Traffic
color histogram values. Support carrier machine and OCR are sign region is extracted from one forth to two forth regions of
used for pictograms and characters [9]. This article proposed a row and right half parts of the column. The following figure 1
method of detecting and recognition Bangladeshi traffic signs. shows the candidate region by using proposed approach. The
They used specific threshold values for each of the red, blue different threshold values are selected dynamically over the
and green colors. An average filter and the largest width, image is called adaptive thresholding. There are various
center and area selected and marked with the specific ROI [10] techniques by which adaptive form of thresholding
are used. The authors presented RGB color thresholding and selection can be utilized, but the aim of all the ways is
four cascaded geometric detectors in detection step. to find the appropriate threshold for different regions in
Morphology-based segmentation method is used for traffic the image [7]. Firstly, the input image frame is applied
sign segmentation from detected image and proposed 10 contrast enhancement method to increase the traffic signs
features are classified by k nearest neighbour classifier. This color. The pure yellow colour is by combining equal values in
proposed U.S speed sign system is not optimized for real-time the red and green channels and zero in the blue channel. So,
[11]. The researchers proposed for 36 classes Irish traffic yellow colors are detected by subtracting blue color intensity
signs images by CIE-LAB colour space method is used for values from red and green color intensity values [15]. After
detection and blob by blob analysis method is for shape getting the intensity values for each pixel by using adaptive
detection. Histogram of Gaussians (HOG) is extracted from form, the maximum threshold value is selected over this
validated traffic sign images and Linear Discriminant image intensity values. The maximum threshold value is
Analysis (LDA) is performed for dimensionality reduction. reduced to 35% because to detect the actual yellow color
They tested HOG features and LDA with three types of SVM traffic sign. The following equation (1) is presented for
classifires and Nearest Neighbors classifier and presented detecting yellow color traffic signs.

ISSN: 2454-5414 www.ijitjournal.org Page 13


International Journal of Information Technology (IJIT) – Volume 4 Issue 4, Jul - Aug 2018
AspRatio =
(7)

where BDWidth is the width, and BDHeight is the height of


the detected traffic sign. AspRatio is the aspect ratio.
AbsRatio is the absolute value of aspect ratio.
(a) Original Frame (b) Localized frame
B. Feature Extraction
Fig. 1 A localized image frame using proposed traffic sign region techniques
The particular form of a traffic signs distinguish it from
many other objects present in the environment. Therefore,
selecting an optimally discriminative feature set for a large
(1) number of traffic sign images is a very important task. The
main four features: crossing points, maximum centroid points,
 YI (i , j ) 1 if (max(R(i , j )  B(i , j ) )  TSrb )  termination and bifurcation points are extracted from

YI (i , j )  (max(G(i , j )  B(i , j ) )  TS gb ) validated images. This TSDR system gets the accurate results
YI  0 by using the proposed features.
 (i , j ) else
1) Crossing Point: It counts the number of transitions
The blue color has value in the blue channel and zeros in red from background to foreground pixels along vertical and
and green channels. Therefore red and green values are horizontal all lines through the image [13]. The following
extracted from RGB channel for detecting blue traffic sign. figure 2(a) shows the horizontal and vertical crossing points.
The following equation is used to find the blue color traffic In this system, the detected traffic sign is divided into four
signs in the input image frame. lines: one third and two third regions of row and column. The
average position values of crossing points on each divided line
are calculated for feature points. The following figure 2(b)
(2) shows the four cross points of the output traffic sign from
detection step [15].

 BI (i , j )  1 if (max( B(i , j )  R(i , j ) )  TSBrb )  VC1 VC1



BI (i , j )  (max( B(i , j )  G(i , j ) )  TSBgb )
 BI  0 HR1
 (i , j )
The red pixel values are greater than blue and green pixel HR2
value in blue color traffic sign. So, we calculate blue color
traffic signs using the following equations. Fig. 2. Crossing Feature Points

2) Centroid Point: The centroid points are the center


point of maximum row and column of detected binary sign
(3) image. In this system, the center point of row and column
 RI (i , j )  1 if (max( R(i , j )  G(i , j ) )  TSBrg )  points are extracted. The following figure shows the row and
 column centroid points of binary traffic sign image.
RI (i , j )  (max( R(i , j )  B(i , j ) )  TSBrb )
RI (i , j )  0

where R, G, and B are the red, green, and blue color
intensities value with range from 0 to 255. TSrb, TSgb, TSBrb,
TSBgb, TSRrb, TSRgb are the threshold values for yellow, blue
and red color traffic signs. In this equation , the threshold
Fig 3 Centroid Point for Left Turn Sign
value is multipied by 0.55 for blue and 0.75 for red by our
experimental results. 3) Termination Point: If any of the 1-connected pixel
value is one, this point is called termination point. The
1) Shape Validation: In this proposed work, aspect ratio following equation is used to find termination points in
1 if ((Cen==1) ^ sum
validated traffic sign [15].
method is used to validate the detected image is traffic sign or (N==1))
not. The following equation is used to calculate the shape of
the candidate region [15]. Tr =
0 else

ISSN: 2454-5414 www.ijitjournal.org Page 14


International Journal of Information Technology (IJIT) – Volume 4 Issue 4, Jul - Aug 2018
the advantages of the rule-base fuzzy system. It combines the
where, Tr =Termination point, Cen= Center point in 3x3 explicit knowledge representation of a FIS with learning
matrix, N=3x3 matrix. capabilities of neural network. The objective of ANFIS is to
integrate the framework of an adaptive fuzzy and neural
4)Bifurcation Point: If any of the 4-connected or 3-
network. The 10 inputs represent the proposed features
connected pixel values are 1s, this point is bifurcation point.
calculated from each traffic sign image. Each input was given
The following equation is used to find the bifurcation points
trapezoidal membership functions and number of membership
from output sign board [15].
functions used 2 for input features. The output was
1 if ((Cen==1) ^ sum represented by output label for each input features set (eg. 1 is
Bif = (N==3)) assigned right turn sign, 2 is assigned straight down sign). The
0 else outputs of the rules are condensed into one single output,
where, Bif =Bifurcation point, Cen= Center point in 3x3 representing that system output for that input image.
matrix, N=3x3 matrix.
The following figures show the termination points and III. EXPERIMENTAL RESULTS
bifurcation points that are extracted for left turn blue color
In our system, we collect data through the use of
traffic sign.
smartphones on highways and highways. The different speeds
(40 km/h and 60 km/h) of the vehicle covered the video. The
resolution of the video is 720x480. The frame rate of the video
is 32 frames / second. In our experiments, this proposed
method can detect between 60 and 80 feet ahead of traffic
(a) Candidate image (b) 3 termination points and 1 bifurcation point signs. Therefore, drivers reduce and control their driving
Fig. 4 Termination and Bifurcation Points
speed in good time.
White blobs are detected in the binary image. The traffic
Some of the traffic signs produced the same number of sign regions are indicated by all blobs in the image. Blobs
termination and bifurcation points (eg. Left turn and Right turn were selected that have an area greater than 200 pixels and an
sign). This system calculates the average values of termination aspect ratio in the range of 0.8 to 1.2. The blobs that met these
and bifurcation points of row and column. So, four features restrictions were extracted the traffic signs shown in the
points for termination and bifurcation are extracted. The figures.
discussed image for ten features is presented in the following The system tested different row and column locations to
table 1. These features points are given as input to ANFIS for locate the range of traffic signs in the input frame. As we
recognition. increase the row and column position, the system detects the
area of the same color as the traffic signs and produces a
TABLE I
PROPOSED FEATURE POINTS FOR LEFT TURN SIGN
blurred image. And the row and column position was also
reduced; the system missed some of the traffic sign areas. The
Name Values proposed location approach is suitable for detecting the traffic
VC1 0.555556
signal region for this system. Further experiments were
VC2 0.555556
HR1 0.916667 conducted with different threshold values. The number of start
HR2 0.916667 detected frames of the threshold range decrease increase and
Cen_R 0.526315 the output zone contains no traffic signs so that it consumes
Cen_C 0.909090 the performance of the system. In this experiment, images
Ter_R 0.537037 which were taken in cloudy weather, drizzle rain and normal
Ter_C 0.666667
Bif_R 0.555555 weather conditions were used in order to analyze robustness
Bif_C 0.727272 of proposed algorithm. This algorithm was tested on a class
of 18 Myanmar Traffic Signs. Out of this, 12 classes
contained ‘Yellow Color “traffic signs, 6 classes contained
where, VC_1, VC_2, VR_1 and VR_2 are average values of “Blue Color” traffic signs and 1 class contained “Red Color”
crossing points of one third and two third regions of row and traffic signs. In this system, the various time lengths of 16
column , Cen_R and Cen_C are centroid point of row and
videos are used and accuracy is showed in the following table
column, Ter_R and Ter_C are row and column average values
2 and 3 of this proposed system. This system produced more
of termination points, Bif_R and Bif_C are row and
accurate results for the blue color traffic signs recognition by
column average values of bifurcation points. using proposed 10 feature points. The informative traffic signs
C. RECOGNITION detection and recognition results decreased. Also, the
prohibitive signs results slightly decreased by compared with
The adaptive neuro fuzzy inference system (ANFIS)
[15]. From the experimental resuts, the centroid point provides
classifier is used to recognize traffic signs by proposed
only good accuracy for blue color traffic signs but not for
features points are given. The ANFIS is a fuzzy Sugeno model
other two signs. This system produces recognized traffic signs
put in the framework of adaptive systems to facilitate learning
output as text with Myanmar language.
and adaptation [14]. The ANFIS is a FIS implemented with

ISSN: 2454-5414 www.ijitjournal.org Page 15


International Journal of Information Technology (IJIT) – Volume 4 Issue 4, Jul - Aug 2018
TABLE III [5] B. Selcan Kaplan, G. Huseyin, O. Ozgur, A. Cuneyt and
DETECTION RESULTS BY USING PROPOSED METHOD
G. Serkan, “On circular traffic sign detection and
Traffic Signs Test Images Correctly classified Accuracy recognition”, Expert Systems with Applications, 48,
signs pp.67-75, 2016.
Mandatory 285 277 98%
[6] D. Yousef-Awwad, and H. Ahmad M., “Accurate Real-
Informative 243 229 94%
Prohibition 160 156 98% Time Traffic Sign Recognition Based on the Connected
Total 688 662 96% Component Labeling and the Color Histogram
Algorithms”, In International Conference on Signal
Processing, 2015.
TABLE IIIII
RECOGNITION RESULTS BY USING PROPOSED FEATURE POINTS [7] P. Parul and A. Pawanesh, “Color Based Image
Segmentation Using Adaptive Thresholding”,
Traffic Signs Test Images Correctly Accuracy International Journal of Scientific and Technical
classified signs Advancements, Volume 2, Issue 3, pp. 151-156, 2016.
Mandatory 232 217 94% [8] D. Abhinav V, “A Brief Overview of Traffic Sign
Informative 222 184 84%
Detection Methods”, International Journal of
Prohibition 150 146 97%
Engineering Research. pp. 141-144, 2016.
Total 604 547 91%
[9] S. Aparna, and A. Deepa, “Multiple Traffic Sign
Detection and Recognition using SVM”, International
IV. CONCLUSIONS
Journal of Advance Research in Computer Science and
The TSDR system is an integral part of the driver Management Studies, Volume 3, Issue 8, August 2015.
assistance system. It warns and directs the situation of the [10] C. Soumen and U. Mohammed Nasir, “Bangladeshi road
road to the driver. Such a system must be fast and accurate to sign recognition based on DtBs vector and artificial
detect traffic signs from video sequences. This article neural network”. InElectrical, Computer and
proposes a new feature ten points for recognition and Communication Engineering (ECCE), International
detection method based on RGB color of traffic signs in real Conference on 2017 Feb 16 (pp. 599-603). IEEE.
time. This system was applied to the red, yellow and blue [11] A. Jafar, Z. Imad and M. Ayman M, “Speed sign
color of Myanmar road signs. This video is captured in the recognition using shape-based features”, International
morning and the weather foggy and drizzle. As described Journal of Computer Applications. 2013 Jan 1;84(15).
above, this system tested several thresholds and localized [12] E. Etienne and Student, “Irish Road sign detection and
positions. As a results, there is not sure to get good accuracy recognition based on CIE-Lab colour segmentation and
by using many feature points. It is need to use the most HOG feature extraction”, 2014.
effective and efficient feature points that provide accurate [13] K. Gaurav, and B. Pradeep Kumar, “A detailed review of
results for traffic signs recognition. In the future study, we feature extraction in image processing systems”, In
will make some important feature points and improve the Advanced Computing & Communication Technologies
processing time of the system. (ACCT), February 2014.
[14] J.-S.R. Jang, “ANFIS: Adaptive-network-based fuzzy
REFERENCES inference system,”IEEE Transactions on Systems, Man,
[1] W. Wint Sandar and T. Myint Myint, “Detection and and Cybernetics, vol. 23(3), pp. 665-685, 1993.
Recognition of Myanmar Traffic Signs from Video”, [15] P. Kay Thinzar and O. Lwin Lwin, “Traffic Sign
ISBN 978-93-84468-20-0, Proceedings of 2015 Recognition System Using Feature Points,” 2018 12th
International Conference on Future Computational International Conference on Research Challenges in
Technologies, (ICFCT'2015), Singapore. Information Science (RCIS). IEEE, 2018.
[2] B. Mustain, W. Sajjad, A. K Kawsar and H. Abu, 2015, [16] Y. Shouyi, et al., “Fast traffic sign recognition with a
“Real Time Traffic Sign Detection and Recognition rotation invariant binary pattern based feature”, Sensors.
using Adaptive Neuro Fuzzy Inference System”, 2015 Jan 19;15(1):2161-80.
Communications on Applied Electronics (CAE) - ISSN :
2394 – 4714, Volume 3 - No. 1, October 2015
[3] K. Jihie et al., “Real-time Korean traffic sign detection
and recognition”. In Proceedings of the 7th International
Conference on Ubiquitous Information Management and
Communication 2013 Jan 17 (p. 19). ACM.
[4] R. Nadra Ben, M. Hazar and H. Mohamed, “An
improved traffic signs recognition and tracking method
for driver assistance system”, In Computer and
Information Science (ICIS), 2016 IEEE/ACIS 15th
International Conference on (pp. 1-6). IEEE.

ISSN: 2454-5414 www.ijitjournal.org Page 16

Anda mungkin juga menyukai