www.ijcrt.

org © 2022 IJCRT | Volume 10, Issue 3 March 2022 | ISSN: 2320-2882

WILD ANIMAL DETECTION IN

AGRICULTURE FARMS USING DEEP
CONVOLUTIONAL NEURAL NETWORK
1
FASEEHA M M, 2AMBILY JACOB
1
Msc Scholar, 2Assistant Professor
1,2
Department of Computer Science,
1,2
St.Joseph’s College (Autonomous), Irinjalakuda, Thrissur, India

ABSTRACT: On account of farmlands or rural terrains reconnaissance is vital to keep unapproved

individuals from accessing the region as well as to shield the region from animals. Different techniques point
just at observation which is basically for human interlopers, yet we will quite often fail to remember that the
primary adversaries of such ranchers are the animals which obliterate the harvests. Crop damage brought
about by animal assaults is one of the significant dangers in lessening the harvest yield. Because of the
extension of developed land into past wildlife territory, crop striking is becoming one of the most alienating
human-wildlife clashes. Effective and solid checking of the wild animals’ right in their natural habitat is
fundamental. This project fosters an algorithm to identify the animals that intrudes into the agriculture land.
Since there are enormous number of various animals physically distinguishing them can be a troublesome
undertaking. This calculation arranges animals in view of their pictures so we can screen them all the more
proficiently. This can be accomplished by applying yolo v3 algorithms which is a powerful real-time object
detection algorithms. YOLOv3 detect an object with the help of the features of deep convolutional neural
network.
Key words: Image Processing, Real-time Object, YOLOv3, Deep Convolutional Neural Network.

INTRODUCTION

One of the significant issues that is faced by farmers are their yields get damaged by wild animals that intrudes
their crops. Wild animal interruption has forever been a continuing issue to the ranchers. A portion of the
animals that go about as a danger to the yields are wild boar, deer, wild buffalo, elephants, tiger, monkeys and
others. These animals might benefit from crops and furthermore go around the field without any rancher and
accordingly make damage those yields. This may thus bring about critical misfortune in the yield and will
make extra monetary security all together arrangement with the result of the damage.
In any case, wildlife-friendly cultivating regularly brings about lower effectiveness. In this manner, endeavors
have been made to foster programmed frameworks fit for identifying wild animals in the harvest without
superfluous discontinuance of the cultivating activity. For instance, a recognition framework in light of
infrared sensors has been answered to lessen wildlife mortality in Germany [1]. In [2] a UAV-based framework

IJCRT2203110 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org a837

www.ijcrt.org © 2022 IJCRT | Volume 10, Issue 3 March 2022 | ISSN: 2320-2882
for roe deer grovel location is introduced. The creators show that warm imaging can be utilized to identify roe
deer grovels based on elevated film, but the location is as yet performed physically.
Here we are using Real-time object detection algorithms which detect an object with the help of deep
convolutional neural network. It helps in identifying wild animal intrusion to the crops. For this project we
need some cameras affixed at agricultural areas, so that it is clearly visible that who is entering to the particular
area. Our system will identify the animal and send a warning notification to the farmer that particular animal
has been entered to his crops. Now this will be easy for farmers to perform he next step to distract those
animals from his crops.

OBJECTIVE

The primary goal of the project is to safe watchman the farming field from wild animals and furthermore to
safeguard them by pushing them away as opposed to killing. The project additionally plans to safeguard
human lives from creature assaults. We are involving an integrative methodology in the field of Deep Learning
to give a checking and repulsing framework for crop insurance against creature assaults.

EXISTING SYSTEM

The current systems essentially give the observation usefulness. Additionally these systems don't give security
from wild animals, particularly in such an application region. They additionally need to make moves in light
of the on the kind of animal that attempts to enter the region, as various techniques are taken on to keep various
animals from entering such confined regions. Likewise the ranchers resort to different techniques by raising
human manikins and likenesses in their homesteads, which is ineffectual in warding off the wild animals,
however is valuable somewhat to avert birds .The other usually involved strategies by the ranchers to forestall
the harvest vandalization by animals incorporate structure actual obstructions, utilization of electric wall and
manual reconnaissance and different such thorough and risky techniques.

STRATEGIES TO PROTECT CROPS

Fruitful farmers generally look to decide the acceptable degree of wild animal harvest assurance utilizing one
of the accompanying innovations:
1. Agricultural fences
 Electric fences
 Plastic fences
 Wire fences
 Wood fences
2. Natural repellents
 Lavender and beans
 Chilli peppers
 garlic emulsion
 Smoke Fish
 Egg based repellent

IJCRT2203110 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org a838

www.ijcrt.org © 2022 IJCRT | Volume 10, Issue 3 March 2022 | ISSN: 2320-2882
3. Chemical repellent
4. Electronic repellent
 Sonic Electronic repellent
 Ultrasonic Electronic repellent

SCOPE OF STUDY

1. To plan a security system for farm assurance

2. Restrict the passage of animal into the farm
3. Use GSM module for cautioning us
4. Plan a system that sounds through solar animal anti-agents when animal attempts to go into the farm
5. In night streak light will zero in on that side.
6. The camera consistently screens the fields and gives the video feed to the farmer at home 24×7 for the
entire day
7. The system guarantees that the caution isn't set off by the presence of a human in the field, or by means of
any arbitrary movement.
8. The system is fit for turning On/Off consequently and averting the animals in this way safeguarding the
fields from any harm additionally we can arrangement a Timer according to farmer's prerequisite

CONCEPT OF THE MODEL

1. Visual recognition - to separate elements from pictures.

2. AI - Supervised learning is required for the model to order the animals.
3. Deep Learning - This gives the mental ability of breaking down the animal recognized to a specific
class and consequently produce the proper result, very much like how human examinations and this
can be accomplished with the assistance of CNN.
4. Keras and OpenCV : which will help in the handling of the information gained including all the above
ideas.
5. Anaconda labelImg : LabelImg is a graphical image annotation tool.It is written in Python and uses Qt for its
graphical interface.Annotations are saved as XML files in PASCAL VOC format, the format used by ImageNet.
It supports YOLO and CreateML formats.
6. Google Colab : Colab is a free Jupyter notebook environment that runs entirely in the cloud. Most
importantly, it does not require a setup and the notebooks that you create can be simultaneously edited
by your team members - just the way you edit documents in Google Docs. Colab supports many
popular machine learning libraries which can be easily loaded in your notebook.

IJCRT2203110 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org a839

www.ijcrt.org © 2022 IJCRT | Volume 10, Issue 3 March 2022 | ISSN: 2320-2882
PROPOSED SYSTEM

The proposed system uses YOLOv3 algorithms to detect real-time object.

YOLOv3 is a real-time object discovery algorithm that distinguishes explicit objects in recordings, live feeds,
or pictures. Consequences be damned uses highlights learned by a deep convolutional neural network to
recognize an object. YOLOv3 is a superior form of YOLO and YOLOv2. YOLO is carried out utilizing the
Keras or OpenCV deep learning libraries

Figure 1. Overall process of proposed system.

For developing this system we have to collect some dataset. Here our datasets are images of wild animals of
8 different classes. After collecting the dataset it will go through an image pre-processing step, it is also called
as annotation and we will get our final dataset for training. These dataset contains the class file, images and
txt files of those images, which is automatically created after successful annotation. Now we can start to train
our system, we are using google colab for training and testing. While the dataset is going through training
process it will generate some weights that is later useful for testing. Training process may take approximately
12-15 hours, but testing can be done easily with the help of those trained weights. While testing the system it
will check an image and look for matching class and it will predict an output which have an accuracy near to
1. After that it will send these predicted class name to firebase which is a real-time database, further the class
name will be notified to the user’s device

IJCRT2203110 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org a840

www.ijcrt.org © 2022 IJCRT | Volume 10, Issue 3 March 2022 | ISSN: 2320-2882

Figure 2. Flow chart of the working model

DATASET

Image dataset is used to this model. We use 8 classes, they are elephant, tiger, leopard, wild boar, deer, wild
buffalo, monkey and peacock. These animals are some of the main intruders of the agriculture area and they
are also threat to human life. Dataset are mainly divided into 2 parts. The majority will go for training and the
other for testing
The training dataset includes:
Elephant – 523 images
Tiger – 550 images
Leopard – 670 images
Wild boar – 520 images
Deer – 560 images
Wild buffalo – 534 images
Monkey – 600 images
Peacock – 533 images

The testing dataset includes:

Elephant – 103 images
Tiger – 100 images
Leopard – 110 images
Wild boar – 90 images
Deer – 98 images
Wild buffalo – 69 images
Monkey – 80 images

IJCRT2203110 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org a841

www.ijcrt.org © 2022 IJCRT | Volume 10, Issue 3 March 2022 | ISSN: 2320-2882
Peacock – 92 images

Figure 3. Sample dataset collection of 8 classes.

YOLOv3

Object grouping frameworks are utilized by AI projects that see the subjects of interest that are explicit objects
in a class. The frameworks sort objects in pictures into bunches where objects with comparative qualities are
set together, while others are ignored except if customized to do in any case. As average for object finders,
the elements learned by the convolutional layers are gone to a classifier which makes the identification
forecast. In YOLO, it uses convolutional layer for the expectation which depends on a layer that utilizes 1×1
convolutions.

YOLO is named "you just look once" on the grounds that its forecast utilizes 1×1 convolutions; the size of
the expectation map is actually the size of the element map before it. YOLO is CNN that is developed to doing
an object location in real-time. CNNs are classifier-based frameworks that can interaction input pictures as
organized varieties of information and distinguish designs between them (view picture underneath). YOLO
enjoys the benefit of being a lot quicker than different networks nevertheless keeps up with exactness. It
permits the model at the test time to take some gander at the entire picture, so it can educate the forecasts by
the worldwide setting in the picture. YOLO and other CNN "score" locales in light of their likenesses to
predefined classes.

IJCRT2203110 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org a842

www.ijcrt.org © 2022 IJCRT | Volume 10, Issue 3 March 2022 | ISSN: 2320-2882

Figure 4. Example of YOLOv3 Computer Vision

HOW TO USE YOLOv3

The first step to using YOLOv3 would be to decide on a specific object detection project. YOLOv3 performs
real-time detections, so choosing a simple project that has an easy premise, such as detecting a certain kind of
animal or car in a video, is ideal for beginners to get started with YOLOv3.In this section, we will go over the
essential steps and what you have to know for using YOLOv3 successfully.

MODEL WEIGHTS
Weights and cfg (or configuration) files can be downloaded from the website of the original creator of
YOLOv3: https://pjreddie.com/darknet/yolo. You can also (more easily) use YOLO’s COCO pretrained
weights by initializing the model with model = YOLOv3 (). Using COCO’s pre-trained weights means that
you can only use YOLO for object detection with any of the 80 pretrained classes that come with the COCO
dataset. This is a good option for beginners because it requires the least amount of new code and
customization.

MAKING A PREDICTION

The convolutional layers remembered for the YOLOv3 design produce a recognition expectation in the wake
of passing the elements learned onto a classifier or regressor. These elements incorporate the class name,
directions of the bouncing boxes, sizes of the jumping boxes, and the sky is the limit from there. In YOLOv3
and its different renditions, the way this forecast map is deciphered is that every cell predicts a proper number
of jumping boxes. Then, at that point, whichever cell contains the focal point of the ground truth box of an
object of interest is assigned as the cell that will be at last liable for foreseeing the object. There is a huge load
of math behind the internal functions of the forecast design.

IJCRT2203110 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org a843

www.ijcrt.org © 2022 IJCRT | Volume 10, Issue 3 March 2022 | ISSN: 2320-2882

Figure 5. Output prediction.

LOSS FUCTION
YOLO uses sum-squared error between predictions (the one with highest IoU) and ground truth to calculate
loss. The loss function composes of:

 The classification loss.

 The localization loss.
 The confidence loss (the abjectness of the box).

 Classification loss
𝑆2
𝑜𝑏𝑗
∑ 1𝑖 .∑𝑐𝑐∈𝑐𝑙𝑎𝑠𝑠𝑒𝑠 (pi(cc) − p̂𝑖 (𝑐𝑐))2
𝑖=0

where 1𝑜𝑏𝑗
𝑖 = 1 if an object appears in cell i, otherwise 0;

p̂𝑖 (𝑐𝑐) denotes the conditional class probability for class cc in cell i.

 Localization loss
𝟐 𝒐𝒃𝒋
𝝀𝒄𝒐𝒐𝒓𝒅 ∑𝑺𝒊=𝟎 ∑𝑩 ̂𝒊 )𝟐 + (𝒚𝒊 − 𝒚
𝒋=𝟎 𝟏𝒊𝒋 [(𝒙𝒊 − 𝒙 ̂𝒊 )𝟐 ] +
𝟐 𝟐
𝒐𝒃𝒋 ̂ 𝒊 )𝟐 ]
𝝀𝒄𝒐𝒐𝒓𝒅 ∑𝑺𝒊=𝟎 ∑𝑩 ̂ 𝒊 ) + (√𝒉𝒊 − √𝒉
𝒋=𝟎 𝟏𝒊𝒋 [(√𝒘𝒊 𝒊 − √𝒘

Where 1𝑜𝑏𝑗
𝑖𝑗 = 1 if the jth boundary box in cell i is responsible for detecting

𝝀𝒄𝒐𝒐𝒓𝒅 Increases the weight for the loss in the boundary box coordinates.
IJCRT2203110 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org a844
www.ijcrt.org © 2022 IJCRT | Volume 10, Issue 3 March 2022 | ISSN: 2320-2882
YOLO predicts the square root of bounding box width and height in order to differentiate large and small
boxes. By setting 𝝀𝒄𝒐𝒐𝒓𝒅 (default: 5), we put more emphasis on the boundary box accuracy.

 Confidence loss

If an object is detected in the box, the confidence loss is:

𝑺𝟐 𝑩
𝒐𝒃𝒋 𝟐
̂𝒊 )
∑ ∑ 𝟏𝒊𝒋 (𝑪𝑰 − 𝑪
𝒊=𝟎 𝒋=𝟎

Where 1𝑜𝑏𝑗
𝑖𝑗 = 1if the jth boundary box in cell i is responsible for detecting the object, otherwise 0;
̂ 𝒊 is the box confidence score of the box j in cell i.
𝑪
However, if an object is not detected:
𝑺𝟐 𝑩

𝝀𝒃𝒂𝒄𝒌𝒈 ∑ ∑ 𝟏𝒊𝒋
𝒃𝒂𝒄𝒌𝒈 ̂ 𝒊 )𝟐
(𝑪𝑰 − 𝑪
𝒊=𝟎 𝒋=𝟎

𝒃𝒂𝒄𝒌𝒈 𝒐𝒃𝒋
Where 𝟏𝒊𝒋 is the complement of 𝟏𝒊𝒋 .
̂ 𝒊 is the box confidence score of the box j in cell i.
𝑪
𝝀𝒃𝒂𝒄𝒌𝒈 weights down the loss when detecting background.
As most boxes do not contain any objects, we weight the loss down by a factor 𝝀𝒃𝒂𝒄𝒌𝒈 (default: 0.5) to
balance the weight.

CONCLUSION

The issue of yield destroying by wild animals has turned into a significant social issue in the current time. It
requires dire consideration and a powerful arrangement. Subsequently this project conveys an extraordinary
social significance as it plans to resolve this issue. Thus we have planned a shrewd installed farmland
protection and observation based framework which is minimal expense, and furthermore consumes less
energy. The principle point is to forestall the deficiency of yields and to shield the region from intruders and
wild animals which represent a significant danger to the rural regions. Such a framework will be useful to the
ranchers in safeguarding their plantations and fields and save them from critical monetary misfortunes and
furthermore saves them from ineffective endeavors that they suffer for the protection of their fields. This
framework will likewise help them in accomplishing better harvest yields consequently prompting their
monetary prosperity.

IJCRT2203110 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org a845

www.ijcrt.org © 2022 IJCRT | Volume 10, Issue 3 March 2022 | ISSN: 2320-2882
REFERENCES
[1]
Tank, V.; Haschberger, P.; Dietl, H.; Lutz, W. Infrarotoptischer Wildsensoreine Entwicklung
zur Detektion von Wild in Wiesen und zur Wildrettung bei der Fr¨uhjahrsmahd. Zeitschrift f¨ur
Jagdwissenschaft 1992, 38, 252–261.

[2]
Israel, M. A UA-Based Roe Deer Fawn Detection System. In Proceedings of the International
Conference on Unmanned Aerial Vehicle in Geomatics (UAV-g), Zurich, Switzerland, 14–16
September 2011; pp. 1–5.
[3]
https://www.itsrm.org/itd-exploring-wildlifedetection-system-in-northern-idaho-to-improve-driver-
safety/
[4]
A Mammerri, “Multi static radar for detection of wild animals” Christ Church University, U.K.
‘News Report about wild animal attacks in agriculture field’, http://www.tribuneindia.com/news/himachal/
[5]

community/wild- animals-causing-heavy-losses-to-farmers-in-state/94186.html, accessed 13 October 2016.

[6]
Green, R.; Stowe, T. The decline of the corncrake Crex crex in Britain and Ireland in relation to
habitat change. J. Appl. Ecol. 1993, 30, pp. 689–695.
[7]
Koffijberg, K.; Schaffer, N. International Single Species Action Plan for the Conservation of the Corncrake.
Crex crex; CMS Technical Series No. 14 and AEWA Technical Series No. 9; Council Directive 79/409/EEC
on the Conservation of Wild Birds (Birds Directive) of the European Union (EU); EU: Bonn, Germany, 1976.
[8]
Kim Arild Steen, Andr´es Villa-Henriksen, Ole Roland Therkildsen, and Ole Green, Automatic Detection
of Animals in Mowing Operations Using Thermal Cameras, Sensors 2012, 12, 7587-7597;
doi:10.3390/s120607587
[9]
‘Wikipedia’,https://en.wikipedia.org/wiki/Jugaad,accessed 24 August 2016.
[10]
DR.R.S. SABEENIAN, N. DEIVANAI, B. MYTHILI, Wild Animals Intrusion Detection using Deep
Learning Techniques, ISSN 0975-2366ÝOI:https://doi.org/10.31838/ijpr/2020.12.04.164.
[11]
Nagaraju Andavarapu and Valli Kumari Vatsavayi,
Wild-Animal Recognition in Agriculture Farms Using W-COHOG for Agro-Security, International Journal
of Computational Intelligence Research ISSN 0973-1873 Volume 13, Number 9 (2017), pp. 2247-2257.
[12]
Hemant Shetty, Hemant Singh, Fazal Shaikh, Animal Detection using Deep Learning, Volume 11 Issue
No.06, ISSN 2321 3361© 2021 IJESC.
[13]
Matuska, Slavomir, et al.: 'Classification of wild animals based on SVM and local descriptors.' AASRI Procedia
9 (2014), pp. 25-30.
[14]
YOLOv3: Real-Time Object Detection Algorithm (What's New?) - viso.ai
[15]
Mini-YOLOv3: Real-Time Object Detector for Embedded Applications | IEEE Journals & Magazine | IEEE Xplore
[16]
You only look once (YOLO) -- (1) | Zhang Handuo's Site

IJCRT2203110 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org a846