Artificial Vision for Vehicle
Monitoring at Tren Urbano Stations
Jaime José Laracuente-Díaz
Advisor: Dr. Manuel Toledo-Quiñones
Electrical and Computer Engineering Department
University of Puerto Rico, Mayagüez Campus
Mayagüez, Puerto Rico 00681-5000
jjlaracuente7@hotmail.com
mtoledo@ece.uprm.edu
Abstract
This paper presents aspects of a license plate
recognition algorithm that is being developed
under the UPR/MIT/Tren Urbano Professional
Development Program. Techniques for license
plate extraction and character recognition are
described. Results of tests in which the license
plate is successfully extracted are presented. An
outline of future work is provided.
monitoring the "publicos" routes and schedule
necessary.
We propose to use a license plate recognition
system (LPRS) to monitor the "públicos"
presence at TU stations as a way to automatically
maintain records on their quality of service. With
this system a database of "públicos" presence at
the stations can be provided. The database can be
used for administration and planning purposes.
1. Introduction
Tren Urbano (TU) is a heavy rail transit system
currently under development and construction in
the San Juan Metropolitan Area at Puerto Rico.
The first phase of the project will provide service
to approximately 115,000 passengers per day in
one of the most densely populated and centrally
developed corridors in the United States. It will
consist of 16 stations and the total extension is
11.8 miles [Wilson02].
A LPRS is used to identify vehicles by
automatically recognizing their license plates.
These systems usually include a video camera
that is installed at a proper place to capture front
or rear images of vehicles. The image is
processed to extract the portion containing the
license plate and to identify the vehicle. Several
techniques can be used for this. Specific methods
should be selected according to the application
requirements.
Today, the San Juan Metropolitan area has
several public transportation services such as,
buses, taxis, and "públicos". "Públicos" is a
privately owned mini-vans system. The van's
driver is usually its owner.
The proposed LPRS system can also be used to
monitor other vehicles. In addition, with minor
modifications it can be used in other
transportation-related applications, such as to
provide enhanced parking lot security.
The "públicos" are a very important part of the
transportation infrastructure of Puerto Rico.
Their integration is essential for the success of
TU. A major challenge to the integration process
is the lack of predictable schedule that
characterizes their current mode of operation.
Some form of additional economic compensation
could be an effective incentive for "públicos" to
cover their routes in a way compatible with the
functioning of a modern transportation system,
such as TU. This will, however, make
This paper describes aspects of a license plate
recognition
algorithm
being
developed.
Techniques for license plate extraction and
character recognition are presented. Results of
tests in which the license plate is successfully
extracted and an outline of future work are
provided.
2. License Plate Extraction
The algorithm used in LPRS consists of the
following steps: 1) capture the car's image, 2)
extract the license plate, 3) recognize license
plate characters, 4) identify vehicle. This section
describes each of these steps.
We have been working with color images
captured with a digital camera and transferred to
computer memory. Our images consist of a
rectangular array of 768x512 pixels. To obtain an
8-bit grayscale image, an algorithm that
eliminates the hue and saturation information
while retaining the brightness or light intensity
information (luminance) is applied.
To determine the portion of the image that
contains the license plate, the different objects
present in the image are found and labeled.
Those that have geometrical characteristics
similar to those of license plate characters are
selected. If the relative position of the selected
objects is consistent with a license plate, it is
assumed that the plate has been found.
To facilitate the labeling process, the image is
segmented according to a threshold level Θ.
Segmentation is the separation of an image's
background and foreground. There are several
segmentation techniques that can be used in
order to achieve the proper license plate
extraction. We explored techniques based on
edge detection and thresholding, and selected the
later because of its superior performance. This
technique takes the grayscale image g and
produces binary image b, by applying a threshold
to each pixel according to
if g[m,n] ≥ Θ then b[m,n]=1
if g[m,n] <Θ then b[m,n]=0
where m and n represent the pixel's row and
column.
In our tests, the threshold level was initially
chosen as the image's mean pixel value. This
produced a black-and-white image of 1-bit pixe ls
in which most of the background is absent, and
facilitates the labeling of the objects. An example
of a segmented image is shown in Figure 1.
In the next step, objects are labeled using a
technique known as connected components
labeling. To recognize objects the algorithm
searches every pixel and classifies groups of
ones surrounded by zeros as different objects.
Figure 1. Binary image
Each object is numbered or labeled. Figure 2
shows an example of a label matrix with three
numbered objects.
L= 1 1 1 1 1 1 1 0000000000
1 1 1 1 1 1 0000002 2 2 2 2
11 000000000000222
00003 3 3 3 3 3 0000000
00003 3 3 3 3 3 0000000
Figure 2. Label matrix
The labeled objects are then analyzed and those
that possess geometrical characteristics similar to
those previously established for license plate
characters are selected. The characteristics we
have used so far are height and width, but others,
such as aspect ratio, are also possible.
Distance between the selected objects is then
calculated. If the number of characters found and
their relative position is consistent with plate
characters, the region of the image containing the
desired information has been identified and the
characters can be extracted. The plate characters
extracted from the image in figure 1 are shown in
Figure 3.
If attempts to find the region containing the plate
fail, the segmentation threshold would have to be
modified and the whole procedure performed
again. This process would continue until
satisfactory results are obtained. In our tests, our
initial choice worked very well, and no iterations
were needed.
Another improvement that we will explore is the
use of color. This will allow us to remove many
unwanted objects from the image, further
improving the robustness of the program. The
use of color, however, requires a color camera.
Although one is available for development, it is
desirable to use a grayscale camera in the device
that will be deployed on the field to reduce cost.
Thus, color should be used only if found to be
necessary.
Figure 3. Extracted license plate
To recognize the plate characters, a possible
algorithm is template matching. This approach is
based on pattern recognition principles. There
are several ways to implement the character
recognition algorithm. One is based on
comparing character patterns in a predefined set
and those in the image being analyzed to select
the one with maximum correlation. This is done
by counting the number of agreements (ones
matching ones and zeros matching zeros) and
choosing the character for which a maximum
number is obtained. The correlation coefficient is
defined as follow:
R = ∑∑ ( Amn − Amean)( Bmn − Bmean)
∑∑ ( A
mn
− Amean) ∑∑ ( Bmn − Bmean)
3. Results and Future work
We have developed an algorithm that reliably
locates license plates in images. An example of
test results is shown in Figures 1 and 3. Figure 4
shows a group of binary images and extracted
license plates from several tests.
Our algorithm relies on plate character size for
detection. As long as the distance between
camera and plate is the expected, the algorithm
works very well. We plan to modify the program
to use character aspect ratio, rather than width
and height, to identify the objects that might be
part of the plate. In this way character size
should not play a significant role and the
program should become more robust to
variations in the distance between camera and
plate.
The next task in our schedule is to implement
license plate character recognition. Since to find
the license plate in the image is considered to be
the most difficult task of a LPRS, we expect
character recognition to be relatively easy.
The recognized vehicle information will be
stored together with other relevant information in
a data base for later use. Early in the project we
decided that to make the system easier to
maintain and use, an embedded implementation
was the best choice. We envision these devices
to be modular and very easy to install, suitable to
be deployed in Tren Urbano's stations. Since the
use of images involve large amounts of data, we
think that to use an embedded computer to run
all algorithms and identify the vehicles locally is
the most practical solution. Only the results
(vehicle identity and time in the station) need to
be transmitted to a conventional computer for
further use. This reduces the required network
bandwidth and makes the use of a wireless link
feasible. In the last phase of our project we will
evaluate the porting of our prototype to such
embedded LPRS system.
References
[Baéz98]Báez-López D., Ramírez J.M. Pattern
Recognition in automotive Plates. In
Midwest Symposium and Circuits, 1998.
[Cui97]Cui Y., Huang Q. Automatic license
extraction from moving vehicles. In The Int.
Conf. On Image Processing, 3- volume set,
1997.
[Cui97]Cui Y.-T., Huang Q. Character
Extraction of license plates from video. In
Conf. On Computer Vision and Pattern
Recognition, 1997.
[Kamat95]Kamat V., Ganesan S. An efficient
implementation of the Hough transform for
detecting vehicle license plates using
DSP'S. In The Real-Time Technology and
Applications Symposium, 1995.
[Naito98] T. Naito, T. Tsukada, K. Yamada, K.
Kozuka. License Plate Recognition Method
for Inclined Plates Outdoors. In Int. Conf.
On Information Itelligence and Systems,
1998.
[Setchall97] Setchall C.J. Applications of
Computer Vision to Road-traffic
Monitoring, PhD Thesis in Electrical
Engineering, University of Bristol,
September 1997.
[Shridhar98] M. Shridhar, J.W.V. Miller,
G.Houle, L. Bijnagte. Recognition of
License Plate Images: Issues and
Perspectives. In the Fifth Int. Conf. On
Document Analysis and Recognition, 1998.
[Wilson02] Wilson, Nigel H. M ., The MIT/Tren
Urbano Project Web Site,
http://www.mit.edu/afs/athena/org/t/trenurbano/
Figure 4: Results.