Australian Journal of Multi-Disciplinary Engineering

ISSN: (Print) (Online) Journal homepage: www.tandfonline.com/journals/tmul20

Swarm of reconnaissance drones using artificial

intelligence and networking

Aashrith A Jain, Akarsh Saraogi, Pawan Sharma, Vibhav Pandit & Shivakumar
R Hiremath

To cite this article: Aashrith A Jain, Akarsh Saraogi, Pawan Sharma, Vibhav Pandit &
Shivakumar R Hiremath (2023) Swarm of reconnaissance drones using artificial intelligence
and networking, Australian Journal of Multi-Disciplinary Engineering, 19:1, 81-88, DOI:
10.1080/14488388.2023.2220154

To link to this article: https://doi.org/10.1080/14488388.2023.2220154

Published online: 18 Jun 2023.

Submit your article to this journal

Article views: 273

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tmul20
AUSTRALIAN JOURNAL OF MULTI-DISCIPLINARY ENGINEERING
2023, VOL. 19, NO. 1, 81–88
https://doi.org/10.1080/14488388.2023.2220154

Swarm of reconnaissance drones using artificial intelligence and networking

Aashrith A Jain, Akarsh Saraogi, Pawan Sharma, Vibhav Pandit and Shivakumar R Hiremath
Department of Mechanical Engineering, PES University, Bengaluru, India

ABSTRACT ARTICLE HISTORY

This paper proposes a novel system for search and recognition in flood-affected areas using a Received 8 March 2023
swarm of autonomous drones. The system is based on a combination of machine learning, Accepted 27 May 2023
artificial intelligence, and networking technologies. The drones are equipped with cameras and KEYWORDS
sensors to identify victims, assess damage, and provide real-time updates to rescue crews. The Drone swarm;
drones can also create a temporary network in areas where the existing network has been reconnaissance; networking;
disrupted. The system has the potential to significantly improve the efficiency and effectiveness natural disaster; search and
of search and rescue operations in the aftermath of a natural disaster. rescue; machine learning;
artificial intelligence; gesture
recognition; prototype

1. Introduction implemented a system of drones to detect people in

distress and rank them according to their distress level.
Natural disasters can have a devastating impact on Currently, we use a simple approach to ranking
communities. In addition to the physical damage, people. The highest level of emergency is a person
natural disasters can also disrupt communication drowning, and the lowest level of emergency is a
and transportation networks, making it difficult to person needing food. As the dataset grows and the
reach victims and provide assistance.
model learns more, we can employ a more complex
In India, an average of 1,650 people die every year due
problem discovery approach.
to floods. This is one of the reasons why it is important to
The proposed system is still in the early stages of
develop new technologies that can help to improve the
development, but it has the potential to revolutionise
efficiency of search and rescue operations.
the way that natural disasters are responded to. The
This paper proposes a new system that uses a swarm
system has the potential to save lives, reduce property
of drones to provide reconnaissance and networking
capabilities in the aftermath of a natural disaster. The damage, and improve the efficiency of search and res­
drones are equipped with cameras and sensors that can cue operations.
be used to identify victims, assess the damage, and Figure 1 illustrates the working of the drone. The
provide real-time updates to rescue crews. The drones drone is equipped with a machine learning (ML)
are also equipped with networking equipment that can model that is capable of detecting humans in flood­
be used to create a temporary network in areas where waters. If the model detects a human, it will then assess
the existing network has been disrupted. the person’s risk of danger. If the person is deemed to
The proposed system has several advantages over be in danger, the model will prioritise them for rescue.
traditional methods of search and rescue. First, the This system can be used to quickly and efficiently
drones are able to access areas that are difficult or identify people in need of assistance in flooded areas,
impossible for humans to reach. Second, the drones which can help to save lives.
can provide real- time updates to rescue crews, which One of the distinctive and key features of this
can help them to make better decisions about how to unmanned aerial vehicle (UAV) is that it is capable
allocate resources. Third, the drones can create a tem­ of providing network connectivity in case of network
porary network in areas where the existing network has outage in a flood-affected area using its onboard net­
been disrupted, which can help victims to communicate working ecosystem. The network system will utilise
with their loved ones and access emergency services. the swarm of UAVs to create a local area network
In addition to these advantages, the proposed system (LAN) that will end at the base station 2021. As
also addresses the challenge of communication in a shown in Figure 2, the base station serves as a network
natural disaster. A simple microphone and speaker provider to the nearby UAV, which then extends the
setup would allow people in distress to send a message, network to the user who is connecting to it by captur­
but it is not guaranteed that the message will be received ing the network and acting as an extension from the
by the intended recipient. This is why we have base. To ensure long-range network connectivity,

CONTACT Vibhav Pandit [email protected]

© 2023 Engineers Australia
82 A. A. JAIN ET AL.

components in images or videos. Because stance

movements are frequently driven by
unique human activities, knowing a human’s
body position is crucial for action identification.
We may deduce what a person is doing by know­
ing the motion of a human body in a succession of
images, which is known as identifying the action.

2.2. Working of MediaPipe

Figure 1. Priority detection. MediaPipe Holistic is a machine learning method

that utilises a combination of convolutional neural
networks (CNNs), recurrent neural networks
(RNNs), and fully connected neural networks
additional hovering UAVs are able to automatically (FCNNs) to detect and track the 3D pose, facial
join this network and expand it. This procedure is only landmarks, and hand gestures of a person cap­
feasible because of network repeaters, which are elec­ tured on camera.The CNNs are responsible for
tronic devices that are capable of capturing network identifying features and patterns in the input
signals and extending them to a greater region. image data (2022).
They are used to detect the body and facial
landmarks by identifying key points and contours
2. Machine learning model on the body and face. The RNNs are then used to
track the movement of these key points and con­
2.1. Mediapipe
tours over time. This allows the method to track
Human detection and human posture recognition are the 3D pose of the person in real time, as well as
two steps that are normally treated independently. We detect and track hand gestures.
explore the difficulty of identifying individuals while The FCNNs are used for classification tasks,
also recognising their stances. To overcome the pro­ such as recognising specific poses or gestures.
blem, we utilise the capabilities of Mediapipe, which They are trained on a dataset of labelled images
employs a multi-task neural network architecture to that contains examples of the poses and gestures
simultaneously detect and track the 3D pose, facial that the model is expected to recognise. During
landmarks, and hand gestures of an individual cap­ inference, the FCNNs take the output of the CNNs
tured on camera. The program must be able to detect and RNNs as input and use it to classify the pose
how many people are in the frame and add key points or gesture that the person is performing (2022).
to the limbs of the body. The goal of human posture MediaPipe Holistic utilises several techniques
estimation is to anticipate the poses of human body such as data augmentation, transfer learning, and

Figure 2. Enabling network connectivity.

 AUSTRALIAN JOURNAL OF MULTI-DISCIPLINARY ENGINEERING 83

model ensembling. Data augmentation involves manoeuverability and machine learning

generating additional training data by applying capabilities.
various transformations to the existing data, such The Raspberry Pi, which consumes an average of 15
as rotation, scaling, and cropping. Transfer learn­ watts of power while in use, can operate for an
ing involves using a pre-trained model as a start­ extended period of time on a 10,000 mAh battery.
ing point and fine-tuning it on a new dataset. MediaPipe is a resource-light model designed for
Model ensembling involves combining the outputs mobile applications, which takes into account the lim­
of multiple models to improve the accuracy and ited power available. The repeaters are the most
reduce the risk of overfitting. To ensure optimal power-hungry components, but the power bank is
performance and maximum efficiency, the method sufficient to keep them powered on. Table 1 depicts
is configured to utilise only the essential land­ estimated power usage by the system.
marks, such as those of the body and hand,
while avoiding the use of superfluous landmarks,
which would only increase the complexity and
2.5. Workflow of the machine learning model
compiling time of the program. It is worth noting
that the complete MediaPipe Holistic system Figure 3 illustrates a complex system that inte­
includes over 540 landmarks, but the ML model grates a companion computer, a Raspberry Pi 4,
utilises only the landmarks that are necessary for and several accessories on a drone to facilitate the
the tasks at hand. processing of machine learning models. The
Pixhawk, which is a flight controller (2022), is
connected to the Raspberry Pi via MAVlink. This
2.3. Building the dataset connection
enables the drone to receive commands for control­
The custom dataset we created includes three unique
ling its movements through MAVlink. A camera is
stances, these stances are represented in Figures 9, 10 ,
also directly connected to the Raspberry Pi via USB,
and 13, and include the designations ‘Person
which provides the essential video input to the com­
Detected!’, ‘Food’, and ‘Help’. This dataset provides
puter for processing through the machine learning
our model with the ability to accurately recognise and
model.
respond to these specific poses with high degree of
To enable the machine learning process, the
accuracy. To generate this custom dataset, we lever­
Raspberry Pi is installed with OpenCV n.d., which is
aged the robust and powerful capabilities of the Sci-kit
a necessary library in our machine learning model.
Learn library 2007, enabling us to create a dataset of
We further utilise this library in Sci-Kit learn to create
over 700 images. This dataset is stored in a CSV file
our dataset, which is then trained on the MediaPipe,
and includes all the necessary information, such as the
enabling the detection of human poses. Once a human
coordinates of body landmarks and the labels for
pose is detected, the essential action can then be com­
specific poses, which the program utilises to accurately
municated to the drone, such as moving it
recognise specific poses.
closer to the person and obtaining their coordinates.
This process facilitates the reconnaissance process.
The integration of these components demonstrates
2.4. Powering the system
the potential for utilising sophisticated technology to
Each drone was equipped with two 5200 mAh Li- facilitate the execution of complex tasks in real-time
Po batteries connected in parallel, along with a scenarios. This showcases the increasing capabilities of
power bank with a capacity of 10,000 mAh. This machine learning and robotics in the realm of
allowed the drones to be powered separately for unmanned aerial vehicles.

Table 1. Estimated Power Consumption check.

84 A. A. JAIN ET AL.

Figure 3. Flowchart depicting workflow of ML Model.

3. How drones communicate? communicate with autopilots. It has a predefined dia­

lect for every version of autopilot, so every flight
To enable communication between drones,
controller uses and understands different types of
messages.
the Raspberry Pi must first be configured to run
telemetry through GPIO pins. This can be done by A MAV-Link 2 protocol can have a max length of
enabling the following parameters in raspi-config: 65.5 kb. This message string also contains SYS_ID,
which is used to differentiate between different
● enable_uart = 1 drones.
● dtoverlay = miniuart-bt The Raspberry Pi can read and write data through
● init_uart_clock = 64000000 telemetry relatively easily, since in Linux everything is
a file (Whether they are devices or storages).
The Raspberry Pi must also be configured to MAV-Proxy is a kit that basically makes it possible
automatically connect to the network. This can be to create a network endpoint for Ground Station to
done by editing the wpa_supplicant.conf file. connect to. It helps in communicating with the
Once the Raspberry Pi is configured, it can be used drones.
to communicate with the Pixhawk using MAV-Link. The Raspberry Pi can request data from the
MAV-Link is a messaging protocol that is used to flight controller, which includes but is not limited

Table 2. The estimated cost of implementing a single drone.

 AUSTRALIAN JOURNAL OF MULTI-DISCIPLINARY ENGINEERING 85

Figure 4. Visualisation on how the Swarm communicates.

to information like the drone’s GPS coordinates, 4. Fabrication of battery housing

velocity, orientation, etc. It is also able to give
The battery is the primary source of power for all of
commands to the flight controller through
the drone’s components. Lithium polymer (LiPo) bat­
MAVlink.
teries are among the most popular battery types used
Figure 4 demonstrates the role of MAVLINK
in drones due to their high energy density relative to
and MAVPROXY in our rescue system. This
their size and weight and higher voltage per cell
shows how a drone can collect data from other
(2022). In our model, we used a 4s2p configuration,
drones and basically how connection is established
which consists of four cells in series and two batteries
with the other drones.
in parallel. This configuration provides a higher vol­
We need every data from FC−1 as it is to be con­
tage and current output than a single LiPo battery,
trolled by this raspberry pi but we only
which is necessary for powering the drone’s motors
need a few parameters from drone 2 as it will be
and other components. However, fitting two batteries
enough to coordinate alignment and movement of
in the F450 drone frame made it very difficult. As a
drones.
result, we opted to increase the area on the base plate
The MAVLINK messages sent by FC is received
to accommodate the batteries. We then 3D-printed a
through mount point/dev/ttyAMA0 which is the
new base plate that was large enough to fit the two
UART port. Since everything is a file in Linux
batteries. This solution allowed us to successfully
including mount points we will easily be able to
install the batteries in the drone frame and power the
read and write data through UART. The data from
drone’s components.
MAVLINK can be proxied through MAVPROXY
into two drone instances and Ground Station can A modified base plate as seen in Figure 5 was
connect to each of the drones through it. designed to accommodate two 5200mAh batteries.
The base plate was extended by 4 mm on each end

3.1. Data transfer from drone to base station

Each drone in a swarm is equipped with a network
repeater. This allows the drones to create a local
area network (LAN) and communicate with each
other. The LAN can also be used to extend internet
access to people in the affected zone by getting
them connected to the drone’s network.
Since the base station is connected to the inter­
net, other drones can route traffic through the base
station and make use of the internet. This allows
the drones to transfer data to and from the base
station, as well as to and from the internet. Figure 5. 3D model of Battery housing and landing gear.
86 A. A. JAIN ET AL.

designers to ensure that a piece of equipment or

structure can withstand the estimated loads.
In this study, we performed structural analysis
on a 3D model of a battery holder using Autodesk
Fusion 360 software. The first step was to create
the geometry of the model. The object was then
meshed, and we were able to achieve very fine
meshing. The weight/force on the object was
then added. The applied force on the model is
shown by the arrow above the central core of the
Figures 7 and 8.
Figure 6. Prototype of drone. The colour coding used in the analysis is as
follows:
Red represents high stress, Blue represents
to create a battery holder and by 7 mm on each stresses that are present but the model can with­
end to create landing gear. Sharp edges were stand them, Green represents regions that are not
avoided and a support was added to prevent the experiencing much stress and are safe. This study
base plate from breaking. The modified base plate demonstrates the effectiveness of using Autodesk
was attached to the F450 drone using 4 mm Fusion 360 software for structural analysis. The
screws. This design improves the aesthetic appeal results of the analysis can be used to improve the
of the drone and enhances its functionality as seen design of the battery holder and ensure that it can
in Figure 6. withstand the estimated loads.
The results of the Safety factor analysis in
Figure 8 shows that there is no significant redness
4.1. Structural analysis of the fabricated battery
around the model, indicating that there is no
house
severe force on one side of the model. Instead,
Structural analysis is a process of determining the there are primarily blue and green regions in the
effects of loads and internal forces on structures, analysis, which appears to be safe for our
buildings, and objects. This allows engineers and applications.

Figure 7. Stress graph.

 AUSTRALIAN JOURNAL OF MULTI-DISCIPLINARY ENGINEERING 87

Figure 8. Safety factor graph.

5. Estimated prototype cost

6. Result and conclusion

6.1. Results
● A prototype of an autonomous drone was suc­
cessfully built and can operate without human
control. The estimated cost of implementing a
single drone is mentioned in Table 2.
● The drone is equipped with a network repeater
that enables us to broadcast the signal from the
base station to a greater distance.
● The drone also has a camera that streams live
footage of the disaster area and feeds it to the
Figure 9. Help gesture being identified by the model.
raspberry pi, which is pre-loaded with the
machine learning model.
● The machine learning algorithm can recognise
humans and estimate their postures, producing
results such as ‘Help’, ‘Food’, and ‘Person
Detected’. These results have been shown in
Figure 9, Figure 10 and Figure 13 respectively.

6.2. Conclusion
● The prototype of the drone can be used as a recon
drone to locate victims in the vicinity and relay their
position so that assistance may be sent as soon as
possible.
● The 3D model for the battery housing was successful
and can sustain the weight of the battery without
bending. Figure 10. Food gesture being identified by the model.
88 A. A. JAIN ET AL.

● The drone’s landing gear provides a perfect Disclosure statement

standoff, with none of the components making
No potential conflict of interest was reported by the author(s).
touch with the ground during landing.
● The Machine Learning model correctly detects the
person within the frame and plots the subject’s body Notes on contributors
landmarks as shown in Figure 11 and Figure 12.
Aashrith A Jain, B.tech in Mechanical Engineering at PES
University, Bengaluru. His areas of interest are in Designing,
Production, Computer Aided Design, Strength of materials
and Supply chain management. He has handled Process
Capibility and Machine Capibility of a particular manufac­
turing process at Bosch.
Akarsh Saraogi is a recent graduate of PES University,
Bengaluru, with a Bachelor of Technology in Mechanical
Engineering and a minor in Computer Science. His minor in
computer science provided him with a solid foundation in
programming and software development. He is now able to
use his expertise in mechanical engineering and computer
science to create novel solutions to complicated issues.
Pawan Sharma is a recent graduate of PES University,
Bengaluru, where he earned a Bachelor of Technology degree
in Mechanical Engineering. Pawan's involvement in the
robotics club and Aeolus Club gave him valuable experience
in building drones and debugging the problems efficiently.
Figure 11. Body Landmarks (Help). Vibhav Pandit is a recent graduate from PES University,
Bengaluru, specializing in Mechanical Engineering with a
minor in Computer Science. With interest in coding languages
like Python and Java, he has demonstrated a strong foundation
in programming. Driven by his profound passion for Artificial
Intelligence and Machine Learning, Vibhav actively engaged in
this project, dedicating his efforts primarily to the ML aspect.
Shivakumar R Hiremath is an Assistant Professor in the
department of Mechanical Engineering, pursuing his Ph.D
at VTU Belgaum. His research interests include engineering
materials, surface engineering, and thermal spray coating.

References
Asaamoning, G., P. Mendes, D. Rosário, and E.
Cerqueira. 2021. “Drone Swarms as Networked
Control Systems by Integration of Networking and
Computing.“ Sensors 21 (8): 2642. https://www.mdpi.
com/1424-8220/21/8/2642 .
Figure 12. Body landmarks.
Chen-Chiung, H., and H. Shu-Shuo. 2007. “A simple
and fast surveillance system for human tracking and
behavior analysis.“ 2007 Third International IEEE
Conference on Signal-Image Technologies and
Internet-Based System. ISBN: 978-0-7695-3122-9.
https://doi.org/10.1109/SITIS.2007.128 .
How to Choose Lithium Polymer Battery for your RC
Drone. Accessed April 20, 2022. https://robu.in/how-
choose-lithium-polymer-battery-for-drone/ .
Neural Network. Accessed April 18, 2022. https://www.investo
pedia.com/terms/n/neuralnetwork.asp#:~:text=A%20neural
%20network%20is%20a,organic%20or%20artificial%20in%
20nature.
Pixhawk Documentation. Accessed April 8, 2022.
https://docs.px4.io/master/en/flight_controller/pix
hawk5x.html .
A review of supervised and unsupervised machine
learning techniques for suspicious behavior recogni­
tion in intelligent surveillance system [Kamal Kant
Figure 13. Person being identified by the model. Verma1 • Brij Mohan Singh2 • Amit Dixit]. n.d.