=Paper=
{{Paper
|id=Vol-2803/paper25
|storemode=property
|title=Development of an automated system for recognizing the parameters of a railway carriage (railway tanks) (short paper)
|pdfUrl=https://ceur-ws.org/Vol-2803/paper25.pdf
|volume=Vol-2803
|authors=Andrei V. Zabrodin,Svetlana E. Frolova,Alexander P. Glukhov
}}
==Development of an automated system for recognizing the parameters of a railway carriage (railway tanks) (short paper)==
https://ceur-ws.org/Vol-2803/paper25.pdf
Development of an automated system for recognizing the
parameters of a railway carriage (railway tanks)
Andrei Zabrodina, Svetlana Frolovaa and Alexander P. Glukhova
a
Emperor Alexander I St. Petersburg State Transport University, Saint Petersburg, 9 Moskovsky pr., 190031, Russia
Abstract
The solution to the problem of automated recognition of the parameters of rolling stock units in
real time is considered. A method for solving the problem of recognizing the main identification
marks of a railway carriage in real time is selected. A specific example of the implementation of
the method proposed in the article is given. The results of the developed program are presented.
The application of the developed system will allow to completely solve the problem of automated
recognition of the parameters of rolling stock units, and save money by avoiding downtime, delays
in return of wagons and conducting reasoned work with contractors.
Keywords 1
identification of standard signs of freight railway wagons, effective recognition methods, neural networks
1. Introduction The search and identification of numbers of
railway rolling stock units is relevant, since every
Today, an important aspect of the efficiency day the need to automate control of entry into the
of using railway transport and monitoring its territory of objects and reduce the influence of the
condition is the identification of rolling stock human factor increases. In this regard, the
(RS) cars by their inventory number. This is due problem arises of automated recognition of the
to the fact that a unique inventory number parameters of railway RS cars in real time, the
assigned to a railway car can determine its main solution of which can be carried out using neural
characteristics and provides code protection for networks. The purpose of developing the
reliable reading of the railway car number [6]. software package is to recognize not only the car
Car number recognition systems find various number, but also such parameters of railway
applications, for example, in commodity rolling stock units as: boiler calibration sign,
production and metrology services, security administration code, etc. The article presents a
services, logistics departments and railway shops. solution to the problem of automated recognition
The main problem of optimizing production of the main identification marks of railway RS
processes at the station is that the process of units in real time based on the use of neural
monitoring the movement of objects of the rolling networks.
stock of railway transport, including their
identification, is not automated at the processing 2. Formulation of the problem
station. Today, hundreds of employees are
involved in the control of rolling stock, who An automated system for recognizing the
ensure the appropriate production and parameters of rolling stock units must create 3
technological processes. Considering the fact that images in three projections of each passing car
the survey of the cars is carried out manually, the (on both sides and on top of the car); and 2 photos
final operations with the car take a lot of time and of the beginning and end of the car.
this process is economically expensive. The problem statement can be presents as
follows: an automated system for recognizing the
Models and Methods for Researching Information System
in Transport, Dec. 11-12, St. Petersburg, Russia
EMAIL: zabrodin@pgups.ru (A.V. Zabrodin),
frolova.svetlana19@yandex.ru (S.E. Frolova);
ORCID: 0000-0002-5578-280X (S.E. Frolova);
©️ 2020 Copyright for this paper by its authors. Use permitted under Creative
Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
178
parameters of rolling stock units, consisting of the applied to the railway carriage). After the train
next components: a frame with cameras, a server passes through the installation with cameras and
with a recognition program, and an automated the processing of all received images is
workstation (AWS) of the user. The work is completed, the information about the train is
carried out as follows: when the train passes transmitted to the user AWS. A general view of
through the installation with cameras, each car is the system operation process is shown in Figure
photographed from several angles, then all 1. The model “Automated system for recognition
images are transmitted in real time to the server, of rolling stock units” is built on the basis of the
where the images are stored and processed objects of the use case diagram.
(search and identification of identification marks
Figure 1: The model “Automated system for recognition of rolling stock units”
Color images from cameras installed on the An example can be the problem of the simplest
frame are fed into the recognition program, at the forecasting, calculating errors, or approximate
output the data is transferred to the user AWS, solution of equations.
basic information is available to the user: date and In cases when problems of high complexity
time of recognition, identification number, boiler arise, the solution of which is not clear in
calibration sign, administration code and load advance, neural networks are used. The class of
capacity. such tasks includes: image recognition, speech
recognition and complex predictions [2].
3. Choosing a method to solve the The range of tasks successfully solved with
the help of neural networks has expanded
problem significantly in recent years. First of all, various
tasks of image processing and video data analysis
To solve the problem, you can use the began to be solved most actively: the detection of
following methods: accurate algorithms, static various objects in the image and their
models, neural networks. classification. This includes the tasks of detecting
Exact algorithms are used to solve problems and recognizing faces and car numbers; searching
of low / medium complexity with a specific and locating people and other objects in the
solution algorithm. An example would be: frame; detecting fire, smoke, water where they
solving a simple arithmetic equation; displaying should not be, etc. [7].
the program window; printing a document on a Another large class of problems in which
printer. neural networks have been successfully used, -
Special statistical methods are used in cases word processing in natural language. These are
where problems of low / medium complexity all kinds of texts classifications (for example,
arise, the solution to which is not fully defined. classification reviews to positive and negative),
179
machine translation, chat bots and task bots - localize objects (search for their areas in the
programs that can replace, for example, ticket image), the second for their identification
sellers [5]. (recognition of numbers). The third, auxiliary
In addition, neural networks are used in the unit performs the function of monitoring the
analysis of scientific data, games, in finance to correctness of the railway car number
assess customers and risks of various types, and recognition.
in many other areas. As a result, neural networks A variety of labels on railway tanks prevents
work well with the data, in which there is a determine the necessary parameters correctly. In
correlation of the measurements (images, sound, such cases it is necessary to apply the detector is
text, time series) [3]. designed to localize objects. The objects localizer
In view of the above, as a method for solving successfully works where objects of the required
the problem, a solution using neural networks type "overlap" each other in the frame or there is
was chosen. a large amount of visual interference. Finding
In the process of modeling and building a areas correctly will allow you to exclude other
neural network, it is important to consider the labels. After training, the detector will filter out
quality of the source data. Stencils applied to the unnecessary "noise" and due to this, the accuracy
railway carriage, have a very large variety of of parameter recognition will significantly
colors, fonts, size and location. This is due to the increase.
fact that the place where the stencil is applied is To solve the subtask described above, a
not fixed and depends on the model of the tank, neural network localizer was designed on the
and the boilers of the tanks are always painted in Mask-RCNN architecture. This architecture has
a different color. While tanks boilers intended for several advantages: recognizes objects in the
the second class of goods, colored in light gray entire image, efficiently consumes computing
(silver) colors, coppers tanks for transportation of resources [4], has a high counting rate, allows to
methanol - highly toxic flammable liquid of the recognize objects in the video, easy to learn.
third class are colored yellow. In addition, you To build a localization model for the railway
can find railway carriage painted in blue, green, car number, in the project added to the class
orange and other colors. AreaSearchNetwork. In the __init__ block,
In addition, when recognizing the constants and a detector with a default
identification marks applied to the railway configuration file are initialized, directories are
carriage, the classical analytical algorithms being configured. The "work" method was
detector finds in the frame any character sets written to work with the localizer: the input is an
similar to numbers and tries to recognize them. image; the output is an array of images, each of
To improve the accuracy of recognition of which is the parameter of a railway carriage
identification marks of freight cars, it is necessary (number, administration code, boiler calibration
to solve such a subproblem as localization of sign, etc.).
objects. After localization, each parameter is
identified using a simple neural network detector,
4. Designing and building a neural which is trained to search and recognize numbers.
It works like this: the input comes from the image
network model area, which has a parameter; at the output - the
number of the recognized parameter.
The current project is based on the Mask- After the railway train passes through the
RCNN-based Nomeroff Net project. Nomeroff installation with cameras and the processing of all
Net is system for recognizing the licence plate the images obtained, information about the train
number of a car, written in Python and open is transferred to the user AWS. The subsystem
source. The entire project recognition of the "Program for recognizing the parameters of
parameters of rolling stock units consists of three rolling stock units" is shown in Figure 2. Class
blocks: two main and one additional. diagram objects and use case diagram objects are
The main blocks are neural network used for display.
detectors. The first detector is designed to
180
Figure 2: Subsystem "Program for recognizing the parameters of rolling stock units"
2. Annotation of data - determination of the
5. Neural network training and required area in the image and its description,
is carried out using the VGG Image Annotator
testing (VIA) [1]
3. Running a script for training a neural
An important property of neural networks is network on a GPU in GoogleColab
their ability to learn from environmental data and, 4. Saving the trained model (file with the
as a result of training, increase their performance. .h5 extension)
Productivity increases over time according to Examples of the work of the trained localizer
certain rules. and identifier are shown in Figure 3 and Figure 4.
A short neural network training algorithm is as During the operation of the localizer, the
follows: following were found in the image: tank, number
1. Data preparation - includes splitting all and sign of the boiler calibration. It can also be
images that will be used for training into test, noted that the number recognized by the program
validation and training sets
matches the real number of the car.
181
Figure 3: The result of the work of the trained localizer
Figure 4: The result of the work of the trained identifier
6. Description of an example of the
recognition program 7. Conclusion
The initial image of the car arrives at the The article discusses an automated system for
entrance to the program for recognizing the recognizing the parameters of a railway carriage
identification parameters of rolling stock units (railway tanks). The proposed solution is
(fig. 5, a). After the work of the localizer, the universal and does not depend on the recognized
areas in which the parameters of a railway objects. It has a rational structure that ensures
carriage are located are highlighted (fig. 5, b). efficient use of computing power.
Further there is identification of parameters and As part of solving the problem, the possibility
forming a data block with information about the of using neural networks for solving problems of
freight railway wagon (fig. 5, c). The correctness finding and recognizing given objects
of the recognized railway car number is checked (signatures) with the required accuracy is shown.
by the checksum. After similar processing of all Application of this solution allows automating
received images, the collected information about the process of monitoring rolling stock at the
the rolling stock is transferred to the user AWS. processing station, which reduces the time of
final operations with the freight railway wagons;
creates conditions for optimizing the time spent
on processing the wagons, as well as the
emergence of material savings due to the
optimization of production processes at the
station.
References
[1] Abhishek Dutta and Andrew Zisserman, The
VIA Annotation Software for Images, Audio and
Video. In Proceedings of the 27th ACM
International Conference on Multimedia (MM
’19), October 21–25, 2019, Nice, France. ACM,
New York, NY, USA, 2019. doi:
10.1145/3343031.3350535
[2] Charu C. Aggarwal, Neural Networks and
Deep Learning, IBM T. J. Watson Research
Center, International Business Machines,
Yorktown Heights, NY, USA, 2018. doi:
10.1007/978-3-319-94463-0
[3] Gafarov F. M., Artificial neural networks and
applications, Kazan, 2018, 121 p. (in Russian).
Figure 3: An example of the recognition program
182
[4] Kaiming He, Georgia Gkioxari, Piotr Dollar, [6] Reference book: Eight-digit numbering
Ross Girshick, Mask R-CNN, 2017-2018. system for 1520 gauge freight cars, Information
arxiv:1409.4842 and Computing Center of Railway
[5] Neural networks: a new breakthrough. Expert Administrations, 2005. 31 p. (in Russian).
opinions, Security systems, 2019, no. 1, pp. 54- [7] Yastrebov A., Neural networks took a break,
56 (in Russian). IT News, 2019.
183