Thermal Imaging with AI for Real‑Time Human Detection in Fire

Emergencies

Savitha P., Dusyanth N. P., Hemanath P. and Mohammed Naveeth P.

Department of Computer Science and Engineering, Nandha Engineering College, Erode, Tamil Nadu, India

Keywords: Infrared Thermal Camera, Convolutional Neural Network Evacuation During a Fire Foggy Fire Scene

Firefighter Safety Human Detection.

Abstract: Forest fires are a significant ecological risk because their smoke is an initial warning of risk but initially smoke

appears in small hardly noticeable quantities it is hard to identify smoke with pixel-based parameters because

the environment is constantly evolving and the dispersion characteristics of smoke are unpredictable this paper

presents an innovative framework that minimizes the susceptibility of many yolo detection methods to minor

visual interference it further analyzes and contrasts the detection capability and processing time of various

versions of yolo ie yolov3 and yolov5the present research scrutinizes a recent human detection scheme for

emergency evacuation in smoke-producing low-visibility fire environment utilizing a thermal imaging camera

tic that meets national fire protection associations 1801 standards we take low-wavelength infrared LWIR

pictures and utilize the yolov4 deep-learning algorithm to recognize objects in real-time while running at 301

frames per second fps the model trained on an Nvidia GeForce 2070 GPU can detect people in smoke

environments with more than 95 accuracy this quick detection capability allows for immediate data

transmission to command centers to facilitate timely rescue and protection of firemen prior to reaching risky

situations.

1 INTRODUCTION

Smoke is a major workplace safety hazard that must

be prevented and promptly addressed in the event of

an emergency. Worker safety, firefighter support, and

preventing property damage all depend on a well-

planned fire evacuation strategy. During an

emergency evacuation, we will use a thermal camera

and the YOLOv4 model to locate people in a smoke-

filled fire scene in accordance with NFPA 1801 rules

to increase fire detection and safety procedures. At a

fire scene, the two greatest dangers are heat and

smoke, with smoke being the most dangerous. It

causes zero vision during firefighter rescues or

evacuations and poses a serious risk of suffocation,

which can be lethal. Humans can only move at 0.5

m/s, yet smoke spreads quickly during a fire at 3–5

m/s. When smoke spreads, fire follows, so finding

people quickly and getting them to safety is essential

for life. LADAR, 3D Laser Scanning, Ultrasonic

Sensors, and infrared thermal cameras are common

solutions for the challenging problem of detecting

people in dense smoke. To provide clear visibility in

a smoke-filled fire scenario, we suggest combining

the YOLOv4 with an infrared thermal camera that

complies with NFPA 1801 criteria. A single GPU

powers an AI-powered human detection system that

uses a Convolutional Neural Network (CNN) to

recognize individuals in a smoke-filled environment.

It helps with emergency evacuations by transmitting

real-time updates to a central control room. Similar

studies on object detection models (Section 2),

dataset and pre-processing information (Section 3),

and the suggested methodology using the YOLOv4

model (Section 4) are all included in this study.

concludes with a discussion of human detection after

presenting experimental data and performance

measures (Section 5).

2 RELATED WORKS

Sai, Liao, and Yuan suggest employing a Deep

Learning model and a thermal imaging camera tic in

the evacuation of people from emergencies involving

smoke-filled fire low-visibility their method

identifies persons in such conditions with over 95

accuracy using the yolov4 model and images from

608

P., S., P., D. N., P., H. and P., M. N.

Thermal Imaging with AI for Real-Time Human Detection in Fire Emergencies.

DOI: 10.5220/0013887200004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 2, pages

608-613

ISBN: 978-989-758-777-1

low-wavelength infrared LWIR. Do Truong and le

suggest a fusion technique utilizing thermal and

infrared imaging to recognize people in real time in

fire conditions where smoke obscures visibility their

technique achieves a mean average precision map of

95 by fusing information from multiple cameras and

processing it with a light-deep neural network. Ai-

powered early fire detection gadget that recognizes

early smoke signs of wildfires through real-time

picture processing and high-definition panoramic

cameras wildfire discovery and response time are

dramatically minimized by solar-powered technology

which is also linked to emergency services. The solar-

powered aid-fire system developed by zhang wang

zeng wu Huang and Xiao employs cloud servers IOT

sensors and ai engines to instantly recognize complex

building fire data2 the system was tested in a full-

scale fire test room and indicated fire growth and

spread correctly with a relative inaccuracy of less

than 15 10 uav technology is transforming emergency

response and fire protection because it enables 3d

mapping in real time and hotspot identification

through the integration of thermal images and ai-

based algorithms drones enhance situational

awareness and allow quicker and more accurate fire

detectio. These drones employ thermal images and

semantic segmentation to efficiently track and

monitor forest fires using deep models such as mask

r-cnn and yolo versions this technique significantly

boosts the accuracy of detection and continuous

monitoring of fires ensuring a more proactive and

efficient emergency response.

3 EXPERIMENT

METHODOLOGY

3.1 Data Collection

The thermal imaging apparatus was selected because

it complies with NFPA 1801 standards, specifically

concerning temperature sensitivity, resolution, and

spectral range. Additionally, the device features an

uncooled microbolometer, which enhances imaging

capabilities and makes it suitable for use in fire

emergencies. To generate more training data, we use

this thermal imaging camera (TIC) to capture images

of various body poses, such as squatting, lying down,

and falling, from every aspect (360°). This makes it

easier to imagine realistic situations in which people

might want assistance during a smoke-filled escape.

The human body temperature in these images is equal

to a Gray Level (GL) of 105, per previous studies on

Thermal making-based human detection in fire

scenarios.

3.2 Training Using Thermal Datasets

Deep learning is highly dependent on the quantity and

quality diversity velocity and authenticity of big data

its limitation is that we do not have large publicly

available thermal image data sets we compensate with

other data similar to the thermal images we recorded

and publicly available data sets like indoor people

dataset in Kaggle and pedestrian thermal images with

more quantity and diversity of data the method

enhances the precision and overall model

performance in thermal image analysis. Figure 1

shows the Thermal Imaging Dataset.

Figure 1: Thermal Imaging Dataset.

3.3 YOLOv3

Joseph Redmon and Ali Farhadi proposed the third

generation yolo you only look once network in 2018

the models inference time is 22 milliseconds and the

mean average accuracy is 282 dimension clusters are

employed to address the ground-truth bounding box

prediction problem in anchor boxes features of

yolov3 are extremely minimalistic however lower

confidence level because its classification layer uses

logistic regression instead of softmax the more

confident the object is the more likely it is to appear

in a given grid cell darknet-53 as the base model

utilized a denser convolution backbone than yolov4

utilized neck layer information through a cross-stage

partial csp network since a detection model yolov3 is

still an excellent one for everyone but the

computationally most demanding applications

Bhattarai and Martinez-Ramon used proficiently

deep models such as yolov4 to spot people in

Thermal Imaging with AI for Real-Time Human Detection in Fire Emergencies

609

thermograms with enhanced intersection-over-union

IOU accuracy Redmon was able to demonstrate in

spite of its deficits that yolov3 is superior and

improved compared to ssd and even its earlier form

yolov2 light yolo variants are also investigated to

optimize object detection methods to utilize for real-

time fire prevention and emergency evacuation

systems along with the above.

3.4 YOLOv4 Model

Even though thermal images tend to have low

resolution and small objects will only occupy

approximately 50 pixels studies demonstrate that yolo

darknet 20 is effective in detecting regular and small

far-range thermal objects the deep architecture of

darknet renders thermal imaging more relevant in a

broad range of conditions which qualifies it as an

appropriate choice for fire emergency response we

utilize the yolov4 object detector which is a cnn-

based object detector for our research as the most

advanced real-time object detection model available

in the market at the time of writing in 2020 it is

suitable for detecting humans in risky environments

such as smoke-filled fire environments a single

Nvidia Geforce 2080gpu is utilized to train the model

which is employed as a real-time head neck and

backbone detector the backbone network is

responsible for feature extraction on various scales

and the neck equipped with spatial pyramid pooling

spp serves to reduce model parameters and increase

training efficiency the head which is founded on

yolov3 is meant for object localization and one-stage

classification.

3.5 YOLOv5

yolov5 was developed to allow for quicker real-time

processing and was initially slated for release in May

2020 it is intended to identify one or more objects

within an image immediately which makes it

extremely efficient for real-time use the model

consists of three primary parts they are the backbone

neck and head. which operate in conjunction with

each other to enhance efficiency and accuracy the

spine is the beginning tasked with detecting important

features in an image accurately irrespective of an

object’s location or size yolov5 uses a path

aggregation network planet which aids in feature

combination and enhances object detection yolo layer

does detection and classification giving each object

features such as its coordinates x-axis, y-axis width

height bounding box and an confidence score

signifying the probability of its existence the model

utilizes the intersection over union IOU method to

limit errors and prevent duplication of detection

choosing the most precise bounding box improved by

these advancements yolov5 is still a light yet very

credible real-time object detection tool.

Figure 2: YOLOv5 Architecture.

3.6 Optimal Model Selection

Three training datasets are employed: our own 360-

degree thermal images captured using the Fluke

Ti300+, the Kaggle AAU TIR image dataset, and

FLIR ADAS. We manually annotate with bounding

boxes the people in these images, as illustrated in

Figure 2. As it provides the best visibility in infrared

imaging, all of the images are within the long-wave

infrared (LWIR) band (8–15μm). With a 1,000 burn-

in time and learning accuracy of 0.0014, the model is

trained for 5,000 epochs. The training was completed

in less than a day on our machine with default

hyperparameters due to the high contrast of thermal

human images and the ability of the model to extract

good features through 53 network layers.

4 EXPERIMENT RESULT

4.1 Finding the Ground Truth for

Objects that Are Partially

Occluded

We aim to recognize and enumerate an individual as

a unique person in our case even though they may be

partially hidden two individuals are counted as one

object if they are extremely close to each other and

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

610

one of them is more than 50 covered by the other five

individuals are employed as ground truth GT in the

left image however since the three individuals on the

right-hand side are so close together that they cannot

be seen as individual units the right photo only

accounts for three people as GT.

Figure 3: Fire Emergency Conditions.

ground truth for partially occluded objects needs

to be known in order to achieve correct object

detection since only a fragment of an object’s features

might be visible when it is being covered by another

precise localization is challenging even when

portions of the piece are concealed ground truth

annotation must be present in this issue in order to

include the pieces intended shape and location one

method is to approximate the overall bounding box of

the item from its visible parts and from prior

knowledge of its typical structure additionally

methods such as segmentation masks and occlusion-

aware labeling aid in optimizing the annotations by

distinguishing between completely visible areas and

partially occluded areas these techniques assist the

model in identifying and classifying objects even with

occlusions more accurately by improving the quality

of the training data. Figure 3 shows the Fire

Emergency Conditions.

4.2 Measures of Detection Accuracy,

Precision, and Recall

when we expect the bounding box to closely resemble

the real object the detection is called a true positive

this is usually determined using the intersection over

union iou threshold of 50 or higher FN is produced if

the iou drops below this level or if the model was

unable to identify an object, on the other hand, an fp

happens when the model labels an item that isn’t there

we may gain a better understanding of the models

capacity for precise item identification and

classification by contrasting precision and recall with

you-based accuracy in order to improve object

detection algorithms for practical uses these metrics

are essential.

Figure 4: PASCAL VOC Metric for Precision and

Confusion Curve.

4.3 PR Curve, Precision, and Recall in

Test Datasets

we concluded that the model weights after 4000

epochs were the best option as they exhibited the

minimum training loss with outstanding accuracy and

accurate object localization on the test set our model

runs very well with a precision and recall of over 97

also even when viewed from various angles and

human positions figure 4 captures a roc score of 98

this high accuracy is attributed to the robust 53-layer

deep cnn which provides precise object detection with

bounding boxes always meeting or surpassing 50

IOU. Table 1 shows the Augmentation Dataset

Image. Table 2 shows the Comparison of Volov4 And

Volov5.

Table 1: Augmentation Dataset Image.

Fire

Condition

Detection

accuracy

(%)

Response

Time(s)

False

Alarm

rate

(%)

Low

Smoke

98 1.2 2

Medium

Smoke

92 1.5 5

High

Smoke

85 2.0 10

Obstructed

View

70 3.0 18

Thermal Imaging with AI for Real-Time Human Detection in Fire Emergencies

611

Table 2: Comparison of YOLOv4 and YOLOv5.

Model

mAP@50

(Accuracy

Inference

Speed

(FPS)

Model

Size

Training

Time

YOLOv4 85–90%

~45 FPS

(GPU)

Large

(244

MB)

Longer

YOLOv5 90–95%

~60 FPS

(GPU)

Small

(14

MB)

Faster

5 CONCLUSIONS

We introduce yolov4 a deep learning model that uses

an NFPA 1801 compliant thermal imaging camera to

show that it can recognize humans in dense smoke

even in high-temperature low-visibility fire situations

greyscale human forms are enhanced by the camera’s

exceptional resolution and low-temperature

sensitivity on an Nvidia GeForce 2070 GPU the

model converges in 4000 epochs with MS coco pre-

trained weights attaining over 95 accuracies at a 50

IOU it is useful for search and evacuation

surveillance because it can identify people who are

squatting standing sitting and lying down even when

there is 50 occlusion at 301 fps real-time detection

operates by mapping settings recognizing people and

locating heat sources to increase firefighter safety

future integration with robotic systems could improve

search and rescue operations

6 FUTURE WORK

Future developments in AI-powered thermal imaging

for real-time human detection during fire situations

can concentrate on a few important areas. First, real-

time performance can be improved by tailoring deep

learning models for edge devices like drones and

Internet of Things sensors. Methods like lightweight

architectures (like YOLOv8-Nano) and model

quantization (like Tenso RT, and ONNX) can lower

computational load without sacrificing accuracy.

Furthermore, combining thermal cameras with RGB

and LiDAR sensors to integrate multi-spectral

imaging might increase the accuracy of detection in

low-visibility situations brought on by smoke or

intense heat. Incorporating Transformer-based

models (such as DETR and YOLO-World) can

further improve AI capabilities by enabling context-

aware human and fire danger detection, and real-time

tracking algorithms can help track human movement

within fire zones. Model resilience can be improved

by adding more various fire situations to the dataset,

such as varying temperatures, smoke concentrations,

and human postures. Overcoming data constraints

can also be aided by the creation of synthetic data

using GANs (Generative Adversarial Networks).

REFERENCES

A. Filonenko, D. C. Hernández, and K. -H. Jo, "Real-time

smoke detection for surveillance," 2015 IEEE 13th

International Conference on Industrial Informatics

(INDIN), Cambridge, UK, 2015, pp. 568-571, doi:

10.1109/INDIN.2015.7281796.

A. Filonenko, D. C. Hernández, Wahyono, and K. -H. Jo,

"Smoke detection for surveillance cameras based on

color, motion, and shape," 2016 IEEE 14th

International Conference on Industrial Informatics

(INDIN), Poitiers, France, 2016, pp. 182-185, doi:

10.1109/INDIN.2016.7819155.

A. Filonenko, D. C. Hernández and K. -H. Jo, "Fast Smoke

Detection for Video Surveillance Using CUDA," in

IEEE Transactions on Industrial Informatics, vol. 14,

no. 2, pp. 725-733, Feb. 2018, doi:

10.1109/TII.2017.2757457. keywords: {Cameras;

Image color analysis;Graphics processing

units;Sensors;Video}

A. Mariam, M. Mushtaq and M. M. Iqbal, "Real-Time

Detection, Recognition, and Surveillance using

Drones," 2022 International Conference on Emerging

Trends in Electrical, Control, and Telecommunication

Engineering (ETECTE), Lahore, Pakistan, 2022, pp. 1-

5, doi: 10.1109/ETECTE55893.2022.10007285.

Bhavnagarwala and A. Bhavnagarwala, "A Novel

Approach to Toxic Gas Detection using an IoT Device

and Deep Neural Networks," 2020 IEEE MIT

Undergraduate Research Technology Conference

(URTC), Cambridge, MA, USA, 2020, pp. 1-4, doi:

10.1109/URTC51696.2020.9668871.

D. Kinaneva, G. Hristov, G. Georgiev, P. Kyuchukov and

P. Zahariev, "An artificial intelligence approach to real-

time automatic smoke detection by unmanned aerial

vehicles and forest observation systems," 2020

International Conference on Biomedical Innovations

and Applications (BIA), Varna, Bulgaria, 2020, pp.

133-138, doi: 10.1109/BIA50171.2020.9244498.

D. K. Dewangan and G. P. Gupta, "Explainable AI and

YOLOv8-based Framework for Indoor Fire and Smoke

Detection," 2024 IEEE International Conference on

Information Technology, Electronics and Intelligent

Communication Systems (ICITEICS), Bangalore,

India,

2024, pp. 1- 6, doi:10.1109/ICITEICS61368.2024.106

24874.

G. M. A, A. Sivanesan, G. Yaswanth and B. Madhav,

"ECOGUARD: Forest Fire Detection System using

Deep Learning Enhanced with Elastic Weight

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

612

Consolidation," 2024 IEEE International Conference

on Big Data & Machine Learning (ICBDML), Bhopal,

India, 2024, pp. 115-120, doi:

10.1109/ICBDML60909.2024.10577367. [7].

H. Prasad, A. Singh, J. Thakur, C. Choudhary and N. Vyas,

"Artificial Intelligence Based Fire and Smoke

Detection and Security Control System," 2023

International Conference on Network, Multimedia and

Information Technology (NMITCON), Bengaluru,

India, 2023, pp. 01- 06, doi: 10.1109/NMITCON5819

6.2023.10276198.

H. Kulhandjian, A. Davis, L. Leong, M. Bendot and M.

Kulhandjian, "AI-based Human Detection and

Localization in Heavy Smoke using Radar and IR

Camera," 2023 IEEE Radar Conference

(RadarConf23), San Antonio, TX, USA, 2023, pp. 1-6,

doi: 10.1109/RadarConf2351548.2023.10149735

R. Donida Labati, A. Genovese, V. Piuri and F. Scotti,

"Wildfire Smoke Detection Using Computational

Intelligence Techniques Enhanced with Synthetic

Smoke Plume Generation," in IEEE Transactions on

Systems, Man, and Cybernetics: Systems, vol. 43, no.

4, pp. 1003- 1012, July 2013, doi: 10.1109/TSMCA.2

012.2224335.

Thermal Imaging with AI for Real-Time Human Detection in Fire Emergencies

613