Real-time Sign Detection and Recognition for Self-driving Mini Rovers

based on Template Matching and Hierarchical Decision Structure

Quang Nhat Nguyen Le, Abir Bhattacharyya, Mamen Thomas Chembakasseril and Ronny Hartanto

Faculty of Technology and Bionics, Rhine-Waal University of Applied Sciences, Kleve, Germany

Keywords:

Real-time Trafﬁc Sign Detection and Recognition, Computer Vision, Autonomous Systems, Mini-rovers.

Abstract:

Sign detection and recognition play vital roles in the ﬁeld of self-driving vehicles. The aim of this paper is to

introduce a real-time methodology that can be implemented on affordable single-board computers to classify

varied trafﬁc signs from the camera feed. The approach is to detect and recognise the colour and shape of

the signs at ﬁrst, then use the acquired information to access a hierarchy structure of the database in order

to extract features of pre-existing templates. Finally, the template matching algorithm is applied to compare

those features with potential Region Of Interest (ROI) based on a threshold value. We installed our system on a

mini rover and tested it on a control urban trafﬁc scenario. The measurements showed processing time ranging

from 230ms to 800ms and 480ms to 1900ms on Jetson Nano and Raspberry Pi 3 Model B+ respectively.

1 INTRODUCTION

Two of the most essential skills for drivers who com-

mute on the roads are trafﬁc signs detection and

recognition. Trafﬁc signs provide information regard-

ing the state of the road (city road, urban road, pri-

ority road, etc.) or indicate the proper behaviours to

drivers and pedestrians (drive slowly, give way, stop,

yield, etc.). While human drivers can perform consid-

erably well those tasks, the accuracy of their decisions

can be affected by many subjective/objective factors

such as physical condition (e.g. tired, drug effect, bad

mood, etc.), environment conditions (snow, rain, ex-

treme illumination, etc.) (Fletcher et al., 2003). In

addition, self-driving vehicles have generated signif-

icant research interest in the last few years as a solu-

tion to resolve trafﬁc congestion, trafﬁc emission as

well as enhance safety and efﬁciency in daily com-

mute and logistic(Dobrev et al., 2017). Nevertheless,

few researchers have addressed the problem of traf-

ﬁc sign detection and recognition in self-driving vehi-

cles and the analysis of adapting those methods in the

dynamic of physical vehicles. Therefore, the require-

ment of a real-time and reliable tool for sign detection

and recognition is critically crucial.

The motivation of this paper is to develop an au-

tonomous self-driving mini rover for the autonomous

vehicle trafﬁc competition – JRC AUTOTRAC 2020:

How the future road transport will look like? (Catta-

neo, 2019). In this competition, there would be an

urban simulation scenario in which the rover must

follow the instruction of trafﬁc signs but still main-

tain the correct paths and avoid crashes with other

rovers. The rover should not exceed the dimension

of 150x200x200mm and weight of 2kg. (Fig. 1).

Several approaches regarding trafﬁc sign detec-

tion and recognition were illustrated in the last decade

such as Support Vector Machine (SVM) (Maldonado-

Bascon et al., 2007) (Greenhalgh and Mirmehdi,

2012), Neural Networks (NNs) (Chiung-Yao Fang

et al., 2003), You Only Look Once (YOLO) (Zhang

et al., 2017), Template matching base on priori knowl-

edge(Piccioli et al., 1996). State of the art meth-

ods including Machine Learning (SVM, NNs, YOLO,

etc.) have demonstrated their outstanding perfor-

mance, however, they either require a bulky stationery

processing unit or might be problematic in real-time

application (Chen et al., 2014). Due to the fact that

our robot - Hammy is a small mobile rover which

travels in scale model of control urban environment,

plus there are a limited number of trafﬁc signs need

to be classiﬁed in the competition, we developed a

system which did not demand excessive computation

and was able to perform adequately on common em-

bedded systems such as Jetson Nano and Raspberry

Pi 3 Model B+.

In this paper, a system for trafﬁc sign detection

and recognition based mainly on template matching

and other techniques of Computer Vision is intro-

duced. Later, experiments of this system on Hammy

208

Le, Q., Bhattacharyya, A., Chembakasseril, M. and Hartanto, R.

Real-time Sign Detection and Recognition for Self-driving Mini Rovers based on Template Matching and Hierarchical Decision Structure.

DOI: 10.5220/0008969702080215

In Proceedings of the 12th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2020) - Volume 1, pages 208-215

ISBN: 978-989-758-395-7; ISSN: 2184-433X

among different control urban environments are dis-

cussed.

2 SYSTEM ARCHITECTURE

One of the priorities of Hammy is the ability to self-

drive and behave associating with trafﬁc signs, lanes

and other objects on the roads. For this reason, the

rover requires an engine unit to deliver torque; a con-

trol unit to process data from sensing system and per-

form computations; and a portable power source.

As for the engine unit, the 2 common selections

are combustion engine and electric motor. The ﬁrst

one can provide greatly high torque but can be con-

siderably heavy and bulky. Furthermore, combustion

engine is also noisy, difﬁcult to control and requires a

large transmission unit to deliver the torque. Electric

motor does not generate as high torque as combustion

engine but has a much smaller weight. It is also eas-

ier to control and more environmentally friendly com-

pared to combustion engine. Because the purpose of

this paper is to develop a reliable solution and algo-

rithm for self-driving scaled-model vehicles, experi-

ments would be conducted in small scale and labora-

tory tracks. Therefore, 2 DC motors equipped with

gearboxes were selected which would drive 2 wheels

separately.

The sensing system consists of an Ultrasonic

Sensor HC-SR04, 2 Reﬂective Infrared IR Optical

Sensors (Optional) and a Camera Raspberry Pi V2.

Hammy would depend mainly on the data from the

camera to adjust its motor speed as well as steer-

ing angle. It should be noticed that humans use

only biological vision without the need of any dis-

tance/infrared/LIDAR (Light Detection and Ranging)

sensors, but still can perform very precisely driv-

ing task. Hence, self-driving algorithm of Hammy

was developed principally focused on computer vi-

sion with the reference of data from external sensors

in an attempt to improve the accuracy of the decision.

Hammy requires a portable but adequate process-

ing unit in order to compute data from the camera and

sensors. In our experiment, Jetson Nano - a small but

powerful computer was selected. Later, a less power-

ful but more affordable single-board computer: Rasp-

berry Pi 3 Model B+ was experimented.

Generally, when Hammy approaches to a poten-

tial trafﬁc sign, it ﬁrstly tries to detect and recognise

the colour and contour of the sign through its camera.

The closer it reaches to sign, the larger of colour area

it detects. If this area exceeds a deﬁned threshold,

the rover will slow down/stop and try to recognize the

sign from the ROI using template matching. Since

Figure 1: CAD design of Hammy.

trafﬁc sign usually follows a size standard, ROI can

be deﬁned utilizing the center of colour area as well

as the sign’s width and length. If the conﬁdence of

classiﬁcation is satisfactory, Hammy will act regard-

ing the indication of the sign. Fig. 2 demonstrates the

software architecture of our system.

3 METHODOLOGY

3.1 Camera Calibration

Distortion and bad representative colour are common

ﬂaws when capturing images from cameras. Those

ﬂaws can generate disturbing effects on the input im-

ages (e.g. straight-line turns to curved line, incorrect

colour level) and deteriorate noticeably the perfor-

mance of Hammy in detecting and recognizing trafﬁc

signs. In an attempt to avoid those unfavourable el-

ements, it is necessary to perform camera and colour

calibration before conducting further steps.

The most common method for camera calibration

is to utilize a planar pattern which contains at least

two distinguishable dimensions (Zhang, 2000). In

this paper, a chess board frame was used as the ref-

erence pattern for Hammy’s camera calibration.

According to (Zhang, 2000), considering (x

, y

)

and (x, y) are respectively processed distortion-free

coordinates and real distorted coordinates, the distor-

tion effect can be solved as below:

= x + x

+ y

) + k

+ y

)

= y + y

+ y

) + k

+ y

)

(1)

in which k

and k

are the radial distortion coefﬁ-

cients. OpenCV library (Bradski, 2000) was adopted

to extracts radial distortion coefﬁcients from sample

images.

Real-time Sign Detection and Recognition for Self-driving Mini Rovers based on Template Matching and Hierarchical Decision Structure

209

Figure 2: Schematic of the Hammy software architecture.

In several circumstances, the illumination of ex-

ternal environment and affect unfavourably to the

recognition ability of Hammy. For instance, high in-

tensity of light can cause the rover mistaking between

”red” and ”pink”. Hammy can avoid that by conduct-

ing colour calibration based on HSV(Hue, Saturation,

Value) colour space to classify different colours. It

was proved that HSV is able to classify colours bet-

ter than both CIELAB and RGB (Reg, Green, Blue)

colour space (Paschos, 2001). Basic colours includ-

ing Red, Yellow, Green, Blue can be converted from

RGB to HSV by below equations(Saravanan, 2010):

With value of RGB in range [0,1]:

H :=











0, if MAX = MIN

⇔ R = G = B

60 ·



0 +

G−B

MAX−MIN



, if MAX = R

60 ·



2 +

B−R

MAX−MIN



, if MAX = G

60 ·



4 +

R−G

MAX−MIN



, if MAX = B

(2)

if H < 0 then H := H + 360

HSV

(

0, if MAX = 0 ⇔ R = G = B = 0

MAX−MIN

MAX

, otherwise

(3)

V := MAX (4)

L :=

MAX +MIN

(5)

3.2 Sign Detection

The process of sign detection contains two steps:

colour contour detection and shape detection. In order

to extract the colour contour of the sign, a state of the

art contour detector technique integrating both local

and global cues was utilized (Maire et al., 2008). This

technique was claimed to deliver best performance

on the Berkeley Segmentation Dataset benchmark in

2008. Hammy would reduce the level of noise by

Gaussian Blur(Nixon and Aguado, 2012) before ac-

tivating contour detector on HSV colour space of the

frame. A pre-deﬁned range was adapted to prevent

the system from wasting its resources on detecting un-

wanted colours. The center of colour contour must be

speciﬁed so that our system can perceive the ROI and

concentrate on it. This could be done by calculating

the central moment (m

i j

) and extrapolating the central

coordinates from it (Yang and Albregtsen, 1996):

∑

x,y

(array(x, y) · (x − x

)

· (y − y

)

(6)

where (x

, y

) is the mass center:

(7)

To classify the shape of a trafﬁc sign, the captured

frame was converted to grayscale(Saravanan, 2010)

in an attempt to optimize resource utilization. Canny

Edge detection (Canny, 1986) was implemented to

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

210

highlight the captured sign contour. Subsequently, the

Douglas - Peucker algorithm was integrated to count

the numbers of vertices from the trafﬁc sign’s con-

tour extraction. This algorithm would convert a curve

created by multiple line segments into a similar curve

with fewer vertices (Prasad et al., 2012). Therefore

the shape of a trafﬁc sign can be determined by count-

ing the number of total vertices. With numbers of ver-

tices equal to three, four and ﬁve the shape will be

triangle, square/rectangle, and pentagon respectively.

In case the numbers of vertices are considerably large,

the shape would be assigned as circle (a circle is com-

posed of vast numbers of line vertices).

3.3 Sign Recognition

In an attempt to interpret the sign, Hammy would

access the pre-existing database associating to infor-

mation extracted from the colour and shape of the

trafﬁc sign and execute template matching algorithm

(Pereira and Pun, 2000). The database should contain

images of trafﬁc sign in which users want Hammy to

recognize. The template T slides over the original Im-

age I and gives a resultant matrix R which contains

the pixel locations in terms of (x, y) where the pixels

matched. If there are no pixels detected this matrix

R would be null. In terms of visualisation, the ma-

trix R is basically a black image except for the region

where it matched, which would then be represented as

a bright spot. If the same template may appear mul-

tiple times the image would then have multiple bright

spots on the same dark image.

OpenCV library (Bradski, 2000) provides several

equations to calculate the region where features of a

template are matched. These equations perform the

appropriate function between the template T and the

original image I, which returns either the minimum or

the maximum values based on the equation chosen.

In this experiment the normalised coefﬁcient correla-

tion was used given by the equation below (Goshtasby

et al., 1984):

R(x, y) =

∑

, y

) · I

(x + x

, y + y

))

∑

, y

)

∑

(x + x

, y + y

)

(8)

The resultant matrix in this case is of the shape of

the original image with a value between 0 and 1 in

each pixel position. A threshold is then used to de-

termine the pattern location of the matching position.

The higher of the threshold, the higher accuracy re-

quired for the template to match the captured images.

4 EXPERIMENTS AND RESULT

The frame of Hammy was designed with CAD soft-

ware and then fabricated with 3D printer. A recharge-

able 5V-2A power bank with storage of 20,000 mAh

is mounted inside the frame. During experiments with

all processing unit, motors and sensors operating con-

tinuously, Hammy could last up to 5-6 hours. The op-

erating duration can be extended by replacing Jetson

Nano with Raspberry Pi 3, yet the difference was in-

signiﬁcant. Hammy weighs approximately 1kg and

has the dimension of 140x190x120mm. Fig. 3 illus-

trates the physical prototype of Hammy which was

conducted in our research.

Figure 3: Prototype of Hammy.

To evaluate the performance of our algorithm, an

urban scenario consisting of Hammy and a trafﬁc sign

stationed along the road was constructed (Fig. 4). The

system was installed on both Jetson Nano and Rasp-

berry Pi 3 Model B+ and tested in different lighting

conditions: lack of illumination (dark - 2 Lux), natu-

ral illumination with daylight (ambient - 64 Lux) and

indoor illumination with artiﬁcial light (lighted - 197

Lux). An experiment is considered successful when

the rover stops before the sign and behaves accord-

ingly to the indication demonstrated by trafﬁc sign.

From recorded data, the temporal performance, as

well as the responsiveness of our system, were evalu-

ated.

4.1 Accuracy Performance

As mentioned in the methodology section, our sys-

tem depends heavily on the colour contour extraction

and shape recognition. One of those processes that

does not function properly will lead to the break of

the whole system. We tested our system on a set of

200 image samples similar to Fig. 5 and recorded

Real-time Sign Detection and Recognition for Self-driving Mini Rovers based on Template Matching and Hierarchical Decision Structure

211

Figure 4: Testing the temporal performance and system re-

sponsiveness on Hammy.

the certainty of our system on Table 1. The result

illustrates that while the accuracy of sign detection

was relatively high (around 90%), this value on sign

recognition was noticeably lower as less than 81%,

especially in dark condition. Fortunately, our sys-

tem architecture does not depend on a single data to

deliver the decision. Conventional cameras usually

record 24 frames per second, which means in one sec-

ond Hammy can access and predict a collection of

24 images. Obviously, it is not efﬁcient to process

all images in one frame throughout, thus a portion of

each image collection in one second is gathered for

the process. The control unit of Hammy could also be

intervened to orient the visual angle of camera, there-

fore new data can be collected in case previous data is

challenging for decision making. Fig. 5a and Fig. 5b

demonstrates the result after applying sign detection

and sign recognition on an image.

Table 1: Accuracy performance of sign detection and sign

recognition on 200 samples.

Dark Ambient Lighted

Sign detection 87.5% 90.5% 92.5%

Sign recognition 75% 80.5% 80.5%

4.2 Temporal Performance

As can be seen in Fig. 2, Hammy perceives the ﬁ-

nal result of trafﬁc sign through consecutive steps:

Colour recognition, Sign recognition and Threshold

ﬁlter. To measure the temporal performance of our

algorithm, a python module called timer was imple-

mented with a starting point and a end point set before

the ﬁrst step and after the last step respectively. The

results are summarised in Fig. 6.

It is evident that the algorithm achieved the best

performance in lighting condition when the process-

(a)

(b)

Figure 5: a) Shape extraction and classiﬁcation of trafﬁc

signs. b) Template matching following the database’s hier-

archy tree.

ing time ﬂuctuated evenly around 238.4ms on Jetson

Nano and 599.5ms on Raspberry Pi 3 Model B+. The

standard deviation was also the smallest among three

conditions, which indicates the system works most

stable in lighting condition. Meanwhile in dark and

ambient condition, it showed poorer performances as

approximately double on Jetson Nano and triple on

Raspberry Pi 3 . The results indicate that our method

depends heavily on the lighting condition since colour

region and sign edges can contrast better to the back-

ground. In addition, Jetson Nano delivers better per-

formance in comparison to Raspberry Pi 3 thanks to

its integrated 128-core Maxwell GPU, yet this differ-

ence is not signiﬁcant. With processing time less than

2s for all three conditions, the system performance on

both of them satisﬁed our requirement in the ability

of real-time operation. Nevertheless, it should be no-

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

212

(a)

(b)

Figure 6: a) Temporal performance in three different illu-

minating conditions. b) Comparison of the temporal per-

formance between Jetson Nano and Raspberry Pi 3 Model

B+.

ticed that the system normally breaks if the sign is

affected by external factors (hidden by obstacles, dis-

colouration, deformation, etc) since its features can

not be exposed to ﬁt with the template extracted from

the database. Such problems can be solved by con-

ducting Machine Learning methodologies including

SVM (Greenhalgh and Mirmehdi, 2012) and Neu-

ral Networks (Zhang et al., 2017) (Zhu et al., 2016)

(Eykholt et al., 2017). Those methods was proved to

deliver reliably and precisely the prediction without

the need of collecting all the features from the tem-

plate. Regardless, Machine Learning methods usually

are resource-intensive and require graphic processing

unit (GPU) to increase the speed of their computation.

4.3 System Responsiveness

One of the most important characteristics of self-

driving vehicles is the ability to rapidly respond to ex-

ternal factor in real-time. In this experiment, Hammy

ran on a straight path with 65% maximum velocity

and a turning sign was installed on its route (Fig.

4). As mentioned above, when Hammy approached

closer to the trafﬁc sign it would try to extract the

colour region and the shape of potential trafﬁc signs

from the captured frame. When the conﬁdence of

the colour and shape recognition reached a deﬁned

threshold (calibrated 25cm in front of the trafﬁc sign),

Hammy would stop, access the database and extract

the features of pre-existing templates based on the hi-

erarchy tree. Subsequently, the rover made its deci-

sion by comparing the extricated features with scep-

tical ROI and selecting the best one. Similar to the

temporal performance experiment, we tested Hammy

in three different illuminating condition and plotted

the data as can be seen in Fig. 7.

It can be observed that the stopping distance of

Hammy ﬂuctuates unevenly in the range of 23cm

to 29cm in three lighting conditions. The tolerance

compared to the desired stopping distance (25cm)

is around ±4cm. The average of responsive dis-

tance in dark condition is approximately 26.13cm

while this number in ambient and lighting conditions

are 25.73cm and 25.67cm. The experiment results

show that the lack of illumination might deteriorate

Hammy’s ability to detect and recognise trafﬁc signs.

However, this tolerance is acceptable since the devi-

ation (±4cm) is relatively small and can be compen-

sated by Hammy’s controller.

5 CONCLUSIONS

In the last decade, many researches in the ﬁeld of

trafﬁc sign detection and recognition have been de-

veloped, still few of them have indicated an effective

real-time solution for self-driving rovers which can be

implemented on small, affordable single-board com-

puters. Therefore, we presented a system for trafﬁc

sign detection and recognition which is cost-effective

and compatible with two of the most popular embed-

ded boards: Jetson Nano and Raspberry Pi. The idea

is to formerly detect and classify colour and shape of

Real-time Sign Detection and Recognition for Self-driving Mini Rovers based on Template Matching and Hierarchical Decision Structure

213

(a)

(b)

Figure 7: a) Spatial performance result among three illumi-

nating conditions. b) Average of spatial performance.

the trafﬁc sign, then use this information as correlated

conditions to go deeper into the database branches

where sign templates are stored. In the next step,

features of all templates are extracted and compared

with ROI from video stream and those which have the

highest similarity will be selected. The advantage of

this methodology is that it consumes a limited amount

of computational power due to the hierarchical struc-

ture of the algorithm. Yet, this system depends heav-

ily on the subtle changes in the external environment

and the sufﬁciency of the database. For example, the

colour can be misclassiﬁed if the real sign discolours

over time, or the system will not able to recognise

shape contour if the sign is hidden by obstacles. The

system shows potential in areas which do not con-

sist of too many appearance-affecting factors such as

hospitals, factories, schools, laboratory, indoor areas,

etc. It can also be implemented as an assisting module

running parallel with other modules on an embedded

system to perform multi-tasks, especially when this

system has limited computational resources.

Future work should focus on algorithms that can

reduce the affection of external environment factors to

the sign detection and recognition process. Additional

sensors can be equipped to improve the accuracy of

decision such as colour sensor or LIDAR (Light De-

tection and Ranging) (Zhou and Deng, 2014).

ACKNOWLEDGEMENTS

This research was supported by Rhine-Waal Univer-

sity of Applied Sciences. We would like to thank

Mr. Frank Kremer and Mr. Freidrich Muhs for as-

sistance with the electronic conﬁguration of Hammy.

We would also like to show our gratitude to HSRW

Fablab Kamp-Linfort, HSRW Robotic Lab for pro-

viding us equipment and assist to develop our re-

search.

REFERENCES

Bradski, G. (2000). The OpenCV Library. Dr. Dobb’s Jour-

nal of Software Tools.

Canny, J. (1986). A computational approach to edge de-

tection. IEEE Transactions on pattern analysis and

machine intelligence, (6):679–698.

Cattaneo, B. (2019). JRC AUTOTRAC 2020 How the fu-

ture road transport will look like ?

Chen, Y., Lin, Z., Zhao, X., Wang, G., and Gu, Y. (2014).

Deep learning-based classiﬁcation of hyperspectral

data. IEEE Journal of Selected Topics in Applied

Earth Observations and Remote Sensing, 7(6):2094–

2107.

Chiung-Yao Fang, Sei-Wang Chen, and Chiou-Shann Fuh

(2003). Road-sign detection and tracking. IEEE

Transactions on Vehicular Technology, 52(5):1329–

1341.

Dobrev, Y., Vossiek, M., Christmann, M., Bilous, I., and

Gulden, P. (2017). Steady delivery: Wireless local po-

sitioning systems for tracking and autonomous navi-

gation of transport vehicles and mobile robots. IEEE

Microwave Magazine, 18(6):26–37.

Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A.,

Xiao, C., Prakash, A., Kohno, T., and Song, D. (2017).

Robust physical-world attacks on deep learning mod-

els. arXiv preprint arXiv:1707.08945.

Fletcher, L., Apostoloff, N., Petersson, L., and Zelinsky,

A. (2003). Vision in and out of Vehicles. Intelligent

Systems, IEEE, 18:12–17.

Goshtasby, A., Gage, S. H., and Bartholic, J. F. (1984).

A two-stage cross correlation approach to template

matching. IEEE Transactions on Pattern Analysis and

Machine Intelligence, (3):374–378.

Greenhalgh, J. and Mirmehdi, M. (2012). Real-Time De-

tection and Recognition of Road Trafﬁc Signs. IEEE

Transactions on Intelligent Transportation Systems,

13(4):1498–1506.

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

214

Maire, M., Arbel

aez, P., Fowlkes, C., and Malik, J. (2008).

Using contours to detect and localize junctions in nat-

ural images. In 2008 IEEE Conference on Computer

Vision and Pattern Recognition, pages 1–8. IEEE.

Maldonado-Bascon, S., Lafuente-Arroyo, S., Gil-Jimenez,

P., Gomez-Moreno, H., and Lopez-Ferreras, F. (2007).

Road-Sign Detection and Recognition Based on Sup-

port Vector Machines. IEEE Transactions on Intelli-

gent Transportation Systems, 8(2):264–278.

Nixon, M. and Aguado, A. S. (2012). Feature Extraction

and Image Processing for Computer Vision. Aca-

demic Press.

Paschos, G. (2001). Perceptually uniform color spaces for

color texture analysis: an empirical evaluation. IEEE

Transactions on Image Processing, 10(6):932–937.

Pereira, S. and Pun, T. (2000). Robust template matching

for afﬁne resistant image watermarks. IEEE transac-

tions on image Processing, 9(6):1123–1129.

Piccioli, G., De Micheli, E., Parodi, P., and Campani, M.

(1996). Robust method for road sign detection and

recognition. Image and Vision Computing, 14(3):209–

223.

Prasad, D. K., Leung, M. K., Quek, C., and Cho, S.-Y.

(2012). A novel framework for making dominant

point detection methods non-parametric. Image and

Vision Computing, 30(11):843 – 859.

Saravanan, C. (2010). Color Image to Grayscale Image

Conversion. In 2010 Second International Confer-

ence on Computer Engineering and Applications, vol-

ume 2, pages 196–199.

Yang, L. and Albregtsen, F. (1996). Fast and exact compu-

tation of cartesian geometric moments using discrete

green’s theorem. Pattern Recognition, 29(7):1061–

1073.

Zhang, J., Huang, M., Jin, X., and Li, X. (2017). A Real-

Time Chinese Trafﬁc Sign Detection Algorithm Based

on Modiﬁed YOLOv2. Algorithms, 10(4):127.

Zhang, Z. (2000). A Flexible New Technique for Camera

Calibration. IEEE Transactions on Pattern Analysis

and Machine Intelligence, 22.

Zhou, L. and Deng, Z. (2014). Lidar and vision-based

real-time trafﬁc sign detection and recognition algo-

rithm for intelligent vehicle. In 17th International

IEEE Conference on Intelligent Transportation Sys-

tems (ITSC), pages 578–583.

Zhu, Z., Liang, D., Zhang, S., Huang, X., Li, B., and Hu,

S. (2016). Trafﬁc-sign detection and classiﬁcation

in the wild. In Proceedings of the IEEE Conference

on Computer Vision and Pattern Recognition, pages

2110–2118.

Real-time Sign Detection and Recognition for Self-driving Mini Rovers based on Template Matching and Hierarchical Decision Structure

215