Implementation of Autonomous Driving Vehicle at an Intersection

with Traffic Light Recognition and Vehicle Controls

Changhyeon Park

and Seok-Cheol Kee

Department of Smart Car Engineering, Chungbuk National University, Seowon-gu Chungdae-ro 1, Cheongju-si, Korea

Smart Car Research Center, Chungbuk National University, Seowon-gu Chungdae-ro 1, Cheongju-si, Korea

Keywords: Autonomous Driving Vehicle, Crossing/Stop Decision, Traffic Light Recognition (TLR), Coordinates Map,

Convolutional Neural Network (CNN).

Abstract: We implemented autonomous driving vehicle system at an intersection equipped with traffic lights. This

system was consisted of a traffic light recognition, crossing/stop decision algorithm, vehicle localization,

vehicle longitudinal/lateral control, and coordinate map generation. The traffic light recognition was

implemented by using camera-based CNN data processing. The crossing/stop decision algorithm decides

vehicle longitudinal control whether drive or not depending on recognized traffic light signal. The vehicle

localization was implemented by using RTK GNSS and dead reckoning. The longitudinal control was

light position/signal recognition results. The overall vehicle control system was implemented based on an

embedded control board. Coordinate map was made by saving vehicle’s position data received from RTK

GNSS. To evaluate the performance of proposed system, we remodeled a commercial vehicle into

autonomous driving vehicle and drove the vehicle on our proving ground. Our own proving ground for the

test vehicle driving performance was located in Ochang Campus, Chungbuk National University. As a result,

the proposed vehicle successfully drove at intersection equipped with traffic lights with a maximum speed of

40 kph on a straight course and a maximum speed of 10 kph on a 90° corner course.

1 INTRODUCTION

As the fourth industrial revolution has progressed in

recent years, research on autonomous driving vehicle

system is actively under way. Autonomous driving

vehicle system can be applied to various industrial

fields such as drone, factory robot, etc. However, the

autonomous driving vehicle system of the vehicle can

bring societal advantages such as convenience, safety

and economically efficiency of people. According to

the Korea police DB in the TAAS(Road Traffic

Authority, 2018), the number of traffic accidents in

the intersection area for one year in 2017 is 102,354.

The number of traffic accidents is a large proportion,

accounting for 47.3% of 2016, 335 cases. Therefore,

if an autonomous driving vehicle system that

guarantees safety of vehicle operation is implemented

at intersection, the proportion of traffic accidents can

be greatly reduced.

https://orcid.org/0000-0003-3612-0306

https://orcid.org/0000-0003-2907-2502

The autonomous driving vehicle system is

constructed by mimicking the driving method of the

actual vehicle driver. The intelligent sensor is

responsible for the driver’s ability to distinguish

objects around the car. Typical examples are Radar,

Lidar, Camera and Ultrasonic sensor. The driver’s

driving control, such as Steering, Brake and

Acceleration, is replaced by an electronically

controlled way. And the brain responsible for the

driver’s judgment is replaced by a high-performance

computing system.

Hardware systems that handle autonomous

driving and judging have already had the capability to

replace the usual driver’s abilities. However, the

vehicle sensor are not able to meet the driver’s

cognitive ability due to the physical limits of the

sensing method and the misdetection problem in the

poor driving environment. Radar sensors are capable

of precise longitudinal distance measurements. But it

542

Park, C. and Kee, S.

Implementation of Autonomous Driving Vehicle at an Intersection with Trafﬁc Light Recognition and Vehicle Controls.

DOI: 10.5220/0007760905420549

In Proceedings of the 5th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2019), pages 542-549

ISBN: 978-989-758-374-2

lack the ability to distinguish objects from lateral

distance measurement (Reina et al., 2015). The Lidar

has the advantage of being able to be used in both

daytime and nighttime situations, but it lacks low

resolution and mass production reliability verification

(Lim, 2014). The camera can acquire a lot of

information about the driving situation, but the image

quality deteriorates seriously due to lighting change,

weather change, and the like (Jo and Lee, 2015).

In this study, the autonomous driving vehicle

system based on the traffic lights recognition was

performed by using a monocular camera installed in

front of the vehicle. Therefore, It is necessary to

develop a TLR algorithm that can operate in a low-

quality camera image environment.

TLR algorithms are divided into Non-Learning

based algorithms and Learning-based algorithms.

Non-learning based algorithms recognize traffic lights

based on color data such as HSV or YCbCr (Jung and

Kim, 2017). When the camera recognizes the traffic

light, it captures a sky area that changes color

sensitively according to time and weather. Therefore,

In this paper, We implemented robust Learning-based

algorithms(Website-WIKIPEDIA, 2019) for various

environments. Also, as a feature of Autonomous

vehicle, the algorithm requires a model capable of real-

time processing, and details are described in ‘2.

TRAFFIC LIGHT RECOGNITION ALGORITHM’.

In the field of the Vehicle Localization, There are

various methods such as only GNSS, 3D Lidar with

IMU (Heo and Park, 2016), Ultrasonic (Kim et al.,

2007), Dead Reckoning (Jung, et al., 2012). 3D Lidar,

IMU and Ultrasonic methods are a solution to solve

the problem of cost and environmental constraints of

GNSS. And guarantee higher performance when used

with GNSS. In case of Dead Reckoning, reliability is

lower than other methods. But real-time is guaranteed.

In this study, We implement Autonomous driving

vehicle system in the area without GNSS obstruction.

Therefore, We constructed the system using GNSS

and to guarantee the real-time, Dead Reckoning is

added. The detail are described in ‘3. VEHICLE

LOCALIZATION’. Meanwhile, the generation of the

vehicle driving path is also implemented using the

GNSS method, and will be described in detail in ‘4.

DRIVING PATH GENERATION USING

COORDINATE MAP’.

The Vehicle Path Following is implemented by

inputting longitudinal and lateral control values to the

embedded control module equipped in the vehicle

based on the generated driving path. At longitudinal

and lateral control, there are various methods such as

PID, Pure Pursuit, Stanley Control, Kinematic Control,

LQR with FF Control, Preview Control (Snider, 2009).

Each of the algorithms has a different advantages in

various environment such as high speed situation, a

low speed situation, a high curvature situation, and the

like. In this paper, we use robust algorithms for low

speed and high curvature conditions considering the

characteristics of intersections. And detail are

described in ‘5. VEHICLE CONTROL SYSTEM

AND PATH FOLLOWING’.

The judgment of the driving state of the vehicle

depends on the recognition result of the TLR. In

particular, the yellow signal decision algorithm which

has two states of stop and go is implemented by using

the speed of vehicle, the distance between vehicle and

stop line, the length of the intersection. The detail are

described in ‘6. STOP/CROSSING DECISION

ALGORITHM’.

Figure 1: System block diagram summary.

In this paper, we used Hyundai Ioniq plug-in

hybrid vehicle for implementing Autonomous

Driving System based on vehicle. And the system is

consisted of the TLR, the Vehicle Localization, the

Driving Path Generation, the Path Following and the

Decision of vehicle driving signal. (Figure. 1)

2 TRAFFIC LIGHT

RECOGNITION ALGORITHM

In this paper, the TLR Algorithm is shown in Figure

2. It extracts the signal information from the front

camera, extracts signal information using CNN

model, detects proximity signal through size

comparison, class integration for the same signal, and

finally recognizes the final signal by extracting signal

information based on the cumulative probability

(Jang and Cho, 2017).

Figure 2: Traffic Light Recognition system block diagram.

Implementation of Autonomous Driving Vehicle at an Intersection with Trafﬁc Light Recognition and Vehicle Controls

543

2.1 Camera Calibration for TLR

The camera sensor is attached near the windshield at

the front of the vehicle. The camera external

parameter calibration is experimentally performed so

that the traffic light area from the intersection entry

area to the intersection passage area is included in the

field of view of the camera. Also, As consider

intersection entry speed of vehicle, traffic light

detection should be performed before 85m from the

intersection area. Therefore, 25.36° narrow angle lens

is used to detect long-range traffic light to improve

image resolution. And to shoot traffic lights in a

variety of outdoor lighting conditions, The Camera is

configured with a 1/1.8" sensor format and a

3.45x3.45um pixel size.

2.2 Classification Model

The recognition of the traffic lights in the autonomous

driving vehicle system requires robust performance in

the weather and lighting changing environment.

Especially, in the backlight situation, the traffic light

vanishing phenomenon is happened by sunlight.

In this environment, the algorithms need very high

recognition performance and reliability in the weather

and lighting changing environment.

In this study, we implemented a Deep Learning

algorithm using a large amount of training data

including general environment image data and poor

environment image data. Also, the autonomous

driving vehicle systems require real-time processing

algorithms. The recently introduced YOLO

V3(Joseph Redmon, Ali Farhadi, 2018) guarantees

robust recognition and real-time in situations with

complex backgrounds and external lighting changes.

Therefore, we adopted it as a detecting model.

2.3 Training Data Configuration

Table 1: YOLO V3 Ground Truth Data Information.

Type Block Type Signal Type Total

3 Signal block

Red 370

1 Yellow 242

2 Green 1119

4 Signal block

Red 4029

4 Yellow 96

5 Green 1914

6 4 Signal block Left Arrow / Green 1101

The images of traffic light used in the training were

taken while driving the vehicle. The images were

taken on a clear day, cloudy day and rainy weather.

The composition and number of the traffic light

images used in the experiment are shown in Table 1

(Park and Kee, 2018).

Training data classes, written Block Type and

Signal Type in Table 1, are divided into as many

classes as possible for the same traffic light signal.

And the ROI(region of interest) of the GT is

constructed by multiplying 1.2 times to width of

traffic light and multiplying 1.5 times to height of

traffic light.

2.4 Training Data Extending

The basic principle of Deep Learning is to train the

weights of neural networks according to training data

and to derive desired results. Therefore, training using

Deep Learning requires training data of various

environments. If the environment of the training data

is limited, the result will be reliable only in the trained

environment.

YOLO V3, a Learning based algorithm, was

implemented using the dedicated frame tool, Darknet

(Website-Joseph Redmon, 2018). The frame tools

provides a Data Augmentation (Salamon and Bello,

2016) methods that reinforces the kind of YOLO V3

input data. Data Augmentation methods modifies the

hue, saturation and exposure of the input training

image to create a new training image.

In this study, The Data Augmentation methods

was optimized for TLR. One of the features of the

TLR is that the position information of the signal can

be used when determining the signal type. Therefore,

in this study, we did not use only the flip option,

which is a symmetric technique, among all Data

Augmentation technologies that Darknet supports by

default.

2.5 Traffic Light Signal Detection and

Post Processing

The YOLO V3 algorithm applied to the TLR outputs

the size, position and signal type information of the

traffic lights. Generally, A large number of Traffic

Lights with different sizes are detected in one image.

Among the many Traffic Lights, what we need to

detect is the nearest one. Therefore, by comparing the

sizes, we decided what we have to recognition.

Refer to Table 1 and Figure 3. The Traffic Light

Signal Information consists of 7 type. And, It is

divided into block and signal type. The block means

the number of signal ball. And the signal means color

VEHITS 2019 - 5th International Conference on Vehicle Technology and Intelligent Transport Systems

544

of ball. In this study, we have integrated the multiple

same signals into represent one. For example, red

signal of 3 signal block and red signal of 4 signal

block is same.

Finally, to improve recognition rate, we

accumulate recognized result. And as a final detection

signal, we select the most frequent one among the

accumulated data. If we accumulate a lot of many

data, the TLR performance will be improved.

However, It requires a large amount of system load.

Figure 3: The result of TLR. The method of naming is TR_

[ B ]_[ C ]. TR mean Traffic Light Recognition. [ B ] mean

the number of circle in Traffic Light. And [ C ] mean sign

al of traffic light. (R : Red, LG : Left arrow and Green, Y :

Yellow).

3 VEHICLE LOCALIZATION

Figure 4: Vehicle Localization System Block Diagram.

In this paper, the vehicle localization is consisted as

shown in Figure 4. This system simply implemented

by using RTK GNSS (Website-WIKIPEDIA, 2019)

and Dead Reckoning.

The GNSS sensor is attached to the vehicle and

receives information of (x, y, theta). However, since

it has a maximum error of 10m, RTK method that has

a maximum error 10cm is used. The error is a range

that can be used for vehicle localization. However,

since the reception frequency is 1 to 5 Hz, real-time

performance in the vehicle cannot be guaranteed.

Therefore, the localization of the vehicle between the

GNSS reception intervals is estimated using Dead

Reckoning. Although the accuracy of the Dead

Reckoning is lower than that of the RTK GNSS, it is

implemented using a wheel encoder and a gyroscope

that are guaranteed real-time. (Woo et al., 2009),

(Langley, 1998)

3.1 Vehicle Localization using RTK

GNSS

We use C94-M8P module to implement RTK GNSS.

And RTK GNSS is implemented using NTRIP.

Receive the GNSS correction signal of Korean

Geographical Information Service through internet

communication and calculate the exact position of

GNSS sensor. The GNSS sensor is attached to the

upper roof of the rear axle of the vehicle like Figure

5. The reception frequency setting and the NTRIP

setting of the GNSS sensor were performed using U-

CENTER software. In this study, we set the receive

frequency to 5Hz.

Figure 5: Attaching sensor for vehicle localization.

3.2 Vehicle Localization using Dead

Reckoning

Dead Reckoning is implemented using the angular

velocity and driving speed of vehicle output from the

embedded control module equipped for autonomous

vehicle. Implementation of the vehicle localization

via Dead Reckoning follows the equations (1), (2) and

(3).

θ ≈ Δθ × Δt (1)

new

= x

old

× Δc × cos(θ) (2)

new

= y

old

× Δc × sin(θ) (3)

At equation (1), (2) and (3), θ means heading of

vehicle. It is calculated using Yaw Rate Sensor value

indicated in Figure 5. And Δc means the distance

traveled of vehicle for a short time. (x

old

, y

old

) is the

previous position of vehicle and (x

new

, y

new

) is the

current position.

4 DRIVING PATH GENERATION

USING COORDINATE MAP

The Driving Path is generated through a pre-made

coordinate map. The coordinate map is composed of

Implementation of Autonomous Driving Vehicle at an Intersection with Trafﬁc Light Recognition and Vehicle Controls

545

points from the beginning to the end of the vehicle to

follow. Each point stores 7 kinds of information for

vehicle control

1) Address of points

2) Longitude

3) Latitude

4) Longitudinal Target Speed of vehicle

5) Stop Point Latitude

6) Stop Point Longitude

7) Stop Point Address

The points shown in Figure 5 are created based on the

GNSS sensor attachment point and uses the RTK

GNSS. At this time, RTK GPS reception frequency is

1Hz and the interval between points is about 2.78m.

5 VEHICLE CONTROL SYSTEM

AND PATH FOLLOWING

Figure 6: Hyundai Ioniq plug-in Hybrid used in the

experiment.

This paper implemented the system based on real

vehicle. The vehicle equipped with an embedded

control module and front camera used in the

experiment is shown in Figure 6. The vehicle control

system is shown in Figure 7.

Figure 7: Vehicle Control System block diagram.

The embedded control module inputs the

longitudinal acceleration and lateral steering angle of

the vehicle. In this paper, the longitudinal control of

the vehicle is implemented using the PD controller.

(Bae, 2013) And the lateral control of the vehicle is

performed using the Pure Pursuit controller. (Park

and Kim, 2015)

In this paper, the longitudinal PD controller

experimentally set the controller Gain to allow

smooth speed rise and stop. In addition, the I

component of the PID controller is not used because

the steady-state error is not significant. As the lateral

controller, Pure pursuit applied Look Ahead Point

method similar to real driver was used. The

calculation formula to get steering angle of vehicle

using Pure Pursuit is shown at (4).

= tan

−1

(k × (2



(α )) / ( + 5))

(4)

At equation (4), δ means steering angle of vehicle. k

means hyper-parameter. W

means wheel-base length

of vehicle. α means an angle between a heading of

vehicle and a Look Ahead Point.  means

longitudinal speed of vehicle.

6 STOP/CROSSING DECISION

ALGORITHM

The Stop/Crossing Decision Algorithm is an

algorithm that judges the driving and stopping

behaviour of the vehicle. And the system is

summarized in Figure 8 as block diagram.

Figure 8: Decision algorithm block diagram. [a] is a result

of adding (5) and (6). [b] is a result of (7).

The Stop/Crossing Decision Algorithm is

implemented based on the recognized signal, the

speed of the vehicle, the distance between the vehicle

and the stop line and the distance of the intersection

section. The recognized signal is finally output in

three states of Driving, Neutral and Stop. And each

means that the result of the TLR is green, yellow and

red.

VEHITS 2019 - 5th International Conference on Vehicle Technology and Intelligent Transport Systems

546

For the Neutral signals, the signal is determined

by the speed of the vehicle and the distance between

the stop line and the vehicle. If the sum of the

traveling time (5) to the intersection passing time (6)

is shorter than the holding time of neutral signal (7),

this algorithm determines it as a Driving signal. If not,

However, this algorithm determines it as a Stop signal.

In this paper, We implement this system at

Autonomous Vehicle Proving Ground equipped with

intersection located in Ochang campus, Chungbuk

National University.(Figure 9) As constant variable,

Therefore, we determines d

Intersection

as 7m,v

Intersection

10kph and t

NeutralTime

as 5 sec.

StopLine

= (d

StopLine

SignalDetection

)×3/2 (5)

Intersection

= (d

Intersection

)

(6)

For the Driving signal, the vehicle follows the stop

line of the traffic light. At this time, the longitudinal

target velocity is linearly reduced to zero.

7 EXPERIMENTAL METHODS

AND RESULT

This experiment is progressed in Autonomous

Vehicle Proving Ground with the intersection

equipped with traffic light located in Ochang campus,

Chungbuk National University.(Figure 9)

Figure 9: Autonomous Vehicle Proving Ground with Inters

ection equipped with Traffic Light.

This system was implemented with maximum

40kph in the straight section and maximum 10kph in

the corner section. The performance of this system is

shown in Table 2, 3, 4, 5 and Figure 10.

Table 2 shows the performance evaluation of the

TLR algorithm. It defines a sequence as a distance of

85 m from a traffic light to a point passing through an

intersection.

Table 3 and 4 shows the vehicle control

performance for lane departure and stopping at stop

point under real driving environment.

Table 5 shows the decision performance

according to the Stop or Crossing in the yellow state.

Figure 10 is a longitudinal and lateral control

performance graph. And This data was written as one

second intervals. At this Figure, Lateral Distance

Error means vertical distance error between the path

heading and the vehicle.

Table 2: The result of the TLR. The confusion matrix is

used for the performance evaluation, and it is defined as one

sequence from the distance of 85m to the moment when the

vehicle passes through the traffic lights (Park and Kee,

2018).

Type Sequence Result

Red / Yellow 30 30TP

Green 30 30TP

Green to Yellow 15 15TP

Red to Green 15 15TP

Recall, Precision 100%

Table 3: Vehicle Lateral Control Performance Result.

90° Corner

section

(kph)

Attempts

Success

(In Lane)

Failure

(Out Lane)

10 25 25 0

15 25 23 2

20 25 19 6

Straight

section

(kph)

Attempts

Success

(In Lane)

Failure

(Out Lane)

20 25 25 0

30 25 25 0

40 25 25 0

Table 4: Vehicle Longitudinal Control Performance Result.

It indicates the distance error between the stop line and the

vehicle when the vehicle is stopped.

Attempts Result

Stop Point Distance Error

(m)

Over

line

2 0.012 0.021 0.017

In line 23 0.011 0.108 0.032

Implementation of Autonomous Driving Vehicle at an Intersection with Trafﬁc Light Recognition and Vehicle Controls

547

Table 5: The result of intersection driving of autonomous

vehicle at Neutral Signal Decision State.

Straight

Course

(kph)

90°

Curved

Course

(kph)

Neutral Signal

Decision

Result

Success Failure

Stop 15 0

Crossing 15 0

Stop 15 0

Crossing 15 0

Stop 15 0

Crossing 15 0

Figure 10: The result of vehicle’s lateral, longitudinal

control according to vehicle’s velocity. Curved sections

mean 90° corner course.

8 CONCLUSIONS

In this paper, the autonomous driving vehicle system

was implemented using real vehicle equipped with

intelligent sensor and embedded control module at

intersection area equipped with traffic light. The

study was conducted on three areas: Traffic Light

Recognition, judgment of behavior of vehicle based

on recognized signals and vehicle control.

The system was successfully implemented at

Autonomous Vehicle Proving Ground located in

Ochang campus, Chungbuk National University.

At traffic light recognition and signal judgment,

the behavior of vehicle recognized as yellow signal

was mainly studied. And this content was written at

‘6. STOP/CROSSING JUDGMENT ALGORITHM’.

In this study, the holding time of the yellow signal

from traffic light was fixed as 5 sec. To implement

robust judgment algorithm, However, the systems

have to receive accurate holding time. For example,

if we use V2X methods, we can get the accurate

holding time to maintaining yellow signal.

At vehicle control, we implemented this system

using localization of vehicle and longitudinal/lateral

control. At part of localization, by using both RTK

GNSS and Dead Reckoning, ‘3. VEHICLE

LOCALIZATION’ was consisted. However, in this

study, lateral control was not stable due to GNSS

heading error which randomly changed maximally up

to 5

. Therefore, it is necessary to compensate the

problems by a technique such as a Kalman filter. Also,

it is difficult to getting trusting data from GNSS

located between high-rise buildings. Also, In case of

Dead Reckoning, Because it has low reliability

compared to GNSS, it is used as complementary

algorithm of GNSS. However, if it is implemented

through a high-quality algorithm, the performance of

the localization will improve.

At control of vehicle, PD controller is used to

longitudinal control. And Pure Pursuit controller is

used to lateral control. More details are covered in ‘5.

VEHICLE CONTROL SYSTEM AND PATH

FOLLOW’. In this study, the PD controller of

longitudinal control is implemented in the low speed

situation. Therefore, the stopping and running

performance cannot be relied on in high speed

situation. Actually, in real road, there are many

challenging environments, such as inclement weather,

high curvature road, and so on. And to comfortable

ride, accurate control is necessary. Also, the Pure

Pursuit controller applied to lateral control is not

stable for all environment. If you want, advanced

control algorithm, such as Kinematic control, LQR

with FF control, Preview control, MPC (Lee and Yi,

2015), and so on, is needed.

As a result of this study, the conclusions could be

summarized that it is experimentally proved that it is

possible to partially implement an autonomous

driving system using a commercial vehicle through

public recognition algorithm and control algorithm

with proposed stop/crossing decision algorithm.

ACKNOWLEDGEMENTS

This work was supported by Institute for Information

& communications Technology Planning and

Evaluation (IITP) grant funded by the Korea

government (MSIT) (No. R7117-16-0164,

Development of wide area driving environment

awareness and cooperative driving technology which

are based on V2X wireless communication).

VEHITS 2019 - 5th International Conference on Vehicle Technology and Intelligent Transport Systems

548

REFERENCES

Road Traffic Authority, Traffic Accident Detailed

Statistics, http://bitly.kr/d54R, Korea, 2018.

G. Reina, D. Johnson, J. Underwood, 2015, Radar Sensing

for Intelligent Vehicles in Urban Environments, MDPI

Sensors Journal, Vol.15, No.6 pp.14661- 14678.

D. J. Lim, 2014, Obstacle tracking and Vehicle Recognition

Based on Lidar and Vision Sensor Fusion, M.S, Ajou

University, Suwon.

Pyeong-Geun Jo, Joon-Woong Lee, 2015, Traffic Light

Detection Using Morphometric Characteristics and

Location Information in Consecutive Images, Journal

of Institute of Control, Robotics and Systems, Vol.21,

No.12, pp.1122-1129.

Taekhoon Jung, Jungha Kim, 2017, Traffic Light

Recognition based on HSV/YCbCr Color Model and

Morphological Feature, KSAE Annual Conference

Proceedings, pp.547-551.

Website-WIKIPEDIA, 2019, CNN, http://bitly.kr/ieEJd.

Sungwoo Heo, Taehyung Park, 2016, Localization of

Autonomous Vehicle Using 3D LIDAR and IMU,

KSAE Annual Conference Proceedings, pp.850-850.

Su Young Kim, Jung-Min Lee, Dong Hwal Lee, Man

Hyung Lee, 2007, Unmanned Navigation of Vehicle

Using the Ultrasonic Satellite System, Journal of

Institute of Control Robotics and Systems, Vol.13, No.9,

pp.875-882.

Hojung Jung, Donghwi Lee, Seungki Kim, Changsup Kim,

Kunsoo Huh, 2012, Development of Vehicle Position

Estimator using Dead Reckoning Algorithm, KSAE

Annual Conference Proceedings, pp.1255-1257.

Jarrod M. Snider, 2009, Automatic Steering Methods for

Autonomous Automobile Path Tracking, Robotics

Institute Carnegie Mellon University Pittsburgh, CMU-

RI-TR-09-08.

Chulhoon Jang, Sungjin Cho, Sangwoo Jeong, Jae Kyu

Suhr, Ho Gi Jung, 2017, Traffic light recognition

exploiting map and localization at every stage, Expert

Systems With Applications, Vol.88, pp.290-304.

Joseph Redmon, Ali Farhadi, 2018, YOLOv3: An

Incremental Improvement, CVPR.

Changhyeon Park, Seok-Cheol Kee, 2018, A Study on the

Traffic Light Recognition Based On Deep Learning for

Urban Autonomous Driving, KSAE Spring Conference

Proceeding, pp.811-812.

Website-Joseph Redmon, 2018, YOLO v3 darknet frame,

https://pjreddie.com.

Justin Salamon, Juan Pablo Bello, 2016, Deep

Convolutional Neural Networks and Data

Augmentation for Environmental Sound Classification,

IEEE Signal Processing Letters, Vol.23, No.3, pp.279-

283.

Website-WIKIPEDIA, 2019, RTK GPS, https://en.wiki

pedia.org/wiki/Real-time_kinematic.

HoonJe Woo, SeungKwon Jung, BongGeun Cho, JungHa

Kim, 2009, Research of navigation algorithm for

Unmanned Gound Vehicle based on Real Time

Kinematic(RTK)-GPS, KSAE Spring Conference

Proceedings, pp.972-976.

Richard B. Langley, 1998, RTK GPS, GPS WORLD

September 1998 INNOVATION, pp.70-76.

Jong-Il Bae, 2013, An Adaptive Cruise Control Systems for

Inteligent Vehicles in Accordance with Vehicles

Distance, The transactions of The Korean Institute of

Electrical Engineers, Vol. 62, No. 8, pp.1157-1162.

Junbum Park, Jungha Kim, 2015, Development of Pure

Pursuit Algorithm using Fuzzy Theory-based Look-

Ahead Distance Function, Kukmin University, Master

degree paper.

Jun-Yung Lee, Kyong-Su Yi, 2015, MPC based Steering

Control using a Probabilistic Prediction of Surrounding

Vehicles for Automated Driving, Journal of Institute of

Control, Robotics and Systems, Vol.21, No.3, pp.199-

209.

Implementation of Autonomous Driving Vehicle at an Intersection with Trafﬁc Light Recognition and Vehicle Controls

549