Grenze International Journal of Engineering and Technology, June Issue

Video Surveillance Fire Detection System using CNN

Algorithm
Prof. U. L. Tupe1, Lakhan Jadhav2, Shivanand Koli3, Prasad Kulkarni4 and Mayur Gaikwad5
1-5
JSPM's Rajarshi Shahu College of Engineering/Information Technology, Pune,India
Email: [email protected]

Abstract— Fires in public places such as shopping malls, hospitals, train stations, and others can
endanger both people and resources. This has been a big concern for the past few decades.
Stopping these accidents should be a priority. So to achieve these, there are various techniques,
but there are some loopholes in those techniques. So to overcome those loopholes, we have
developed a model that detects fire in images and video frames, and as soon as it detects fire, it
sends an alert message to the nearest fire station and related authorities. The main purpose of
these models is to prevent accidents due to fire and minimize human work load. This paper uses
the CNN algorithm to build these projects.

I. INTRODUCTION
Fire accidents are one of the biggest threats to industries, social gathering places, hospitals, malls, and various
densely populated areas across the world. These kinds of incidents may cause damage to property and the
environment and pose a threat to human and animal life. According to the recent National Risk Survey Report
[1], fire was ranked third in terms of its impact on various fields related to problems, among many others. There
were some fire accidents in different countries that resulted in the impending ecological disaster, claiming
millions of lives and resulting in billions of dollars in damage. Early detection of fire can save many lives, as
well as various resources, and prevent damage caused by fire. This is one of the advantages of detecting fire at
an early stage. In order to achieve high accuracy and robustness in dense urban areas, detection through local
surveillance is It is both necessary and effective. There were numerous issues with traditional fire systems, such
as false alarms, detecting fire when there was none, and ringing alarms when there was no fire. Maintenance was
difficult. The use of sensors in hot, dusty industrial conditions is also not possible. Thus, detecting fires through
surveillance video One of the most feasible and cost-effective solutions is streaming, which is suitable for the
replacement of existing systems without the need for large infrastructure installations or investments. The
existing domain knowledge is heavily used in video-based machine learning models.
As a result, they must be updated to meet new threats. So our model can detect fire in a video or image frame
and send an alert message as soon as it detects fire. It can be used to detect fires in surveillance videos. Unlike
existing systems, this neither requires special infrastructure for setup, like hardware-based solutions, nor does it
need domain knowledge and prohibitive computation for development.

II. PROPOSED FRAMEWORK

We investigate deep neural networks for potential fire detection in the early stages of the suggested system's
surveillance. For the objective problem, we investigate various deep CNNs while accounting for accuracy, the
embedded processing power of CCTV systems, and the frequency of false alarms. The computer vision problems

Grenze ID: 01.GIJET.9.2.334

© Grenze Scientific Society, 2023
and applications where CNNs have shown promise include object detection and localization, image
segmentation, super-resolution, classification, and indexing and retrieval. Their hierarchical method, which
automatically learns very potent qualities from raw data, might be credited for this widespread success.
A typical CNN design is made up of three well-known processing layers. 1) Numerous feature maps are created
as a convolution layer when various kernels are applied to the input data. 2) A pooling layer that, in order to
achieve some translation invariance and dimensionality reduction, chooses the maximal activation from a small
neighbourhood of feature maps acquired from the previous convolution layer. 3) A fully interconnected layer
that builds a global representation of high-level data from input information. The high-level characteristics of the
input data are produced by this layer after a sequence of convolutional and pooling layers. These layers are
arranged in a hierarchy, with the output of one layer serving as the input for the following one.
In convolutional kernels and fully connected layers, the weights of every neuron are adjusted and learned
throughout the training period. By simulating the characteristic features of the input training data, these weights
can carry out the objective classification. Pre-processing refers to all the adjustments made to the raw data in our
project before it is provided to the deep learning or machine learning algorithm. A CNN model, for instance,
would almost likely produce subpar classification results if trained on raw images. A neural network called CNN
extracts the features of the input images. The input image is used by a network for feature extraction. The feature
extraction signals are used by the neural network for categorization.

III. RELATED WORK

THE SDLC MODEL TO BE USED each software development life cycle model begins with an analysis, in
which the technologies employed in the project and the team load are specified. Software Development Life
Cycle models, or SDLC models, are one of the fundamental ideas in the software development process. The
SDLC is a continuous process that begins when the choice to start the project is made and ends when it is
completely removed from exploitation. No single SDLC model exists. Here, the waterfall model is used in the
suggested system. The steps of the waterfall model that are necessary for creating the suggested system project
are shown in Figure 3.1.
The phases of the proposed system's waterfall model are listed below. Requirement Analysis
A. Waterfall Model:
Figure 3.1 the most crucial and essential stage of the SDLC is the requirement analysis. With input from all
Department of Computer Engineering, ACEM, 2021–22 18 Video Surveillance Fire Detection System
stakeholders and subject matter experts (SMEs) in the sector, the senior members of the team carry it out. At this
point, planning is also done for the requirements for quality assurance and for the identification of project-related
risks. A meeting is scheduled with the client by the business analyst and project manager to obtain all the
necessary information, such as what the customer wants to construct, who will be the end user, and what the
product's goal is.
B. System Design:
System Design is the following phase that will compile all of the information on requirements, analysis, and
software project design. This phase is the result of the previous two, such as requirement collection and client
input.
C. Implementation:
The actual development process starts here, and the programming is created. Coding represents the start of
design implementation. Programming tools including compilers, interpreters, debuggers, and other similar tools
are used to generate and implement the code, and developers must adhere to the coding standards outlined by
their management.
D. Testing:
After the code is created, it is tested in comparison to the requirements to ensure that the solutions are addressing
and accumulating the demands throughout the requirements stage. Unit testing, integration testing, system
testing, and acceptability testing are carried out at this level.

1023
E. Deployment:
After the software has been certified and no problems or errors have been reported, it is deployed. The software
may then be delivered as is or with proposed improvements in the object portion depending on the assessment.
The maintenance of the software starts once it has been deployed.
F. Maintenance:
Once the client begins utilising the built systems, the real problems surface and periodic problem-solving is
required. Maintenance is the process when the developed product is given attention.

IV. PROPOSED WORK

When classifying images, convolutional neural networks have produced results with extremely high accuracy.
Convolutional neural networks are the most widely used deep learning architecture and are quite powerful.

Figure 1. Using deep CNN Early flame detection in surveillance videos.

Figure 2: Operation of CNN architecture

This built model will collect data from CCTV or surveillance footage and process it gradually in real time.
Frame by frame, the video is processed, and then the processed frames are fed into the pretrained CNN model.
This pre-trained CNN model will categorise frames into two groups in real time: one with fire and the other
without fire. This pre-trained CNN model may be set up to operate on a distant server using data from various
video surveillance systems. After processing these inputs, the pre-trained CNN model outputs a real-time
prediction on the real-time streaming data. To ensure that no frames are lost, streaming frames will be kept in
data storage. As the knowledge grows, the model will shortly
By doing this, a rich dataset is produced, and the model is trained using a large number of frames from the
dataset. As a result, the model's frame prediction accuracy will grow. Since the architecture of the monitoring
system won't need to be changed, this fire detection will be affordable. The model uses information from current
CCTVs or surveillance systems to forecast the presence of fire. This architecture is shown in Fig. 1.

1024
V. CNN
An example of an eager learner algorithm is CNN. One of the Deep Learning models is CNN. The classification
performance of CNN is excellent. The most effective algorithm for classifying images at the moment is CNN.
Since LeNet, a Deep Learning algorithm, CNN has gained enormous popularity for classifying images. Huatan
Watan Jishu Wutan AUG/2020, Volume XVI, Issue VIII, ISSN 1001-1749 Pages: 99 There were five concepts
utilised, all of which had very good accuracy, to classify handwritten digits. The convolution layer, the Relu
layer, the Pooling layer, and the fully connected layer are the four processing layers that make up a convolution
neural network. One layer's output is used as the input for another layer.
1. The convolution layer, which forms the foundation of CNN, maps several kernels onto the input data before
performing a dot product to produce a feature map. 2. The Relu layer replaces negative numbers with 0 while
leaving other values alone. 3. To minimise and produce translation invariance, select the maximum value for the
pooling layer in a limited area around these maps. The fourth layer of CNN is known as the fully connected
layer. This layer receives as its input the output of the previous three levels. A fully connected layer will classify
the photos using the training data it has collected using the provided weights.

Implementation

VI. ADVANTAGES
1) Easy to detect Fire or NOT.
2) Improve Accuracy.
3) Time Saving.
4) Easy to use.
5) It is User Friendly Application.

VII. APPLICATIONS
1) Fire Detection System.
2) Helps in Fire Fighting

VIII. CONCLUSIONS
In conclusion, a sizable archive of recordings of forest fires in a range of scene conditions has been used to
investigate an aerial-based forest fire detection approach. The chromatic and motion characteristics of a forest
fire are first extracted, and then they are rectified using a rule to highlight the burning region, in order to increase
the detection rate. Second, smoke is also extracted using our suggested algorithm to address the issue of dense
smoke that almost completely engulfs the fire. In the actual application of aerial forest fire monitoring, the
proposed System framework demonstrates its robustness with a high accuracy rate of detection and a low false
alarm rate. Making difficult and specific scene comprehension data sets for fire detection methods and in-depth
trials may be the main focus of future research.
Additionally, fire detection systems can be integrated with reasoning theories and information-hiding algorithms
to intelligently observe and authenticate the video stream and start the necessary actions in an autonomous
manner.

1025
ACKNOWLEDGMENT
Sincere appreciation to Prof. U. L. Tupe and HOD Dr. Nihar Ranjan for their assistance in resolving project-
related issues. A particular thanks to the computer wizards who created lovely libraries that are time-saving.
Additionally, we appreciate the IEEE community's assistance with the use of libraries.

REFERENCES
[1] C. Kao and S. Chang, “An Intelligent Real-Time Fire-Detection Method Based on Video Processing,” IEEE (2003 Int.)
[2] C. Ha, U. Hwang, G. Jeon, J. Cho, and J. Jeong, “Vision-based fire detection algorithm using optical flow”, 2012.
[3] C. E. Premal and S. S. Vinsley, “Image Processing Based Forest Fire Detection using YCbCr Colour Model”, 2014.
[4] N. I. Binti Zaidi, N. A. A. Binti Lokman, M. R. Bin Daud, H. Achmad, and K. A. Chia, “Fire recognition using RGB
and YCbCr color space”, 2015.
[5] K. Poobalan and S. Liew, “Fire Detection Algorithm Using Image Processing Techniques,”, December, 2015.
[6] Khan Muhammad, Jamil Ahmad, Zhihan Lv, Paolo Bellavista, Po Yang, and Sung Wook Baik, “Efficient Deep CNN-
Based Fire Detection and Localization in Video Surveillance Applications”, 2018,IEEE.
[7] Oxsy Giandi, Riyanarto Sarno “Prototype of Fire Symptom Detection System”, ©2018 IEEE.
[8] Sneha Wilson, Shyni P Varghese, Nikhil G A, Manolekshmi I, Raji P G, “A Comprehensive Study on Fire Detection”,
(2018). [9] Jiang Feng, Yang Feng, “Design and experimental research of video detection system for ship fire”, ©2019
IEEE. [10] Ke Chen,Yanying Cheng, “Research on Image Fire Detection Based on Support Vector
Machine”,(2020),IEEE [11] UANG HONGYU1, KUANG PING1, LI FAN1, SHI HUAXIN1, “An Improved
MultiScale Fire Detection Method Based On Convolutional Neural Network”, © 2020 IEEE.

1026