Autonomous Drone for Defence Machinery

Maintenance and Surveillance

P.R. Sriram,Sandeep Kumar Ramani,Ram V Shrivatsav,Muthu M.Mankiandan,Nithin Ayyappaa
Department of Electronics and Communication Engineering
Sri Venkateswara College of Engineering, Sriperumbudur, Tamil Nadu, India
[email protected],[email protected],[email protected],[email protected],
[email protected]

Abstract—This proposed research work focuses on the im-

plementation of an autonomous unmanned aerial vehicle (UAV)
which is controlled using a pix hawk flight controller. The
Quad Copter is capable of navigating autonomously without
any real-time input from the user and also programmed to
follow a specified path autonomously. The algorithm enables a
control technique by which quad copter is empowered to fly
autonomously, trajectory tracking, graceful motion and accurate
altitude hold performance. Surveillance and Machinery mainte-
nance application are the primary applications designed for the
Defense purposes in the Line Of Control and War-zones. This
work is aimed to design a quad copter that will follow a command
to fly through specified way points. The deep learning algorithm
detects human motions and from data acquired camera and
ultrasonic sensors to the cloud. On deviations from the standard
protocol which is detected using a Sjcam 5000x elite Camera.
Also, here a backup auxiliary mini drone is programmed to eject
along with the data stored on the primary drone’s memory on an
unforeseen calamity or an attack that would damage the primary
drone’s ability to fly..
Keywords—Autonomous Drone,Deep Learning,Wireless Com-
munication,Auxiliary Drone,image Processing,Line Of Con-
trol,Sense Hat,UAV.
Fig. 1. Architecture Of the Drone

I. I NTRODUCTION
is pre trained by humans and stored on cloud. The primary
In todays developing world of technology the impact aim of the research in autonomous drones and navigation is to
they have created on the human life has been extraordinary. minimize the time and requirement of human operators. The
Vehicles have become smart and autonomous which paves way advantage of this is an increased reconnaissance ability at a
to incorporate the same into aerial and space vehicles too. lower risk and cost in terms of time and economics. Machinery
Drones are relatively small sized flying vehicles that can cater and armed equipment pose a serious threat to humans when
to numerous human as well as the needs of the ecosystem. it comes to monitoring and maintaining them in the event of
Autonomous drones are a boon to the human race. They are an error or a malfunction. An autonomous drone handles this
capable of navigating on their own without any user input. with ease as they have on bard hardware to communicate with
Machine learning and artificial intelligence play a key role in the device placed on the machinery that reads out its status
the structure of an autonomous drone design. In this research wirelessly to the drone when in flight around its radius of
work the primary focus is about the implementation of a quad communication.
copter in the Line of Control (LOC) for defense machinery
maintenance and surveillance. To increase the level of data II. STRUCTURE OF THE DRONE
security and backup security during times of calamity or attack
from the enemies an auxiliary drone is programmed to eject A. SOFTWARE STRUCTURE
automatically and render the data captured to the base station.
In order to control the motion of the drone, we have used
The drone is designed to navigate around with four brush-
Mission Planner software where we assign the drone way-
less motors and a carbon fiber frame for stability and robust
points and actions. We have used Pix-hawk flight-controller
structure.Unmanned aerial vehicles or UAVs have become an
which sends all the required control signals to electronic speed
integral part of modern warfare. They are of low cost and risk
controller to vary the propeller rotation speed.
to the users. In recent times, the abilities of an UAV has been
extended to be on par with those of manned fighter air-crafts. For surveillance, we take images every seconds using
But even now, most UAVs are still used for camera and video Sjcam 5000x elite camera attached on the drone. On arrival at
filming purposes. Moreover, UAVs that can remain abode for the base station we transfer all the images to the AWS cloud
extended periods have been developed. This would be almost were we have our Yolov3 deep learning pre-trained model to
impossible to design for traditional air-crafts. Autonomy in which we pass all the images to predict human and objects
controlling these UAVs is a good way of reducing cost and motions. If human motions are detected we alert the military
yet maintaining the quality of human judgment as the model along with GPS coordinates. In addition to surveillance, we

978-1-7281-3780-3/19/$31.00 2019
c IEEE 288
Fig. 2. Constructed Image

Fig. 4. SCJAM 500x Camera

shown in Figure 6,the theoretical analysis closely matches the

experimental results obtained for the carbon ber blade. The
minor differ-ences between the experimental and analytical
results may be due to a less than perfect approximation of the
exact Rotor properties. Future work will include rening the
measurement of geometrical parameters such as airfoil shape,
twist,and chord using laser scanning techniques

IV. MACHINERY HARDWARE AND SOFTWARE

STRUCTURE

Fig. 3. Top Angle Over View of the Drone

The machinery data are collected using a various sensor
such as sense hat, NEO-6m GPS sensor attached to raspberry
zero microprocessors. The Sense hat module contains multiple
can update the software in the machine via drone Xbee module sensors in a single module which collects all the necessary
which sends the updated software code which is transmitted information about the machine environment such as tempera-
to the machine and the receiving Xbee stores the updated code ture, absolute pressure from HTS221 IC and 3D accelerometer,
and executes it. gyroscope and magnetometer through LSM9DS1 IC and logs
these data every 5 mins. The NEO-6m GPS module is used
The stability of the drone is constantly monitored and any to give the exact location of the machine which is used to see
deviation from the ideal state ignites a counter correction code if the machine is moved to a different location or not.
to act that would stabilize the drone.
V. DEEP LEARNING AND OBJECT DETECTION
III. HARDWARE STRUCTURE
We have used the most compact and lightweight HD
Our autonomous quad-copter drone is controlled using camera Sjcam 5000x elite on our drone to take images for
PixHawk flight controller which is powered using Lithium surveillance. The camera is capable of taking 8 Megapixel res-
Polymer 2200mAh battery. We have used the lightweight S500 olution images with 4x digital zoom and its lens are adjustable
frame with carbon fiber power distribution board which is used wide-angle fish-eye lens between 170 140 110 and 70. It has
to equally distribute the power to all the four DC motors. a battery life of 80 minutes which is powered by 900mAh
”Gyro stabilization technology is one of the most important battery and the overall weight is just 78g making this camera
components, allowing the drone to fly super smooth even in an ideal choice for drone as our quadcopter drone payload
strong winds and gusts. ” On the drone, we have the receiving is just 1kg. It has an inbuilt timer circuit which is set to take
Xbee module which stores the cipher encrypted message in images every 2 seconds during the flight of drone and we store
EEPROM of the Arduino Uno which is decrypted in the base them in an SD card. On arrival, we transfer all the images to
station which is analyzed for finding any deviations from AWS S3 bucket using a python script and pass these images to
preset ranges of values.An SCJAM 5000x Elite Camera is AWS EC2 instance where we have our trained neural network
used for object detection and deep learning input data.As model. Object detection on aerial images are challenging as an

Third World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4) 289
 Fig. 8. YOLOv3 object detection output

only top view of the object is available. We used UCF Aerial

Action data set and manually annotated the image frames from
the video to recognize humans and vehicles. We trained over
several models such as YOLOv3, Mobile net, and SSD object
Fig. 5. Combined Plot detection algorithm and found YOLOv3 to outperform others
and give high validation accuracy.

VI. DEEP LEARNING METHODOLOGY

Out of various object detection neural network algorithms
we took YOLOv3 because it makes detections at three differ-
ent scales. Detections at different layers help address the issue
of detecting small objects which are seen by the drone camera
by preserving the fine-grained features which help in detecting
small objects. The 13 x 13 layer is responsible for detecting
large objects, whereas the 52 x 52 layer detects the smaller
objects, with the 26 x 26 layer detecting medium objects.
The model has 106 layers fully convolutional with first 53
stacked layers trained on Image net data-set as shown in the
figure [ ]and the remaining 53 layers for detection. The custom
transfer learning model was trained on over 400 annotated
images using Pytorch framework to get a validation accuracy
of 91In order to further increase the accuracy, we added
data augmentation technique such as horizontal and vertical
translation, 5-degree rotation and scaling to the image along
Fig. 6. Power Vs RPM (Experimental vs Analysis) with the automatic tracking and updating of bounding box
coordinates over a 416x416 dimension image to get a score of
94During inference, the test images collected from drone are
preprocessed by converting the images into video and passing
into an optical flow image processing methods to find relative
motion frames and these frames are re-sized into 416x 416
dimension image before passing to the Amazon web services
cloud. There is a trade-off between taking images every one
second on the drone and a high-definition video as memory
storage requirement are large for video and higher resolution
images are possible. Therefore, we take images of the terrain
surface every second from the drone camera. On the cloud, the
test image is passed to the YOLOv3 model and prediction are
made to assist the defense official for surveillance purpose.

VII. AMAZON WEB SERVICES CLOUD STORAGE

AWS S3-Amazon Simple Storage Service is a cloud drive
for the world wide internet. It is designated to enable web-scale
computing simpler for engineers and programmers. Amazon
Fig. 7. waypoints directed path S3 has a simple web services platform that one can utilize
to store and retrieve any quantity of data, at any moment,
from anywhere on the internet. It gives any programmer access

290 Third World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4)
 to the same highly scaleable, reliable, fast, inexpensive data
storage interface that Amazon utilizes to run its own global
network of web sites. Maximum benefits can be obtained with
the scale and be utilized through the developers. The cloud
storage could be utilized through Amazon S3. Through AWS
regions you could upload documents and multimedia data
by creating a bucket. Data and information can be uploaded
via objects marked in the buckets. APIs are provides to
Amazon S3 to manage the necessary resources to implement
the objects. The real time results are generated through a
programmed neutral network. The neural network is trained by
a dedicated complex neural algorithm. This enables to capture
time to time images employed by the spectral camera. These
real time images are then uploaded via APIs and a python
coded script to the web services bucket. The above is requested
by specific standard 3 protocols. The web services bucket
obtains the captured images with the help of a instructed neural
networks and APIs. This helps to provide a regulated survey
and precise output generation.

VIII. IMPLEMENTATION
Fig. 9. Data collection and preprocessing steps The drone proposed here is designed to navigate through
the way points that are specified to it. All defense machinery
that is deployed in the line of control or in the sensitive
borders where they might be prone to threats. To overcome the
external threats a hardware setup that consists of a raspberry pi
along with a sense hat, which is programmed to communicate
using zigbee protocol. When the deployed drone approaches
defense machinery within a threshold radius, the drone can
gain access to the modules present in the machinery and
will be able to receive updates regarding the functionality
of the said machinery which is facilitated using the zigbee
protocol. The drone then carries the data and returns back to
the base station to report the status of the machinery which
is under surveillance. Modern day smart defense machinery
which require periodic software updates can be carried out
using these autonomous drones which act as a catalyst that
would incorporate the required updates to the machinery to
enhance its performance. The auxiliary drone is programmed
to eject along with the harvested data in case of a hostile attack
and return with the data to the base station autonomously based
on the pre- programmed hijack protocol. The 4k resolution
camera escalates the image processing and object recognition
using the pre trained models that are stored on the cloud
with optimum accuracy. The carbon fiber frame of the drone
provides a nimble structure which increases the aero dynamical
performance of the drone. This provides substantial payload
capacity. Although titanium blades of the quad copter increases
the overall production cost, they increase the ability of the
drone to sustain heavy turbulence and haphazard wind streams.
The brush-less motors ability to render high torque to weight
ratio and enhance torque per watt is a distinct feature of our
cutting edge design. This also increases the reliability and
reduces the noise output that would camouflage the drone from
the enemies. They have a longer life time and substantially
reduced electro-magnetic interference which is essential to
isolate the motor from the wireless protocols in the functional
parts of the drone.

Fig. 10. YOLOv3 first 53 layer architecture IX. PAYLOAD

There are a number of factors that play a role in deter-
mining the overall payload of the UAV. The main factors
that stand out are the power of the motor enclosed, number

Third World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4) 291
 of the propeller blades, the variety of battery used and most R EFERENCES
importantly the overall weight of the frame of the drone. What [1] Suet-Peng Yong,Yoon-Chow Yeong, Human Object Detection in Forest
we kept in mind while we designed this drone was that the with Deep Learning based on Drones Vision4th International Confer-
propeller thrust should be twice that of the total weight of our ence on Computer and Information Sciences (ICCOINS) Year: 2018
drone as well as its payload. A fifty percent allowance is to Page s: 1 - 5.
account for the effect of the wind, the air pressure and also [2] Sunyoung Cho, Jihun Park, Young Sook Shin, Sang-ho Lee, Rec-
the moisture in the air all of which can prove to test and push ognizing Human-Object Interactions via Target LocalizationControl
Automation and Systems (ICCAS) 2018 18th International Conference
the limits of the drone. on, pp. 836-840, 2018.
To get a payload of higher magnitude, a battery of greater [3] px4, Pixhawk https://www.pixhawk.org
capacity would be required. However the larger the size of the [4] D. Mellinger, M. Shomin, and V. Kumar, Control of quadrotors for
battery, the weight of the drone will also increase. To tackle the robust perching and landing,International Powered Lift Conference,
October 5-7, 2010, 2010.
increase in the payload, the drone is equipped with a frame that
[5] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, You Only Look
is crafted out of a durable carbon fibre material. The drone we Once: Unified, real-time object detection,Proceedings of CVPR, pp.
have designed has a weight capacity o 45 minutes. The total 1697-1700, 2016
flight time of the drone is estimated to be around forty five [6] UCF Aerial Action Data Sethttp://crcv.ucf.edu/data/UCF Aerial Ac-
minutes to an hour depending upon the operational payload. tion.php
[7] R. Girshick,Fast R-CNN, Proceedings of ICCV, pp.1440-1448, 2015
It is obvious that the higher the power-to-weight ratio of
[8] Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S. Fu, C., YSSD:
the drone, the more payload it can carry. Our drone has been Single Shot MultiBox Detector”,European Conference on Computer
designed to accommodate a higher value of power-to-weight Vision, pp. 21-37, 2012
ratio. The heavier the drone greater the impact on the power [9] S. Bouabdallah Design and control of quadrotors with application to
required to carry a big payload. This will also have an effect autonomous flying, Lausanne Polytechnic University, 2007
on the overall flight duration. [10] Z. Zaheer, A. Usmani, E. Khan and M. A. Qadeer, Aerial surveillance
system using UAV,Thirteenth International Conference on Wireless and
Our innovative design followed here justifies the use of Optical Communications Networks (WOCN),Hyderabad, 2016, Pp.1-7.
rotor blades of greater dimensions. Consider a hypothetical [11] C. G. Srcin, I. Drago and A. I. Chiril, IEEE, Powering aerial surveil-
situation were we consider two drones of the same mass, one lance drones.10th International Symposium on Advanced Topics in
will have rotors of greater dimensions and the the other will Electrical Engineering (ATEE),Bucharest, 2017, pp.237-240.
have rotors of smaller dimensions. The one with the large [12] Han, S., Shen, W., Liu, Z. Deep Drone: object detection and tracking
rotors will generate a greater magnitude of air thrust. However for smart drones on embedded systemStanford University,2012.
the one with the rotor blade of smaller dimensions will not be [13] LIU Xu,HU Haoquan, A Miniature Dual-Mode Microstrip Bandpass
Filter At 15.5 GHZ. IEEE,2011
able to emulate such a thrust.
[14] E. Mingkhwan and W. Khawsuk Digital image stabilization technique
for fixed camera on small size drone,Third Asian Conference on
X. C ONCLUSION Defence Technology (ACDT), Phuket, 2017, pp.12-19
This paper proposes an innovative addition to the pre- [15] General Thrust Equation NASA.Available from
https://www.grc.nasa.gov/www/k-12/airplane/thrsteq.htmlAccessed
existing UAVs and drones by equipping it with cutting edge au- on 11 SEPT 2017
tonomy. These changes that we proposed to make will display [16] D. Mellinger, M. Shomin, and V. Kumar , Control of quadrotors for
significant elevation in terms of performance and efficiency robust perching and landing.International Powered Lift Conference,
when compared to its predecessors. The deep learning aspect October 5-7, 2010, 2010.
of the drone is where the heart of the innovation lies. The [17] V. Kumar and N. Michael, Opportunities and challenges with au-
YOLOv3 algorithm is the feature that makes this UAV stand tonomous micro aerial vehicles Intl Symposium on Robotics Research,
out when compared to the existing models.The autonomous 2011.
operation would open doors to wide spread applications in [18] I. Kara, B. Y. Kara and M. K. Yetis, Energy minimizing vehicle rout-
ing problem, in Combinatorial Optimization and Applications (LNCS
Defense and Military regions that are life threatening to 4616), A. Dress, Y. Xu, and B. Zhu, Eds. Berlin, Germany: Springer,
manned soldiers to access.The high sensitive borders and latest 2007, pp. 6271.
generation military machinery can put under high accuracy top [19] Amazon Web Services Amazon Web Services. [Online]. Available:
notch surveillance using these Autonomous Unmanned Aerial https://aws.amazon.com/
Vehicles. [20] SJCAM SJ5000X ELITE SONY https://sjcam.com/sjcam-sj5000x-elite-
sony-imx078-gyro-4k24-2k-action-camera-pre-order/

292 Third World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4)