Fire and Smoke Detection for Wildfire using YOLOV5 Algorithm

Computer Science & Engineering: An International Journal (CSEIJ), Vol 15, No 1, February 2025
DOI:10.5121/cseij.2025.15106 45
FIRE AND SMOKE DETECTION FOR
WILDFIRE USING YOLOV5 ALGORITHM
Bharathi K, Jawahar Jonathan K, Satyabhama, Shobharani C N,
Sridevi, Bhagyashree K Itagi
Department of Computer Science and Engineering, Acharya Institute of
Technology, Bangalore, India
ABSTRACT
This research paper presents a realtime fire and smoke detection system using the YOLOv5
object detection algorithm. The system aims to detect fire and smoke in images and video
streams captured by a camera in real-time, without the need for any preprocessing or
manual intervention. The proposed system uses the YOLOv5 algorithm to detect the fire
and smoke regions in the input images and videos. The YOLOv5 model is trained on a
dataset of annotated images to recognize fire and smoke patterns accurately. The proposed
system has been tested on different datasets and has achieved high accuracy and precision
in detecting fire and smoke in real-time. The experimental results demonstrate that the
proposed system is robust and efficient, and it can detect fire and smoke in real-time with
high accuracy and low latency. The proposed system can be used in various applications,
such as early warning systems, fire safety, and disaster management. It can also be
integrated with the CCTV network directly.The testing findings show how reliable and
effective the suggested system is, as well as how accurately and quickly it can detect fire
and smoke in real time. Numerous applications, including early warning systems, fire
safety, and disaster management, are suitable for the suggested system. Direct integration
with the CCTV network is another option.
1. INTRODUCTION
The development of human civilization has been greatly aided by fire. Conversely, it is one of the
world's worst calamities, with a staggering loss of life and property worldwide. Early fire
detection can save lives and property by raising awareness of the danger and averting these
tragedies. Usually, smoke is released before an object catches fire and starts to burn. Smoke can
thus be used as a signal to spot fires early. In India, the number of fire events has increased within
the last several years. An analysis by Accidental Deaths and Suicides in India (ADSI), kept up to
date by the National Crime Records Bureau, shows that in the five years between 2016 and 2020,
fire-related incidents claimed the lives of 35 individuals on average every day.
Division of Records. Such incidents demolish properties valued at millions of rupees. Early fire
detection without false alarms is essential to reducing such tragedies. Numerous autonomous fire
detection technologies are being developed and are commonly employed in real-world situations
because traditional detection systems proved to be unsuccessful in outside contexts.
The advent of computer vision-based detection systems may be attributed to the rise of artificial
intelligence and its offshoots. These vision based detection systems provide greater surveillance
coverage and overcome size and location constraints. They also provide less false alarms,
quicker detection, and less need for human intervention. When it comes to computer vision

46
technology, researchers have worked very hard to overcome the problems of false detection and
system complexity.
Therefore, we want to apply this development for the good of society by creating a more
straightforward system for detecting fire and smoke using convolutional neural networks. This
system will be able to analyze a picture and determine whether or not it includes smoke or fire.
Enhancement of data has been used to add slightly altered copies of previously existing data to
increase the amount of data. These copies are derived from original images that have undergone
some minor geometric transformations (such as flipping, translation, rotation, or the addition of
noise) to increase the diversity of the training set. or using alreadyexisting data to create new,
synthetic data. When the machine learning model is being trained, it serves as a regularizer and
lessens overfitting.
2. LITERATURE SURVEY
[1] The accuracy, efficiency, and speed of flame and smoke detection algorithms have
been addressed by recent advances. Low precision and lengthy processing are common
problems with traditional approaches, especially for tiny targets like smoke plumes or flames.
Researchers have used YOLOv5s, a state-of-the-art object identification model, into their
algorithms to get around these restrictions. Accuracy is increased and real-time detection
capabilities are made possible by this integration. Furthermore, the backbone network's
technologies like ODConvBS improve feature extraction and make it possible to identify faint
patterns of smoke and flame. The spatial interaction and feature extraction efficiency are further
improved by methods such as Shuffle Attention (SA) module and recursive gated convolutions
(Gnconv). By taking into account the vector angles between regressions, specialized loss
functions like SIOU speed up model training and produce faster and more accurate detections.
[2] The literature study included in the material supplied highlights how crucial it is to
identify wildfire smoke early in order to reduce casualties and property damage. It emphasizes
the use of camera networks for prompt detection, highlighting the significance of clever video
smoke detection algorithms and cost-effective camera placement techniques. Two modules
make up the suggested effective video smoke detection framework: one handles wildfire
camera location optimization as a binary integer programming issue, while the other uses dense
optical flow and local binary patterns for smoke detection in video frames. The framework
makes use of deep learning methods such as MobileNetV2 in conjunction with physics-based
features to improve detection accuracy while preserving computational and storage economy.
The paper also discusses the computational difficulties that deep learning-based methods face
in situations with limited resources, arguing in favor of simple solutions ideal for embedded
local applications as wildfire camera systems.
[3] Research on improving fire detection systems has increased significantly in recent
years, especially in response to the growing threat of forest fires. The failure of conventional
sensor-based techniques has led to research into computer vision algorithms. There are benefits
to vision-based methods in terms of precision, coverage, and instantaneous response. By
extracting complex information from photos, the development of Deep Learning— particularly
convolutional neural networks like YOLOv5 and U-Net—has revolutionized the detection of
fires. In order to overcome issues like false alarms and lengthy inference times, this article
suggests a novel architecture for fire detection and segmentation that blends YOLOv5 with U-
Net.The suggested architecture integrates YOLOv5 with U-Net in an attempt to overcome the
obstacles that persist in obtaining high accuracy and speed despite developments. This
integration offers enhanced performance in detecting and segmenting fires, including small-size
fire objects, while decreasing false alarms. These results highlight the increasing interest in

47
deep learning (DL) approaches to fire-related problems and point to a viable path for further
fire detection and monitoring system development.
[4] The two primary strategies for smoke detection included in the literature review are
those based on computer vision and those based on deep learning. Computer vision techniques
make use of classifiers like rule-based systems, SVMs, and shallow neural networks, as well as
characteristics including color, motion, texture, form, and energy. For smoke detection tasks,
deep learning approaches, on the other hand, use CNN architectures such as AlexNet and Faster
R-CNN. In order to achieve equivalent detection accuracy but requiring less computing and
storage, the suggested approach combines lightweight CNNs with conventional feature
extraction techniques. The methodology solves the shortcomings of previous methods and can
be used in resource-constrained wildfire monitoring camera systems by integrating
mobilenetV2 CNN for classification with local binary patterns (LBP) and dense optical flow for
feature extraction.
3. METHODOLOGY
The process for detecting fire and smoke starts with recognizing the special difficulties that come
with being in a region that is prone to wildfires and real-time and accurate monitoring. Gathering
various datasets of videos and photos from wildfires and annotating them is the process of
collecting data for model training. YOLOv5 is chosen due to its precision and instantaneous
response times; it is trained and refined to efficiently identify fire and smoke occurrences.
Deployment in wildfire-prone areas is made easier by integration with a real-time surveillance
platform, which also ensures compatibility and extensive testing across a range of situations.
Testing and validation studies are used in assessment to evaluate performance, and continuing
maintenance and enhancements guarantee dependability and efficacy. The goal of this
methodical approach is to improve safety protocols in regions susceptible to wildfires by
tackling the intricacies of smoke and fire detection.
Fig.3.1 Methology of fire and smoke detection.
• Problem Understanding and Requirement Analysis:
Recognize the unique difficulties and specifications related to smoke and fire detection in
regions vulnerable to wildfires. Determine the necessity of robustness, high precision, and real-
time monitoring for the detection of fire and smoke incidents.

48
• Gathering and Preparing Data
Compile a varied dataset with pictures and videos of wildfire episodes that show different
situations, the weather, and the time of day. To produce ground truth data for training the
YOLOv5 model, preprocess the dataset by categorizing the areas of fire and smoke in the
photos and videos.
• Model Selection and Training:
Due to its accuracy and real-time performance, choose YOLOv5 as the object identification
model.
To modify the YOLOv5 model to detect fire and smoke incidents, train it on the annotated
dataset using transfer learning techniques. To boost the variety of the training set and enhance
the model's capacity for generalization, apply data augmentation approaches.
• Model Enhancement:
In order to retain real-time performance and get high detection accuracy, adjust the YOLOv5
model's parameters and hyperparameters. Optimize the model architecture and training
procedure to enhance the model's capacity to identify fire and smoke occurrences in wildfire
situations.
 Integration and Deployment:
Combine a real-time surveillance platform appropriate for detecting wildfires with
the trained YOLOv5 model.
Install the integrated system where wildfires are likely to occur, making that it is
compatible with the infrastructure and monitoring systems already in place.
 DATASET ACQUISITION
The dataset used for training and validation had a major impact on the deep learning
model's accuracy level. From open-access sites like GitHub and Roboflow, we collected
fire photos that showed a variety of situations (form, color, size, inside and outdoor setting).
To verify the system's dependability and efficacy in identifying fire and smoke incidents, test it
in a variety of lighting and environmental scenarios.
• Assessment and Confirmation
Analyze the deployed system's performance in terms of processing speed, dependability, false
positives, and detection accuracy
Conduct validation studies and field testing to verify the system's efficacy in identifying fire
and smoke occurrences in wildfire zones.
• Constant Maintenance and Improvement
Track the functioning of the implemented system and gather input from users and stakeholders.
Update and enhance the system often in response to user input and fresh advancements in

49
object and wildfire detection technologies. To guarantee the system's continued functioning and
performance, undertake routine maintenance and upgrades.
4. TESTING
Fig.4.1 Wildfire smoke detection operation using deep learning.
There were 2462 fire photos in our fire image dataset.
 DATA PREPROCESSING
Roboflow was used to store the dataset. Pictures without annotations were eliminated. Then
create a dataset by combining them. The roboflow framework was used to preprocess the
photos . There was no augmentation used by us. They were extended to 416 × 416 after being
resized. Lastly, split them into 50% training and 50% validation.Next, we export the dataset in
Pytorch format (YOLO v5), and it creates an API connection that we immediately utilized in
our model.

50
5. IMPLEMENTATION
Fig.5.1 Overall design of the proposed system.
 Data Preparation:
Compile a varied dataset with pictures and videos showing smoke and fire incidents in different
settings. Using programs like Roboflow or Makesense, annotate the dataset to identify the fire
and smoke zones of interest.Divide the dataset into subgroups for testing, validation, and
training.
 Model Training for YOLOv5:
To train the YOLOv5 model on the annotated dataset, use Google Colab.To ensure effective
training, set up a Google Colab laptop with GPU capability.For training, use the YOLOv5
repository from GitHub and adhere to the guidelines given.Utilizing the provided dataset, fine-
tune the YOLOv5 model to maximize detection accuracy.
 Validation:
Using the dataset's validation subset, validate the trained YOLOv5 model. Analyze the model's
performance measures, including mean average precision (mAP), recall, and accuracy.To
enhance performance, modify the model's hyperparameters or training settings as necessary.
• Testing:
To evaluate the model's performance in the actual world, test it using the testing subset of the
dataset.Use inference on unseen photos and videos to find occurrences of smoke and
fire.Analyze the model's processing speed, false positive rate, and accuracy of detection.

51
• User Interface Development:
Provide an intuitive user interface so that users can communicate with the smoke and fire
detection system.Create an interface that can display detection results and take in photos or
videos as input.Provide tools for monitoring detection confidence ratings, changing detection
thresholds, and handling alarms.
• Integration with Makesense and Roboflow
Use Makesense and Roboflow to integrate data preparation and annotation into the training
workflow.For dataset maintenance, augmentation, and format conversion, use Roboflow.To
improve annotations and confirm label correctness, use Makesense.
• User Examination:
Allow testers to experience the operational smoke and fire detection system.Give the testers
preestablished test scenarios to work with while they are doing the tests.Get tester input about
the performance, efficacy, and usability of the system.Utilize the input to pinpoint areas that
need work and make iterative changes to the implementation.include tools for controlling
warnings, monitoring detection confidence ratings, and changing detection thresholds.
• Refinement and Optimization
Adjust the implementation iteratively in response to user testing input.Optimize the YOLOv5
model even further to boost robustness and detection accuracy.For real-time deployment,
optimize the system's performance and resource use.
• Deployment:
Implemented smoke and fire detection system should be put to use in an actual setting.Make sure
there is a smooth interface with any current networks or surveillance systems.Keep an eye on the
system's dependability and performance in use, and take appropriate action when problems
occur.
6. RESULTS
Fig.6.1.Confusion Matrix

52
A table that lists a classification model's performance in summary form is called a confusion
matrix.
It displays the number of forecasts that are true positives (TP), true negatives (TN), false
positives (FP), and false negatives (FN). TP:
Properly anticipated positive cases. TN: Accurately forecasted adverse events. FP: False alarms
caused by incorrectly projected positive cases. FN: Negative incidences that were incorrectly
anticipated (missed detections).
Fig.6.2.F1 and Confidence Curves
The model's performance in the event of a class imbalance is assessed using the F1 score, which
is the harmonic mean of accuracy and recall. Plotting the F1 score against various forecast
confidence criteria is known as the F1 curve. It aids in figuring out the ideal cutoff point that
strikes a balance between recall and accuracy. The F1 score's variation with the confidence
threshold is depicted by the curve.
Fig.6.3.Precision and Confidence Curves
The precision of a prediction is the percentage of accurate positive forecasts among all positive
predictions. Plotting precision versus various confidence criteria is done using the precision
curve. It facilitates the visualization of the model's accuracy at various degrees of confidence.
More accurate detections result from fewer false positive predictions, which are shown by a
greater precision.

53
Fig.6.4.Accuracy and Recall Curves
At various confidence levels, accuracy and recall are traded off. These curves show this
tradeoff. Plotting recall is done on the xaxis and precision on the y-axis.
The graph illustrates how, with fluctuating confidence levels, accuracy changes when recall
varies or vice versa.
It aids in evaluating how well the model captures all positive occurrences while producing
accurate positive forecasts.
Fig.6.5.Confidence curves and recall
recall calculates the percentage of accurate positive predictions among all real positive
occurrences. Plotting recollection versus various confidence criteria is done using the recall
curve. It facilitates the visualization of the model's recall at various confidence levels. More
accurately detected positive cases are indicated by a better recall, which also means fewer
incorrect negative predictions.
7. DISCUSSION
The YOLOv5 object identification method is used in the research article to propose a realtime
fire and smoke detection system. The introduction emphasizes how important it is to identify
fires early since they offer serious hazards. It is decided that conventional detection techniques
are insufficient, which leads to the investigation of automated, higher level systems. The
suggested solution uses YOLOv5 architecture and deep learning approaches to try to
overcome these drawbacks. It highlights how crucial improved accuracy, on-the-spot
monitoring, and automation are to the detection process. High detection accuracy, the ability

54
to monitor in real-time, and easy connection with current surveillance systems are important
advantages. The research is driven by the pressing need to enhance fire detection systems,
particularly in settings where conventional approaches are inadequate. The project's scope
includes all phases, from development and research to assessment and integration. The study
presents a clear architectural foundation for the proposed system by identifying modules and
using UML diagrams. The efficacy of the YOLOv5-based smoke and fire detection system is
emphasized in the conclusion, underscoring its capacity to improve safety precautions and
lessen the risks connected with fire occurrences.
8. CONCLUSION
In summary, the YOLOv5-developed fire and smoke detection system has shown to be an
efficient and successful means of accomplishing the project's goals and functional requirements.
YOLOv5 has been chosen as the foundational technology for object detection. Allowed the
system to precisely identify incidents of fire and smoke in pictures and videos in real time,
achieving the project's main goal. The accuracy and performance of the system have been greatly
enhanced by YOLOv5's cutting-edge object identification capabilities and architecture, which
guarantee dependable detection even in challenging and dynamic contexts.
Additionally, the monitoring interface, alert notification module, and surveillance system
integration with YOLOv5 have made things easier.Smooth cooperation and communication,
enabling thorough observation and prompt reaction to fire and smoke occurrences. The
technology has improved operating efficiency and reduced the danger of possible disasters by
automating the detection process, setting off alarms, and notifying the appropriate authorities.
Overall, it has been quite successful in meeting the project's goals and functional requirements to
use YOLOv5 as the primary technology for smoke and fire detection. Due to its precision,
efficiency in real-time, and integration potential, the system has established itself as a dependable
means of improving safety protocols and lowering the risks connected with fire-related incidents.
REFERENCES
[1] Jingrun Ma , Zhengwei Zhang , Weien Xiao, Xinlei Zhang, And Shaozhang Xiao. “flame and smoke
detection algorithm based on odconvbsyolov5s”
[2] Sushma C*1, Hemanth Kumar BN*2 “placements and intelligent smoke detection algorithm
in the wildfire- monitoring cameras”
[3] Wided Souidene Mseddi L2TI, Institut Galilée Université Sorbonne Paris Nord Villetaneuse,
France sercom Laboratory University of Carthage Tunis, Tunisia. “fire detection and segmentation
using yolov5 and unet”
[4] Jie Shi , (Graduate Student Member, IEEE), Wei Wang,(Graduate Student Member,
IEEE),Yuanqi Gao ,(Graduate Student Member, IEEE), and Nanpeng Yu ,(Senior
Member, IEEE)“optimal placement and intellegent smoke detection algorithm for wildfire-
monitoring cameras”

Fire and Smoke Detection for Wildfire using YOLOV5 Algorithm

More Related Content

Similar to Fire and Smoke Detection for Wildfire using YOLOV5 Algorithm (20)

More from CSEIJJournal (20)

Recently uploaded (20)

Fire and Smoke Detection for Wildfire using YOLOV5 Algorithm