You are on page 1of 5

Fire Detection System with Utilization of Industrial

Video Surveillance System

SABARI V VIGNESHWARAN P PRUSHOTHAMAN


Electrical and Electronics Engineering Electrical and Electronics Engineering Electrical and ElectronicsEngineering
Kalaignarkarunanidhi Institute of Kalaignarkarunanidhi Institute of Kalaignarkarunanidhi Institute of
Technology Technology Technology
(Affiliated to Anna University) (Affiliated to Anna University) (Affiliated to Anna University)
Coimbatore,India Coimbatore,India Coimbatore,India
Sabariveerasekar2004@gmail.com vwrv8825@gmail.com Prushothaman2001@gmail.com

Abstract - Fire occurrence is the most destructive hazard for The software product was developed and implemented in
production, as it causes irreparable damage to the enterprise, the high-level general-purpose programming language Python
equipment, health and life of people, as well as the environment. with additional libraries for image processing.
Therefore, recognizing fires and preventing them in a timely
manner is very important. Modern production facilities are Most of the equipment of a typical industrial oil refinery
equipped with widely distributed video surveillance systems. By faces strict operating conditions and a high risk of fire and open
analyzing the transmitted images, the algorithm is able to fire. The main risk is related to the delayed response of service
recognize the source of fire and activate an alarm, thereby personnel to emergency situations. This happens usually due to
ensuring the safety of the enterprise. The algorithm deploys the absence of the operator at the workplace or his inattentive
various characteristics of flame images during data processing. observation of the process. For immediate response of
The mixed Gaussian model is used to distinguish the dynamic operators and special services, it is better to automate the
area from the static background, and the color characteristics of process of fire recognition. If a fire is detected, the system's
pixels in the RGB model are analyzed to detect fire. In order to response must be accompanied by a sound alarm and a
increase the accuracy of the results obtained, the frequency of message sent to the security services.
changes in the contour of the dynamic area (flickering) in the
target area of fire is checked. It is not always possible to use smoke detectors or gas
analyzers on industrial sites, especially in open spaces.
Keywords - industrial video surveillance, petrochemical Therefore, the task of monitoring the state of the process is
industry, fire·detection, color model, gaussian distribution assigned to the operator. This work is routine and requires
constant presence of the operator at the workplace. The human
I. INTRODUCTION factor impact, for example, inattention, fatigue or drowsiness of
Industrial video cameras of explosion-proof design are the employee on night duty should also be kept in mind.
installed at certain oil refining facilities, which are most often
The main goals of human replacement in the process safety
prone to fire or emergency situations. These facilities include:
monitoring are as follows:
 Pump stations;
 Releasing the operator from monotonous routine
 Compressor stations; operations to solve other more important tasks;
 Process heaters;  Reducing material losses through timely response;
 Reactors;  Improving the speed of emergency response.
 Fractionators; The implementation of a fire detection software package
will reduce the response time to local fires and mitigate
 Feedstock tank parks. material and human losses. The implementation proposed in
The main task of video surveillance is to record emergency the project will also reduce the operator's activity, leaving him
situations in order to further clarify the causes of their only partial control over the system itself as a whole.
occurrence. However, due to the human factor, it is not always II. MATERIALS AND METHODS
possible to respond to emergencies in a timely manner and
avoid major damage or casualties. Let us imagine the structure of the recognition system. This
system processes and analyzes images received from
Within this project, various methods of image recognition optoelectronic systems. Given the nature of the task, the
were studied and a software package was developed for following limitations and features must be considered:
automated recognition of open fire or ignition of oil refining
equipment.  Fire can occur at any time of the day, so there may be
errors in the form of sun glares or lanterns;
 Due to the limited time to make decisions and the The main part of the fire recognition algorithm is connected
presence of a closed loop of traffic control, processing, analysis with the background detection among moving areas or dynamic
and recognition must be performed in the real time; objects, such as flame.
 The system must operate in an independent mode. One approach to implementation comprises the use of a
neural network that receives the spatial and temporal
Recognition will take place in several stages, the main ones characteristics of fire as input data. The flame area is
of which are searching for changes in the moving pixels, computationally cut out of the image using the statistical
determining the found pixels by color, and checking the parameters of the color models described above. Then the
movement of the checked area of fire (flickering). defined contour is linked to the polar coordinates of the initial
In order to determine the fire, two main approaches to the image to transform a series of frames into a series of
algorithm formation can be used. The first approach pertains to fluctuations in time. In order to obtain the frequency spectrum,
initial identification of colors similar to fire in the image, and the received signal is converted into Fourier space. The
then search for dynamic areas, i.e. areas where movement frequency spectrum is transmitted to a trained neural network.
occurs or there is a presumable flame. A training table should be created at the development stage for
the respective training process. The described method has a
The second approach provides for the initial search for high accuracy of the results obtained, but due to the high
dynamic areas, and after that, the identification of the area of computational complexity, it is not applicable in this work.
the presumable fire by the number of pixels similar to the fire
in color in different stages of its activity. Another way to separate dynamic and static objects is to
use the mixed Gaussian model. It is simple and effective at the
The selection of the implemented approach should be made stage of modeling a static background, by tracking changes in
based on the characteristics of the controlled area. For example, all pixels over time and mixing К normal distributions for a
if the scenery has a lot of dynamic objects, but the color sequence of frames in the video stream. К-distributions are
scheme does not have colors similar to the flame, then the first divided into two groups: the foreground group and the
approach should be selected, because this solution would background group. If a pixel belongs to the background
reduce the computational time of the process, completing the distribution, it is considered background. The same is true for
selection at the first stage. Conversely, if the controlled area the group of foreground pixels. This method can analyze the
has a lot of colors similar to the flame in color, for example, background of any complexity, but it has a low calculation
illuminating lamps, then the second approach should be speed, since the analysis takes place at the pixel level and at
implemented. high video resolution, the implementation of the method
First, let us consider the stage of determining the color requires a lot of time.
characteristics of pixels. The method is based on the analysis of Let us consider an improved method based on the Gaussian
the space of color mathematical models. Four color models can model. We denote the vector of pixel color characteristics in
be distinguished. the RGB color model in coordinates (x, y)

X xy,t  R xy,t , Gxy,t , Bxy,t ,


The RGB model that is the most common. Color T
reproduction is performed by adding three main colors: red,
green, and blue.
where t is any moment in time.
YCbCr color space. It is one of the variations of the RGB
Over time, a series of observations is formed as
 
model, where Y-brightness, Cb and Cr are the blue and red
chrominance components. A fuzzy logic system is used to X xy,1 , X,..., X xy,t , which is represented as a stochastic
xy,2

describe the fire image. The image is fuzzified - it is converted process relative to other pixels. The generated series has a
into a fuzzy set. Then the degree of membership of the probability distribution density function. In subsequent
elements in the fuzzy set changes in accordance with the fuzzy analysis, if each pixel is described by this function sufficiently,
rules. the area will be considered background. It ought to be noted
that pixel parameters are subject to complex distributions,
HSL color model. A model in which color coordinates are
which makes it more difficult to select and adapt the
described by three components – tone, saturation, and intensity,
distribution density function described above. Therefore, over
respectively.
time, the pixel is described К by normal distributions. With К
HSV color model. Analogous to the HSL model, only the increasing, the complexity of calculations increases as well.
last coordinate expresses the color value.
We define the probability of observing a vector of color
The parameters of HSL and HSV models represent a characteristics as
knowledge base that the operator creates by specifically K

these methods, a large knowledge base, which is not always  xy,k ,t



highlighting P( X xy,t )  w ( X xy,t , u xy,t , xy,k ,t )


convenient, istherequired,
flame inbecause
the image. However,
there in order
is an infinite to apply
number of
wxy ,k ,t k 1

the flame variations. where is the weight factor, which is the probability
that the observation belongs to k-th component;
u xy ,t – vector of mathematical expectations; of the mixed model should be updated. If there are matches
with K-distribution at the moment of time t, then u xy ,t and
 xy,k ,t – covariance matrix of the k-th normal distribution at
 xy ,k ,t are updated according to the current values X xy,t , but
the moment of time t; remain unchanged for non-matching distributions. The
 - normal distribution density function. described adaptation is implemented by the following
equations:
1
( X xy ,t , u xy ,t , xy ,k ,t )  n 1 e
1/ 2( X xy ,t uxy ,k ,t )T 1
xy , k ,t
( X xy ,t uxy , k ,t )
wxy ,k ,t1  (1   )wxy ,k ,t  M xy ,k ,t ,
(2 ) 2 xy ,k ,t 2 
 u xy ,k ,t 1  (1   xy ,k ,t )u xy ,k ,t   xy ,k ,t X xy,t ,
 2 xy ,k ,t 1  (1   )2 xy ,k ,t   ( X xy ,t  u xy,k ,t 1 )T ( X xy,t  u xy,k ,t 1 ),
where n is the number of channels, since we use the RGB  xy ,k ,t xy ,k ,t
model, it is equal to three. 

 xy ,k ,t   ( X , u , 2
xy ,k ,t I ).
In order to reduce the calculation time, we will assume the xy ,t xy,k ,t

red, green, and blue color channels for each pixel as where α is the background update rate, varying from 0 to 1.
independent and having the same dispersion. If there are no matches for the current observation vector
  diag 2 , 2 , 2   2 I, X xy,t with K-distribution, the component with a small value
xy,k ,t R G B xy,k ,t
w /  is updated in accordance with the expectation set by
where I is the unity matrix; X xy,t large variation, and low weight.
 – standard deviation (mean square deviation). In order to improve the accuracy of the results, the flicker
The greater the value of the weight (w) of the normal detection method will be applied. Flickering, or flickering
distribution, the more observations will fit the normal movement of the flame is one of the characteristics of fire that
distribution. The stability of the distribution ensures a small allows distinguishing the flame even against objects close to it
deviation  . The ratio w /  determines the status of the pixel, in color. The phenomenon of flickering is a random change in
whether it belongs to the background or to the foreground, by the boundaries of the flame from frame to frame. Flickering of
K-distribution. The dynamic area is described by several the flame occurs on average with a frequency of 10 Hz. Thus,
distributions with small weights. Therefore, the first В in the by selecting pixels with a given color change frequency, more
distribution of the total number of К are determined for the accurate data on the presence of fire can be obtained.
background assessment, while the rest ones as the foreground.
III. RESULTS
Bxy ,t  arg min B  wxy,k ,t  T 
B
. As a result of the methods described above, the fire
recognition algorithm is shown in Figure 1.
 k 1 
The value T indicates the threshold, the minimum part of As a result of the algorithm, the image of the video stream
which is considered the background. For each new pixel, T can is displayed on the screen. If the program detects a fire, its
presumable area is highlighted with a rectangle. Various flares,
match one of the largest components of K-distribution. This
reflections, and direct sunlight can be mistaken for flames, so
match is possible if the distance between X xy ,t and u xy,t is less the program processes only flickering light sources. Also, to
than the threshold value  . filter out small noises, the minimum threshold area of the
dynamic area was set.
To find the first matching distribution, each new pixel
observation is compared with the existing K-distribution In order to test the program, video files with images of the
component. fire, which are freely available, were used. The results of the
program processing are shown in Figures 2 and 3.
1, при X xy,t  u xy,t  
M   xy,k ,t ,

xy,k ,t 
0, при X xy,t  u xy,t   xy,k ,t ,


where M xy,k ,t is the matching function of the k-th
distribution at the moment of time t;
 - a number in the range {2.5; 3} to achieve an accuracy
level of 95%.
If the analyzed pixel matches B-distribution, it is considered
background. Otherwise, it will be a foreground pixel.
For practical use, the change in illumination of the
Fig. 2. Results of the software product testing
controlled object should be taken into account. The parameters
IV. DISCUSSION
Получение на вход
Image obtained The implemented program can process a single video
изображения
at input
stream at any frame rate. For the future use in enterprises, the
program must be updated by increasing the amount of input
data. Reporting on detected fires should be planned as well.
During testing, the filter range of the color model and the
Поиск
Search for
динамических flickering frequency range of the fire were adjusted. In addition
dynamic areas
областей to the ability to process video recordings, signal processing has
been implemented using the RTSP Protocol.
At present, only one program presented in the consumer
market can recognize fire, this is a program created by the
No
Нет Динамическая Macroscop company. There are no other registered programs
Dynamic area found?
область найдена? with implementation based on our algorithm. Accordingly, the
developed solution can be considered new in the recent times.
ДаYes
The developed program has demonstrated its efficiency in
practice and coped with the assigned task. The program
recognizes flame in the video sequence and does not react to
Извлечение
Extractingцветовых
color
характеристик
characteristics other dynamic objects. The presented software product
interacts with only one video stream, but in the future, it is
planned to expand its functionality to operate with multiple
video streams, since one computer is most often connected to a
multi-stream industrial video surveillance system in the
No
Нет Обнаружен цвет
Color similar to
production conditions.
подходящий
flame found?под
пламя? For its operation, the program does not require high power
and is suitable for computers with average computing power.
ДаYes
ACKNOWLEDGMENT
Оценить
Assess
мерцание
flame The work was carried out with the financial support of the
пламени
flickering Russian Foundation for Basic Research, Project No. 19-08-
00261.

No
REFERENCES
Нет Мерцание
Flickering
подтверждено?
confirmed? [1] O.E. Bandeev, “Classification of flame detection methods by video
sequences,” Actual problems of aviation and astronautics, no. 13, 2017,
ДаYes pp. 91-93.
[2] HSL. (n.d.). Retrieved from https://ru.wikipedia.org/wiki/HSL.html
[3] HSV (color model). (n.d.). Retrieved from
Включить тревогу,
Activate alarm, https://ru.wikipedia.org/wiki/HSV_(цветовая_модель).html
выдать сообщение
send message to
оператору [4] G.A. Polte, and A.P. Saenko, “Quantitative evaluation of image quality
operator
using fuzzy logic methods. Instrumentation,” vol. 54, no. 3, pp. 32-36,
2011.
Fig. 1. Block diagram of algorithm implementation [5] Industrial and process television systems. (n.d.). Retrieved from
http://www.itmicro.ru/industrial-tv.html
[6] Z. Zivkovic, “Improved adaptive Gaussian mixture model for
background subtraction,” Proceedings of the 17th International
Conference on. IEEE, vol. 2, pp. 28-31, 2004.
[7] T. Celik, Y. Demirel, H. Ozkaramanli, and M. Uyguroglu, “Fire
detection using statistical color model in video sequences,” IEEE
International Conference on Acoustics, Speech and Signal Processing,
vol. 2, pp. 213-216, 2006.
[8] Macroscop. (n.d.). Retrieved from https://macroscop.com
[9] SecurOS video management integration platform. (n. d.). Retrieved from
https://iss.ru/products/securos-vms.html
[10] G. Bradski, Learning Open CV, O’Reilly Media, 2008, p. 576.
[11] P. Piccinini, S. Calderara, R. Cucchiara, “Reliable smoke detection
Fig. 3. Results of the software product testing system in the domains of image energy and color,” in 15ᵗʰ International
Conference on Image Processing. San Diego, California, USA, 2008, pp.
1376-1379.
[12] B. U. Toreyin, “Wavelet based real-time smoke detection in video,” 15th European Signal Processing Conference. Poznan, Poland, 2007, pp.
Signal Processing: Image Communication, EURASIP, Elsevier, vol. 20, 1794 – 1798.
pp. 255-26, 2005. [18] R. Yasmin, “Detection of Smoke Propagation Direction Using Color
[13] D.K. Kim, Y.-F. Wang, “Smoke Detection in Video”, World Congress Video Sequences,” International Journal of Soft Computing, vol. 4, no.
on Computer Science and Information Engineering. Los Angeles, USA, 1, pp.45-48, 2009.
2009, pp. 759-763. [19] M.M. Kaziev, E.V. Zubkova, and V.I. Bezborodov, “Features of the fire
[14] B. Ugur Toreyin, Yigithan Dedeoglu, A. Enis Cetin “Contour based protection transparent constructions with water irrigation.
smoke detection in video using wavelets”, In European Signal Pozharovzryvobezopasnost,” Fire and Explosion Safety, vol. 25, no. 8,
Processing Conferenc. Florence, Italy, 2006, p. 5. pp. 27–33, 2016. [DOI: 10.18322/PVB.2016.25.08.27-33].
[15] F. Comez-Rodriuezl, “Smoke Monitoring and measurement Using [20] M.M. Kaziev, A.M. Podgrushnyj, and A.V. Dudunov, “A pane of glass
Image Processing. Application to Forest Fires”, Automatic Target destruction under thermal influence in the conditions of a fire,” Fire and
Recognation XIII, Proceedings of SPIE, vol. 5094, pp. 404-411, 2003. Emergencies: Prevention, Elimination, 2009, no. 2, pp. 5–10.
[16] D. Krstinić, T. Jakovčević, D. Stipaničev, “Histogram-Based Smoke [21] A.V. Yaroslavtsev and O.V. Matukhina, “Creating the Web Service of
Segmentation in Forest Fire Detection System,” Information the Laboratory Information and Management System for Analytical
Technology and Control, vol.38, no. 3, pp. 237-244, 2009. Quality Control,” Lecture Notes in Networks and Systems, vol. 115,
[17] Turgay Çelik, Hüseyin Özkaramanli, and Hasan Demirel, “Fire and Industry Competitiveness: Digitalization, Management, and Integration.
smoke detection without sensors: Image processing based approach”, Switzerland: Springer, vol. 1, pp. 643-649, 2020.

You might also like