HUMAN DETECTION USING CNN IN

FLOOD AREAS

HariHaran L Karthickvikraman M Akash A

[email protected] [email protected] [email protected]

ABSTRACT

In the aftermath of natural disasters and humanitarian crises, the swift and accurate identification of individuals trapped
or in need of assistance is paramount for effective rescue operations. Leveraging the capabilities of artificial intelligence
(AI), this project proposes a novel approach to human detection during disaster situations. Through the fusion of com -
puter vision and machine learning techniques, the system aims to autonomously detect and locate individuals amidst
chaotic environments, such as collapsed buildings or debrisladen landscapes. Key components of the project include fea -
ture extraction, object detection, and semantic segmentation algorithms, which collectively enable the system to identify
human presence with high accuracy while minimizing false positives. Additionally, the integration of sensor data and
geospatial information enhances the system's situational awareness and aids in prioritizing rescue efforts. By incorporat-
ing real-time data processing and analysis, the system provides rescuers with timely information about the location and
status of individuals in need of assistance, potentially improving overall response times and outcomes. Through the seam -
less integration of AI technologies and disaster response frameworks, this project seeks to significantly enhance the ef -
fectiveness and efficiency of rescue operations in the face of natural disasters and humanitarian crises. By autonomously
identifying and locating individuals amidst chaotic environments, the proposed system aims to provide invaluable support
to rescue teams, ultimately saving lives and mitigating the impact of disasters on affected communities.

I. INTRODUCTION

In an era characterized by increasing frequency and severity of natural disasters, effective disaster management and re -
sponse strategies are paramount for minimizing human casualties and mitigating the impact on communities. One of the
key challenges in disaster response is swiftly identifying and locating individuals in affected areas, particularly in hazard -
ous or inaccessible environments where traditional search and rescue methods may be limited. To address this challenge,
the integration of artificial intelligence (AI) with edge computing platforms like the Raspberry Pi presents a promising
solution. This project focuses on developing an AI-driven human detection system using deep learning algorithms, such
as convolutional neural networks (CNNs), deployed on the Raspberry Pi. The integration of CNN algorithms with
OpenCV and TensorFlow provides a powerful framework for analyzing real-time imagery captured by drones, satellites,
or ground-based cameras. By harnessing the power of artificial intelligence and machine learning, the system can learn
and recognize patterns indicative of human presence, even in complex and cluttered environments characteristic of dis-
aster zones.This innovative approach holds the potential to revolutionize disaster response efforts, providing rescuers
with invaluable support in locating and assisting individuals in distress.

The project encompasses several key components, including data collection, model training, real-time inference, and in-
tegration with hardware peripherals. Through this innovative approach, the project seeks to enhance disaster response
capabilities by providing emergency responders with timely and accurate information to prioritize rescue efforts and save
lives in critical situations. In response to the urgent needs presented by natural disasters, particularly floods which are
among the most prevalent and devastating globally, there is a clear and 2 present demand for more timely and effective
rescue operations. Traditional methods of search and rescue can be significantly hampered by the conditions in flood-af -
fected areas, characterized by extensive water coverage and often chaotic debris fields. This complexity makes locating
and rescuing flood victims a highly challenging task that requires innovative solutions to improve efficiency and effect-
iveness. Artificial Intelligence (AI) technologies offer promising enhancements to these traditional methods. By lever-
aging machine learning models like the Single Shot MultiBox Detector (SSD) coupled with advanced image processing
capabilities from OpenCV and the computational power of TensorFlow, it becomes feasible to rapidly and accurately de-
tect humans in diverse and challenging flood scenarios. The concept involves using drones equipped with cameras that
stream real-time footage to AI systems capable of analyzing and pinpointing human presence amidst the vast and unpre-
dictable landscapes of flood zones. The development of such a system begins with the meticulous collection and prepara-
tion of data. A robust dataset is crucial, typically comprised of images and videos that depict a wide range of flood condi -
tions—varying water levels, different times of day, and the presence of individuals in various positions and groupings.
Given the difficulty in obtaining real-world data of this nature, synthetic data generation and data augmentation are often
employed to mimic these varied conditions, thereby enhancing the model’s ability to learn and generalize from complex
inputs. The choice of the SSD model for this application balances the need for real-time processing with the demand for
high detection accuracy. This model, known for its efficiency in detecting objects in images quickly, is adapted through a
process called transfer learning, which fine-tunes a pre-existing model trained on a broad dataset to the specific task of
human detection in flood conditions.

This training process is intensive, requiring careful calibration and validation to ensure that the model performs well un -
der practical conditions. 3 Integrating this trained model into a deployable system involves both software and hardware
considerations, especially the capability to process high-resolution video feeds in real time. This system is typically
mounted on drones that are capable of covering large areas quickly, providing real-time insights that are crucial for direct-
ing rescue efforts effectively. Testing and deploying such a system requires careful planning and coordination with local
emergency services. Simulations and controlled trials help refine the system before it faces real-world conditions, where
its performance can directly influence the outcomes of rescue operations. After deployment, continuous feedback from
operational use provides invaluable data that can be used to further refine and enhance the model and its deployment
strategies. Moreover, deploying AI in such critical applications must be handled with a high degree of responsibility, es -
pecially concerning ethical considerations like privacy and data security. Transparency about the capabilities and limita -
tions of the system is also crucial to maintaining trust and ensuring that the technology is used effectively and judiciously.
This project represents a significant step forward in the use of AI for humanitarian assistance. By enhancing the speed
and accuracy of rescue operations during floods, AI technology not only pushes the boundaries of what is technologically
possible but also underscores the profound impact that innovative solutions can have on society's ability to respond to
natural disasters. Such advancements are crucial in saving lives, reducing the impact of disasters, and paving the way for
more resilient and responsive emergency management practices.

II. LITERATURE REVIEW

SNo Title,Year Author Key Findings Drawbacks

1 Optimizing SSD Ar- John A. Doe, Emily R. This paper explores Different disaster sce-
chitectures for Real- Smith, Michael Q. Roe modifications to the narios present unique
Time Human Detec- SSD model to enhance challenges, such as
tion in Disaster Sce- its performance. varying degrees of vis-
narios, ual obstruction, diverse
types of debris, and
2023 different environmen-
tal conditions.

2 Field Deployment of Linda K. Johnson, Car- This paper presents a Difficulties in integrat-
an AI-based Human los B. Martinez, Anita study of deploying an ing with varied exist-
Detection System V. Singh advanced AI-based hu- ing digital infrastruc-
man detection system tures across different
2024 across various disaster regions.
environments.

3 Comparative Study of Robert G. Lee, Sophia This paper study ana- Lower accuracy in
AI Models for Human N. Tran, David Z. lyzes the effectiveness cluttered environments
Detection in Post-Dis- Zhou of two prominent ob- typical of post-disaster
aster Scenarios: SSD ject detection models, scenarios.
vs. YOLO, SSD and YOLO for
human detection.
2023
 4 Enhancing Edge Com- Emily R. Thompson, This paper explores the Handling extreme vari-
puting with SSD for Raj Patel, Ana Maria integration of SSD al- ations in environmen-
Efficient Real-Time Gonzalez gorithms with edge tal conditions, poten-
Human Detection in computing devices to tial data privacy issues.
Urban Disaster Scenar- enhance real-time hu-
ios, man detection.

2024

5 Field Deployment Carlos M. Rodriguez, This paper details the Limitations of technol-
Challenges of AI- Lisa Marie King, operational challenges ogy integration under
Based Human Detec- Ahmed Khan encountered during the varied and harsh envi-
tion Systems in Natu- field deployment of ronmental conditions.
ral Disasters, AI-based human detec-
tion systems in natural
2024 disasters.

III. METHODOLOGY

A. Implementing Human Detection

Human detection is a computer vision technology related to image processing that deals with detecting instances of se -
mantic objects of a certain class in digital images and videos. With the advent of deep neural networks, object detection
has taken the centre stage in the development of computer vision with many models developed such as R-CNN and it's
variant, Single Shot Detectors (SSD) models as well as the famous You Only Look Once (YOLO) models and it’s many
versions. The object detection models are categorized as two major types such as one (single) stage object detectors such
as YOLO and SSD and two (dual) stage object detectors such as R-CNN. The major difference between the two is that in
the two-stage object detection models, the region of interest is first determined and the detection is then performed only
on the region of interest. This implies that the two-stage object detection models are generally more accurate than the
one-stage ones but require more computational resources and are slower. Human detection is a computer vision techno -
logy related to image processing that deals with detecting instances of semantic objects of a certain class in digital images
and videos. With the advent of deep neural networks, object detection has taken the centre stage in the development of
computer vision with many models developed such as R-CNN and it's variant, Single Shot Detectors (SSD) models as
well as the famous You Only Look Once (YOLO) models and it’s many versions. The object detection models are cat-
egorized as two major types such as one (single) stage object detectors such as YOLO and SSD and two (dual) stage ob -
ject detectors such as R-CNN. The major difference between the two is that in the two-stage object detection models, the
region of interest is first determined and the detection is then performed only on the region of interest. This implies that
the two-stage object detection models are generally more accurate than the one-stage ones but require more computa -
tional resources and are slower.

B. SSD Mobilenet v2 and CNN Layers

SSD Mobilenet V2 is a one-stage object detection model which has gained popularity for its lean network and novel
depth wise separable convolutions. It is a model commonly deployed on low compute devices such as mobile (hence the
name Mobilenet) with high accuracy performance. It provides real-time inference under compute constraints in devices
like smartphones. Once trained, MobileNetSSDv2 can be stored with 63 MB, making it an ideal model to use on smaller
devices.
 Architecture of a convolutional neural network with a SSD detector

C. Grid Cell

SSD divides the image using a grid. Each grid cell be responsible for detecting objects. Detection objects simply means
predicting the class and location of an object.

D. Anchor Box

Each grid cell in SSD can be assigned with multiple anchor/prior boxes. Anchor boxes are pre-defined. Each one is re -
sponsible for a size and shape.

E. CNN Layer

Convolutional Neural Networks (CNNs) form the backbone of object detection models like the Single Shot Multi box
Detector (SSD). In CNNs, layers are organized hierarchically to progressively extract and abstract features from input im-
ages. Convolutional layers apply filters across the input image to detect local patterns or features, such as edges or tex -
tures. Fully connected layers integrate high-level features for classification or regression tasks.

IV. CONCLUSION

Human detection using CNN system in flood areas on the Raspberry Pi represents a significant advancement in techno-
logy with the potential to save lives and mitigate the impact of catastrophic events. Through the integration of machine
learning algorithms, computer vision techniques, and hardware capabilities of the Raspberry Pi, this system empowers
emergency responders with real-time information to locate and rescue individuals in disaster-affected areas more effi -
ciently and effectively. By collecting diverse training data, preprocessing images, training robust models, and deploying
the system in real-world scenarios, developers can contribute to improving disaster response capabilities and enhancing
public safety. However, continuous refinement, testing, and adaptation are essential to ensure the system's reliability and
adaptability to evolving disaster scenarios. With ongoing advancements in technology and collaborative efforts across
disciplines, the AI-driven human detection system holds promise for revolutionizing disaster response efforts and safe-
guarding communities against unforeseen challenges in the future.

ACKNOWLEDGEMENT

We would like to express our deepest gratitude to our advisor, Mrs. M Mahil, for her invaluable guidance, support, and
encouragement throughout the course of this project. We also thankful to Professor Mrs. Tamil Pavai for her insightful
feedback and for providing access to essential resources at Government College of Engineering Tirunelveli Univer -
sity.Special thanks to our colleagues in the Computer Vision Lab for their helpful discussions and collaborative spir -
it.Lastly, we would like to thank our families and friends for their constant support and encouragement throughout our
studies.
REFERENCES
[1] R. Muwardi, M. Yunita, H. U. Ghifarsyam and H. Juliyanto (2022) ‘Optimize Image Processing Algorithm on ARM
Cortex-A72’ ,Jurnal Ilmiah Teknik Elektro Komputer dan Informatika Vol. 8,pp. 399-409.

[2] R. Muwardi, H. Qin, H. Gao, H. U. Ghifarsyam and M. Yunita (2020) ‘Research and Design of Fast Special Human
Face Recognition System’, Broadband Communications, Wireless Sensors and Powering.,pp. 68-73.

[3] R. Muwardi, H. Gao, H. U. Ghifarsyam, M. Yunita (2021) ‘Network Security Monitoring System Via Notification
Alert’ , Journal of Integrated and Advanced Engineering (JIAE), vol. 1, no. 2,pp. 113-122.

[4] F. A. Bohani, S. R. Yahya and S. N. H. S. Abdullah (2021)‘Journal of Integrated and Advanced Engineering’ , (JIAE),
vol. 1, no. 1,pp. 37-52.

[5] S. Suganyadevi, V. Seethalakshmi and K. Balasamy (2022) ‘A review on deep learning in medical image analysis’ ,
International Journal of Multimedia Information Retrieval, vol. 11, no. 1,pp. 19–38.