IEEE INTERNET OF THINGS JOURNAL, VOL. 5, NO.

2, APRIL 2018 747

IoT-Driven Automated Object Detection Algorithm

for Urban Surveillance Systems in Smart Cities
Ling Hu and Qiang Ni, Senior Member, IEEE

Abstract—Automated object detection algorithm is an impor- i.e., the extraction of a license plate region from an image, is
tant research challenge in intelligent urban surveillance systems the key module in a VLPR system [3], which influences the
for Internet of Things (IoT) and smart cities applications. In accuracy of the VLPR systems significantly. Different algo-
particular, smart vehicle license plate recognition and vehi-
cle detection are recognized as core research issues of these rithms have been proposed for identifying a vehicle license
IoT-driven intelligent urban surveillance systems. They are key plate using image processing [4]. One typical way is vertical
techniques in most of the traffic related IoT applications, such as edge matching [5]. The idea is to first locate the two ver-
road traffic real-time monitoring, security control of restricted tical edges of a license plate, and hence to detect its four
areas, automatic parking access control, searching stolen vehicles, corners. In this way, the license plate can be extracted accu-
etc. In this paper, we propose a novel unified method of auto-
mated object detection for urban surveillance systems. We use rately. Using the contrast between the grayscale values, [6]
this novel method to determine and pick out the highest energy proposed a vertical edge-based license plate recognition
frequency areas of the images from the digital camera imaging method.
sensors, that is, either to pick the vehicle license plates or the Another technology is morphology-based license plate
vehicles out from the images. Our proposed method can not only detection. This method is to extract important features of con-
help to detect object vehicles rapidly and accurately, but also can
be used to reduce big data volume needed to be stored in urban trast as guidance to search the license plates [7]. In [8], to
surveillance systems. extract potential text information from the image, a method is
proposed using adaptive threshold, fractal filter and morpho-
Index Terms—Internet of Things (IoT), smart city, smart
transportation, urban surveillance system, vehicle, vehicle license logical analysis. In [9] and [10], edge statistics in combination
plate. with morphological approaches are proposed to eliminate the
undesired edges in the images.
Color-based methods are also attempted which make use
I. I NTRODUCTION
of the colors of the vehicle license plate. In [11], a color-
MART transportation and urban surveillance systems
S are important Internet of Things (IoT) applications for
smart cities [1], [2]. In these smart transportation and urban
based method combined with the texture characteristics is
proposed to try to detect license plate from the color image.
In [12] and [13], the combination of edge information and
surveillance applications, cameras/imaging sensors are com- plate color are utilized to identify the vehicle license plates.
monly installed to automatically detect and identify potential Based on neural network techniques, other recognition
vehicles/cars through automated object detection methods. methods of vehicle license plates are proposed. These meth-
Usually, such automated object detection methods demand ods are designed to train classifiers to offer a proper response
high-complexity image/data processing technologies and algo- to the license plate images. Sun et al. [14] applied genetic
rithms. Hence, the design of low-complexity automated object algorithm to the training process and combine the statistic fea-
detection algorithms becomes an important topic in urban tures together with structure features. In [15], a vehicle license
surveillance systems. Among these researches, both vehi- plate detection method using neural network approaches is
cle license plate recognition (VLPR) and vehicle recognition proposed. The proposed scheme utilizes a neural network
are hot research topics worldwide, which can be applied chip named as CogniMem to detect the vehicle license plates.
to many IoT applications, such as road traffic data col- Sharma et al. [16] proposed a method using wavelet transform
lection/monitoring, automatic parking charging and access technique to decompose the images into different layers, and
control, and searching stolen vehicles. then utilize the low frequency images to combine with neural
It is known that a license plate number is a unique identifi- network technique.
cation of a vehicle. Specifically, the license plate recognition, While most approaches have attempted to deal with the
Manuscript received April 7, 2017; accepted May 4, 2017. Date of publi- VLPR with optical characteristics recognition, another method
cation May 18, 2017; date of current version April 10, 2018. This work was is proposed based on the horizontal and vertical frequency
supported by the Institute of Software Application Technology, Guangzhou energy differences. Rajput et al. [17] used Daubechies wavelet
and Chinese Academy of Sciences and Dongguan CAS Smart City Software
Co. Ltd. under Smart City Urban Surveillance Systems project under Grant transform method to calculate the vertical and horizontal
CSA7035. (Corresponding author: Qiang Ni.) frequency energy curves, considering the frequency is the
The authors are with the School of Computing and Communications, highest near the number plate.
Lancaster University, Lancaster LA1 4WA, U.K. (e-mail:
l.hu@lancaster.ac.uk; q.ni.@lancaster.ac.uk). Other researchers focus on the vehicle recognition.
Digital Object Identifier 10.1109/JIOT.2017.2705560 Mendoza-Schrock et al. [18] proposed a combination of the
2327-4662 c 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
748 IEEE INTERNET OF THINGS JOURNAL, VOL. 5, NO. 2, APRIL 2018

Fig. 1. Original car image. Fig. 2. Gray scale image.

transfer subspace learning technique with the manifold learn-

ing approach to allow a more systematic search of tuning
parameters for the purposes of cross-data and cross-domain
electro-optical vehicle recognition. Pei et al. [19] proposed
a vehicle recognition and retrieval system using the so-called
bag of words approach. In this way, the system may deter-
mine automatically the manufacturer and the type of the
captured vehicle images. Burresi and Giorgi [20] presented
a vehicle recognition system prototype using magnetic sensor
techniques, where more complex hardware and processing are
utilized.
However, all the above mentioned existing methods in the
Fig. 3. Gradient of the gray image.
literature are still too complex which demand high computa-
tional processing. Furthermore, each method can only be used
to recognize a vehicle license plate or a vehicle individually. The formula to convert the image is illustrated as follows:
None of them can be used as a unified method to detect both 3Ar + 6Ag + Ab
vehicle license places and vehicles. Agl = (1)
10
In this paper, we propose a novel, simple, and unified
method to search the objects by filtering out the vehicle and/or where Ar denotes the red spectrum of the color image, Ag is
license plate images rapidly from the digital camera imaging the green spectrum of the color image, and Ab is the blue
sensors. We design a simple filter to effectively detect either spectrum of the color image. The result Agl is the converted
vehicle license plates or vehicles, motivated by the observa- gray scale level. The gray scale image is shown in Fig. 2.
tion that in the most cases the object is the highest energy In the second step, we calculate and obtain the gradient of
frequency part of an image. This process can be easily imple- the gray image. We consider the 2-D gray scale image as a 2-D
mented in any urban surveillance systems in smart cities to matrix P, and calculate the numerical gradient of this matrix.
pick out comparably important areas from the images cap- Px corresponds to the differences in x (horizontal) direction.
tured by any camera/imaging sensors in urban environments, Py corresponds to the differences in y (vertical) direction. The
which will be a useful method not only to rapidly detect the definition to calculate the Px and Py are

important information but also to reduce the large data volume Px(i,j) = I(i + 1, j) − I(i, j)
(2)
required to be stored because only those selected impor- Py(i,j) = I(i, j + 1) − I(i, j)
tant (smaller) data will be stored as compared to huge raw where I is the gray scale value of the gray scale image, and
data captured/generated from any cameras/imaging sensors (i, j) is the position of the pixel.
24 h per day, 7 days a week, and 365 days per year. For every pixel, after getting the gradient of both the hori-
zontal and vertical directions, we calculate the overall gradient
II. O UR P ROPOSED A LGORITHM of this pixel
A. Pretreatment of the Image 
P = Px 2 + Py 2 . (3)
As an instance, let us look at Fig. 1. From this image, we
need to pick out the car plate, which is the highest energy The gradient result is shown in Fig. 3. In Fig. 3, some large
frequency area. Before formal treatment of the image, we carry areas of constant gray scale fade into darkness because of the
a pretreatment of the image. First, we convert the original slow movement of gradient. Now, the car image has finished
RGB (R: red, G: green, B: blue) image into gray scale image pretreatment, the image has been well prepared to implement
since the color information is not needed in our method, and our method. The next step is how to pick out the car license
the size of the image file can be reduced to a lower level. plate area automatically.
HU AND NI: IoT-DRIVEN AUTOMATED OBJECT DETECTION ALGORITHM FOR URBAN SURVEILLANCE SYSTEMS IN SMART CITIES 749

Fig. 4. Line array filters to filter the pretreated image.

B. Design of Novel Filter

Our idea is to design a 2-D smart filter to pick out the object
area of an image. We first investigate the characteristics of Fig. 5. Horizontal filtered image.
the object. We found that, within the object boundaries, the
energy frequency is high, and the energy frequency curves
down sharply outside the object (e.g., the car license plate)
boundaries. This interesting characteristic inspires us to design
a new 2-D filter to figure out the horizontal and perpendicular
frequency energy curves, since the meaning of filtering is to
calculate the intercorrelation of the filter and the image matrix.
Mathematically, the definition of intercorrelation is a mutual
relationship or connection between two or more things. The
formula for the correlation coefficient is
n
(xi − x)(yi − y)
r =  1  . (4)
n 2 n
(x
1 i − x) (y
1 i − y) 2

Our objective is to implement a new filter to calculate the

correlation coefficient for every pixel of the car image and
Fig. 6. Horizontal axis’s brightness values.
obtain a new image with the car license plate area standing out.
To achieve this, we investigate the function for intercorrelation
which is expressed as follows: In order to get the intuitive impression, we plot the 1-D bright-
 ∞ ∞
◦ ∗ ness values for straightforward view by choosing the middle
f (x) g(x) = f ∗ g = f ∗ (u, v) g(x − u, y − v)dudv.
−∞ −∞ horizontal array line of the filtered image. The values obtained
(5) in the x-axis are shown in Fig. 6.
From Figs. 5 and 6, we find that the brightness values show
This motivates us to utilize the pulse function δ(x) which off the plate area in the horizontal direction roughly, with the
holds an important characteristic as follows: higher value representing high energy frequency. The short-
 ∞
coming is the noise, which makes how to detect the threshold
δ(x − a)ϕ(x)dx = ϕ(a). (6)
−∞
difficult. In order to solve the problem of the noise, we need
to improve our filter to smooth the noise while keeping this
That is, the function δ(x−a) can be used to pick out another
filter simple and effective.
function ϕ(x) at the x-axis when x = a. We use this property
to pick out the highest energy frequency area of a car image.
Specifically, we define line arrays to simulate the pulse func- C. Improvement of the Filter
tion δ(x). After the filtering process, the result will be shown By researching the filtered image, we find that the noise
in a new image. comes from the thin lines of the filter. In order to merge the
As for our proposal, a perpendicular line array filter (in sharp edges between the lines, one way is to expand the width
the array, all the numbers are set as 1’s) is used to filter the of the filter. On the other side, the filter width cannot be too
pretreated image in horizontal direction, and another horizontal wide, since it will lead to blurred filtered image. In the hor-
line array filter (in the array, all the numbers are set as 1’s) is izontal direction, we expand the original line array filter into
used to filter the pretreated image in perpendicular direction. 30 lines of 1’s, that is, to expand the filtered line into a matrix.
The diagrammatic sketch is shown in Fig. 4. In this way, the filter becomes a long and narrow rectangle.
We first filter the pretreated image in the horizontal direc- This perpendicular matrix (all the numbers are set as 1’s) is
tion, since the vehicle plate (or a vehicle) normally holds larger used to filter the image in horizontal way. The diagrammatic
size in this direction. Hence, it is easier to eliminate noise in sketch of this filter is shown in Fig. 7.
this direction. The results are shown in Fig. 5. In this way, the We use this improved filter to filter the pretreated image in
high energy frequency areas are shown off by brighter lines. horizontal direction; the results are shown in Fig. 8. In this
750 IEEE INTERNET OF THINGS JOURNAL, VOL. 5, NO. 2, APRIL 2018

Fig. 7. Matrix filter in horizontal direction.

Fig. 10. Object area in horizontal direction.

Fig. 8. Horizontal filtered image using improved filter. Fig. 11. Matrix filter in perpendicular direction.

plate is within an already-known area, we can cut out the car

license plate in x-axis correctly.
We know that the width of the filter matrix brings some blur
and expands the width of the car plate, and the filter scans the
image from left to right. Hence, we only need to gently adjust
the final results by shrinking the detected area and move the
results a little left.
After obtaining the x-axis value, we cut this horizontal
area from the original image, and the result is shown in Fig. 10.

E. Use Filter to Detect the Perpendicular Direction Area

Fig. 9. Horizontal axis’s brightness values using improved filter.

Similar to the horizontal direction, we use matrix filter in the
perpendicular position. Since, the car license plate area in hor-
izontal direction has already been picked out, in order to avoid
direction, the high frequencies areas again are figured out by the interference of unwanted noise, we only need to filter the
brighter areas. pretreated image in the area shown in Fig. 11. We notice that
In order to get the intuitive impression, we plot the 1-D in the perpendicular direction, there is no break between the
figure. By choosing the middle horizontal brightness values of characters like the horizontal direction; hence the height of the
the filtered image, we plot the results in Fig. 9. The results are filter does not need to be big. Here, we choose the height num-
smoothed and noise is reduced while the main shape remains. ber as 8 to filter the object image. The diagrammatic sketch
of this filter is shown in Fig. 11.
After filtering the pretreated image within the area shown
D. Object Area in the Horizontal Direction in Fig. 10, the perpendicular results are shown in Fig. 12. In
Since, the car image has been filtered perfectly, to cut out this direction, the high frequencies areas are shown out by
the plate area is not difficult now. As for the horizontal axis, brighter areas.
the whole plate area is the brightest area. According to this In order to get the intuitive impression, we plot the 1-D fig-
feature, we only need to scan the horizontal axis values from ure for straightforward view as well. By choosing the middle
left to right and simply choose the threshold brightness of 0.5. perpendicular brightness values of the filtered image, we plot
In order to eliminate the interruption of unwanted small bright- the results in Fig. 13. Because the width of the filter is small,
ness, such as the edges of the car, we check the picked width the results are smoothed and noise is reduced while the main
between the thresholds. Since, the width of the car license shape remains.
HU AND NI: IoT-DRIVEN AUTOMATED OBJECT DETECTION ALGORITHM FOR URBAN SURVEILLANCE SYSTEMS IN SMART CITIES 751

Fig. 12. Perpendicular filtered image.

Fig. 15. Process of our proposed method.

Fig. 13. Perpendicular axis’s brightness values.

Fig. 16. Detected car license plate-2.

The result shows the correctly picked car license plate.

Consider that we only use a simple filter to get the result,
Fig. 14. Detected car license plate-1. this demonstrates that our method is effective and practical.

F. Flow Chart of Our Proposed Method

Now, we need to cut out the plate area in the perpendic- The flow chart of our proposed method is shown in Fig. 15.
ular direction. As for the perpendicular axis, the whole plate It is a simple method to pick out the highest energy frequency
area is the brightest area although there are some interruptions area of a vehicle image.
of noise coming from the edges of the car, with the noise
brightness normally shown as narrow pulse. According to this
III. E XPERIMENTAL R ESULTS OF
feature, we only need to scan the perpendicular axis’s bright-
O UR P ROPOSED F ILTERS
ness values from left to right and simply choose the threshold
value at 0.5. In the scan process, we check the picked width A. Use Our Filter to Detect Vehicle License Plates
between the thresholds. Since, the height of the car license In this section, we test our filter to see how it works on
plate is within an already-known area, we can cut out the car more car image with license plates. Our objects are the vehicle
license plate in perpendicular direction correctly. license plates. Some results are shown in Figs. 16 and 17 with
We then gently adjust the final results by shrinking the the same parameter settings.
detected area and move the results a little up. After getting In many situations, the car license plates are not such big
the four edge value, we cut this area from the original car as shown in Figs. 16 and 17. They may appear like in Fig. 18,
image, and the result is shown in Fig. 14. a whole car with a smaller car license plate.
752 IEEE INTERNET OF THINGS JOURNAL, VOL. 5, NO. 2, APRIL 2018

Fig. 21. Detected car license plate-6.

Fig. 17. Detected car license plate-3.

Fig. 22. Original image of a car.

Fig. 18. Whole car with the car license plate.

It is a simple method to pick out the highest energy

frequency area of a vehicle image, in our cases which are the
vehicle license plates. For other similar sized vehicle images,
the parameter settings may be chosen as the same, hence it is
useful for urban surveillance systems. For example, at a car
parking lot, the camera takes the vehicle images at a fixed
place and fixed distance with the vehicles.

B. Use Our Filter to Detect Vehicles

Fig. 19. Detected car license plate-4.
In the following, we will show that our unified filter can be
used to detect the vehicles from the images, since the vehicles
also show high energy frequency within the images.
This time we begin from a car image shown in Fig. 22.
Our aim is to pick out the car from the image. The first step
is to convert the color image into a gray image. Then, similar
to the car license plate searching, we calculate and obtain the
gradient of the gray image.
After we complete the pretreatment process of the car
image, the next step is to pick out the car using our proposed
method.
We first filter the pretreated gradient image in horizon-
tal direction. In this direction, the high frequencies areas are
Fig. 20. Detected car license plate-5. shown out by brighter area. In order to get the intuitive impres-
sion, we plot the 1-D brightness values for straightforward
view by choosing the middle horizontal array line of the
In these cases, we test our proposed filter, and find that filtered image. The array values are shown in Fig. 23.
our method can still detect the car license plates by simply Since, the prepared work has been done perfectly, to cut out
adjusting the threshold of the x-axis and y-axis. We adjust the car area is not difficult now. As for the horizontal axis, the
the threshold of x-axis to 0.7 and keep the threshold of car area is the brightest area. According to this feature, we
y-axis unchanged (0.5). Then our method matches that kind of only need to scan the horizontal axis array from left to right.
median size car license plates. The results of our experiment We simply choose the threshold at 0.5 for the horizontal axis.
are shown in Figs. 19–21. In order to eliminate the interruption of unwanted brightness,
HU AND NI: IoT-DRIVEN AUTOMATED OBJECT DETECTION ALGORITHM FOR URBAN SURVEILLANCE SYSTEMS IN SMART CITIES 753

Fig. 26. Detected car-1.

Fig. 23. Horizontal axis’s brightness values of a car.

Fig. 27. Detected car-2.

Fig. 24. Detected car area in horizontal direction.

Fig. 28. Detected car-3.

Now, we are to cut out the car area in perpendicular direc-

tion. As for the perpendicular axis, the whole car area is the
brightest area although there are some interruptions of noise.
According to this feature, we only need to scan the perpendicu-
lar axis’s brightness values from left to right and simply choose
the threshold as 0.5 for the perpendicular brightness values.
In the scan process, we check the picked height between the
thresholds. Since, the height of the car is within an already-
known scale, we can cut out the car in perpendicular direction
Fig. 25. Perpendicular axis’s brightness values of a car. correctly.
After getting the four edge value, we cut out this area from
the original car image, and the result is shown in Fig. 26.
we check the picked width between the thresholds. Since, the The result is amazing. Note that we use the same simple
width of the car is within an already-known area, we can cut and effective filter to get the result.
out the car in horizontal direction correctly; the result is shown Then, we test our filter to see how it works on more car
in Fig. 24. images. Similar results are shown in Figs. 27 and 28 with the
Similar to the horizontal direction, we use matrix array filter same software code. We find that our simple filter can filter
in the perpendicular position. Since, the car area in horizontal out the cars from their images correctly. This filter is not only
direction has already been picked out, we only need to filter suitable for searching the vehicle license plates from the vehi-
the pretreated image in the area shown in Fig. 24. cle images, but also suitable for detecting the vehicles from
By choosing the middle perpendicular brightness values of different environments, which is very useful in the IoT appli-
the filtered image, we plot the results in Fig. 25. cations of smart cities, since this filter can flexibly find out
754 IEEE INTERNET OF THINGS JOURNAL, VOL. 5, NO. 2, APRIL 2018

the vehicle information based on different demands of urban [10] B. Hongliang and L. Changping, “A hybrid license plate extrac-
surveillance systems. tion method based on edge statistics and morphology,” in Proc. 17th
IEEE Int. Conf. Pattern Recognit., vol. 2. Cambridge, U.K., 2004,
pp. 831–834.
IV. C ONCLUSION [11] G. Li, R. Yuan, Z. Yang, and X. Huang, “A yellow license plate location
method based on RGB model of color image and texture of plate,”
This paper designs a novel filter which is unified to detect in Proc. 2nd Workshop Digit. Media Appl. Museum Heritages, 2007,
either the vehicle license plates or the vehicles from the digital pp. 42–46.
[12] A. Ahmadyfard and V. Abolghasemi, “Detecting license plate using tex-
camera imaging sensors of urban surveillance systems in smart ture and color information,” in Proc. Int. Symp. Telecommun., 2008,
cities. To the best of our knowledge, we are the first to design pp. 804–808.
this kind of filter to detect the vehicle/license plate objects. We [13] L. Luo, H. Sun, W. Zhou, and L. Luo, “An efficient method of license
plate location,” in Proc. 1st Int. Conf. Inf. Sci. Eng., 2009, pp. 770–773.
tested our filter with different images. The results show that [14] G. Sun, C. Zhang, W. Zou, and G. Yu, “A new recognition method
our filter can automatically detect the highest energy frequency of vehicle license plate based on genetic neural network,” in Proc. 5th
areas out from the images, which makes our proposed algo- IEEE Conf. Ind. Electron. Appl., 2010, pp. 1662–1666.
[15] Y. Liu, D. Wei, N. Zhang, and M. Zhao, “Vehicle-license-plate recog-
rithm a simple and effective method to automatically detect nition based on neural networks,” in Proc. IEEE Int. Conf. Inf. Autom.,
vehicle objects for IoT and smart cities applications. Our 2011, pp. 363–366.
method can also be used to reduce the big data volume which [16] J. Sharma, A. Mishra, K. Saxena, and S. Kumar, “A hybrid technique
for license plate recognition based on feature selection of wavelet trans-
is generated every day from urban surveillance systems in form and artificial neural network,” in Proc. Int. Conf. Rel. Optim. Inf.
smart cities. Technol., Faridabad, India, 2014, pp. 347–352.
[17] H. Rajput, T. Som, and S. Kar, “An automated vehicle license plate
recognition system,” Computer, vol. 48, no. 8, pp. 56–61, Aug. 2015.
R EFERENCES [18] O. Mendoza-Schrock, M. M. Rizki, and V. J. Velten, “Manifold and
[1] S. Chen, H. Xu, D. Liu, B. Hu, and H. Wang, “A vision of IoT: transfer subspace learning for cross-domain vehicle recognition in
Applications, challenges, and opportunities with China perspective,” dynamic systems,” in Proc. 18th Int. Conf. Inf. Fusion, Washington,
IEEE Internet Things J., vol. 1, no. 4, pp. 349–359, Aug. 2014. DC, USA, 2015, pp. 1954–1961.
[2] M. Handte, S. Foell, S. Wagner, G. Kortuem, and P. Marrón, “An [19] W. Pei et al., “A rapid vehicle recognition and retrieval system,” in Proc.
Internet-of-Things enabled connected navigation system for urban bus 2nd Int. Conf. Syst. Informat., Shanghai, China, 2014, pp. 748–753.
riders,” IEEE Internet Things J., vol. 3, no. 5, pp. 735–744, Oct. 2016. [20] G. Burresi and R. Giorgi, “A field experience for a vehicle recognition
[3] B. Li, B. Tian, Q. Yao, and K. Wang, “A vehicle license plate recog- system using magnetic sensors,” in Proc. 4th Mediterr. Conf. Embedded
nition system based on analysis of maximally stable extremal regions,” Comput., 2015, pp. 178–181.
in Proc. 9th IEEE Int. Conf. Netw. Sens. Control, Beijing, China, 2012,
pp. 399–404.
[4] C.-N. E. Anagnostopoulos, I. E. Anagnostopoulos, I. D. Psoroulas,
V. Loumos, and E. Kayafas, “License plate recognition from still images
and video sequences: A survey,” IEEE Trans. Intell. Transp. Syst., vol. 9,
no. 3, pp. 377–391, Sep. 2008.
[5] M. Yu and Y. D. Kim, “An approach to Korean license plate recognition
based on vertical edge matching,” in Proc. IEEE Int. Conf. Syst. Man
Cybern., vol. 4. 2000, pp. 2975–2980.
[6] A. M. Al-Ghaili, S. Mashohor, A. R. Ramli, and A. Ismail, “Vertical- Ling Hu, photograph and biography not available at the time of publication.
edge-based car-license-plate detection method,” IEEE Trans. Veh.
Technol., vol. 62, no. 1, pp. 26–38, Jan. 2013.
[7] J.-W. Hsieh, S.-H. Yu, and Y.-S. Chen, “Morphology-based license plate
detection from complex scenes,” in Proc. 16th IEEE Int. Conf. Pattern
Recognit., Québec City, QC, Canada, 2002, pp. 176–179.
[8] A. A. Lensky, K.-H. Jo, and V. V. Gubarev, “Vehicle license plate detec-
tion using local fractal dimension and morphological analysis,” in Proc.
1st IEEE Int. Forum Strategic Technol., 2006, pp. 47–50.
[9] S.-Z. Wang and H.-J. Lee, “Detection and recognition of license plate
characters with different appearances,” in Proc. IEEE Intell. Transp. Qiang Ni (M’04–SM’08), photograph and biography not available at the time
Syst., vol. 2. Shanghai, China, 2003, pp. 979–984. of publication.