Proceedings of the International Conference on Inventive Research in Computing Applications (ICIRCA 2022)

IEEE Xplore Part Number: CFP22N67-ART; ISBN: 978-1-6654-9707-7

Efficient Wildlife Intrusion Detection System using

Hybrid Algorithm
1 2 3
Divya M eena Hari Krishna P Chakka Naga Venkata Jahnavi
School of Computer Science and School of Computer Science and School of Computer Science and
Engineering, VIT-AP University Engineering, VIT-AP University Engineering, VIT-AP University
Amaravathi, India Amaravathi, India Amaravathi, India
2022 4th International Conference on Inventive Research in Computing Applications (ICIRCA) | 978-1-6654-9707-7/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICIRCA54612.2022.9985684

divyameena.s@vitap.ac.in hari.19bce7675@vitap.ac.in venkata.19bce7082@vitap.ac.in

4 5
Patri Lalithya M anasa J Sheela
School of Computer Science and School of Computer Science and
Engineering, VIT-AP University Engineering, VIT-AP University
Amaravathi, India Amaravathi, India
lalithya.19bcd7013@vitap.ac.in sheela.j@vitap.ac.in

Abstract— Human-wildlife conflict arises when the needs and situation. A variety of s tudies are being developed to
behavior of animals have a detrimental influence on humans or address the issue, and some of them look to be viable. To
when humans have a negative impact on the needs of wildlife. assemble ev idence of malicious assaults or intrusions,
The primary causes of Man-Wildlife Conflicts include cutting-edge community forensics is currently unable to
agricultural expansion, human settlement, livestock
reliably monitor and analyze community site visitors [3].
overgrazing, deforestation, illegal grass gathering, and
poaching. Each year, human-animal conflict in human habitats
causes a massive loss of sources and put lives in jeopardy. As The ab ility of a system to distinguish between a dog and
the global human population continues to force wildlife out of a cat is referred to as image classification [4]. Despite being
their natural habitats, conflicts are unavoidabl e, which is why the oldest computer approach, it remains important in
habitat loss is one of the most prevalent dangers to endangered today's culture. Image classification remains a tough
animals. So, it is necessary to detect animals and identify the calculation due to impediments in images, like posture
animal detected to reduce the effects of human-animal conflict. variations, occlusion, illu mination, camouflage, and more.
This research study has developed a hybrid algorithm, which Deep neural learn ing is a form of machine learn ing that
classifies animal images into multiple groups using YOLO v5
enables a model to classify informat ion directly fro m of the
(You only look once) combined with CNN. The proposed
system distinguishes whether the animal is in human
train source, such as images, text, or speech [4]. The
environment or not, and then reliably distinguishes which creation of a Deep Neural Net work is required for this
animal class it belongs to using CNN. The model has been (DNN) [5]. Structuring a model fro m the ground up can
tested through its paces on a variety of tasks in analysis to yield better outcomes, but it’s also more difficult and time
define how well it performs in various scenarios. The system is inefficient. Transfer Learning, on the other hand, may be
being fine-tuned with the goal of attaining the most accurate utilized to develop extremely efficient neural network
results possible in recognizing and decreasing hazards posed system. Th is study has proposed a novel technique with
by animal invasions into human land. These experimental Yo lov5 and Alex net Architecture wh ich is a Convolutional
findings show that the yolov5 coalescing technique paired with
Neural network which used GPU to boost performance Here
CNN can properly categorize animals in habitats, with a 92.5%
accuracy from the proposed model.
these two networks are combined and confer us better
accuracy and standard model
.
Keywords— YOLO v5, Convolution Neural Network,
Deep Neural Networks, Labels, Adam optimizer II. LIT ERAT URE REVIEW
The desire is to make the habited liv ing area a wild
I. INT RODUCT ION animal-free zone. Early on, the sites that became man-made
W ild life is an essential part of the natural environment dwelling zones were forests, which have now evolved into
as it helps to maintain the ecological balance. Wildlife cities and towns. Due to drought and starvation in forests,
resources have been effectively secured, and the population animals prefer to migrate via habited reg ions surrounding
of all sorts of wild animals has been recovered, due to the forest zones during these periods. Siddhant kumar et al. [26]
adoption and implementation of wild life conservation have proposed their idea with standard solution, using g -
legislation [1]. Hu man-animal conflict manifests itself in Crop which shows up the growth of crops which is getting
several ways, such as urban monkey annoyance, wild pig yielded and gives the information through IOT techniques
crop raiding, and so on [2]. They lose their whole life's hard and machine learn ing. The drawback is lack of the technique
work in a matter of seconds if they are attacked by an detection due to a smaller nu mber o f datasets. The
animal, making it incredib ly difficu lt for them to return to advantage is they use dynamic datasets which gives real
normalcy. Many efforts are being developed to help ease the time values.

978-1-6654-9707-7/22/$31.00 ©2022 IEEE 536

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on May 22,2024 at 09:45:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Inventive Research in Computing Applications (ICIRCA 2022)
IEEE Xplore Part Number: CFP22N67-ART; ISBN: 978-1-6654-9707-7

K Balakrishna et al [27] proposed the concept of ML animals into general and different categories. Here they use
and IOT based technique which prevents animal intrusion as Multi-part Convolutional Neu ral netwo rk with Fisher Vector
wild animals cause major d isaster in habited areas. The based Stacked Auto Encoder and Animal Breeds are
raspberry-pi is used to run using machine learning separated and categorised into different classes. In Existing
algorith ms which is inherited with Wireless Fidelity module, approach electricity powered fences are used to defend
picamera, led and buzzer. They use CNN and M L plants from the animals. The major defect of this approach
algorith ms wh ich plays a major role for detection. The is due to high power supply the animals get electrocuted not
experiment reveals that Single Shot Detection algorith m only wild animals the domestic animals get affected.
Twilio API have shown up software sends the information
to the user to take further actions in the areas. Sheela et al. Bhambri et al. [29] proposed a detection system using
[8] proposed their idea o f A lert system using IOT devices. supervised learning approach and convolutional neural
The proposed idea of PIR tower sensor which senses the network (CNN). Their mot ive is detecting the suspicious
mo ments of wild an imals in habited areas and USB wh ich is human activity system. They have train network and the
connected passes the alert via internet to the user. To moves for detection classifier. here the trained network and
process images “OpenCV” is proposed on raspberry pi detection classifier takes up the data and check for
electric lines solar power is used to reduce the cost. tower suspicious activities in the arena and classifies it according
with sensors the solar panel will collect sunlight and use to to the algorith m and not only hu man activit ies it also is used
charge battery and power sensor. power consumption is low to detect the environmental causes.
and stores battery power all the time. human-an imal
interaction can be hazardous to both species, an intelligent Chen et al. [30] proposed the concept of Internet of
surveillance and alarm system is required. The flir e4 was Things in which latest technologies are used here PIR
chosen for this project since it is a single camera with a sensors, Raspberry pi3, Webcams are used, in this raspberry
reasonable cost. pi3 relates to PIR sensors and a webcam which detects the
suspicious activities a facial data base is stored which
Jeevitha et al. proposed [10] The develop ment of recognizes face which goes through pre-processing,
sensors has been brought to a whole new level thanks to recognizing, feature extraction and classifying at the end.
image processing and IoT sensor monitoring network Alarms are set up with email authorized. Here if the
technologies. Conflicts between wild animals and hu mans suspicious activity is detected alarm turns up through this
have become a huge issue in Habited zones, risking human paper, we can learn the usage of motion detection which can
lives and causing a massive loss of resources. Xue et al. be used for animal detection at an arena.
proposed [11] UW B technology which is a wireless sensor
used to detect movements over habited areas, forests,  The electric powered fences are used for stopping the
agriculture lands such as wildlife intrusion. evaluating the plants however in present approach digital changed into
properties of Ultra -wide band (UW B) signals, convolutional used for detecting the animals that's economically
neural networks are utilized to automat ically learn their excessive price.
features. Finally, to distinguish people fro m animals, the  The indication is to be had with inside the machine
SVM classifier is used [28]. however it sends the message handiest to the woodland
officer now no longer to the leav ing humans with ins ide
Nirit Datta et al. [6] show how to utilize automated the farmland.
monitoring and alarm systems to alleviate the concerns of  These charges very excessive whilst it's far achieved
wild an imals moving out of wildlife areas and causing with massive farms as it isn't always viable for hu mans
human and animal injury and fatalit ies. Automatic machine to display animal actions constantly at some point of the
and alert machine incorporating GSM and GPS era have day.
been applied inside the shape of a tool that could be
connected to the frame of an animal. Mark O. A folabi. The assault of animals with inside the farm or robbery
recommended an alarm system that primarily serves to scare of plants via way of means of people reasons a heavy loss in
ruffians off the field. If the ru ffian is around for an extended cultivation. Wild life intrusion in regions with excessive
amount of time, a GSM message is sent to the glazier, human mobility is deadly for every human and animal. The
informing him that there is a ruffian or fowl on the field. surveillance and monitoring of the wild animals are tough
The metallic sensor, on the other hand, delivers information because of length and nature of their motion. Due to loss of
to individuals who enter the field. detection machine ends in assaults kill v illagers and ruin
their plants
Bhagyashree et al [12] have used the algorith m that
resulted was stored and tested on a validation set. The III. PROPOSED M ODEL
detection process took an average of 1.5 seconds per image, In this part, we use Alex net Architecture which is an
during which time the photos were simply fed through the object classification with Convolutional neural network also
trained algorith m. The algorith m that scored best on Support vector machine is used for classification. we'll go
evaluating set was then tested on a test set that included over the model's steps how the idea detects and recognize
complete pictures as well. Divya Meena et al. [15] proposed the wildlife. The model pre-processes the pictures using
the idea of A Multi-part Convolutional Neural Network different strategies in the first step, which is image pre-
algorith m classifies around 40,000 animal photos from the processing and noise removal [9]. Pseudo-labels are used to
dataset into 30 different classes. The algorithm div ides retrain animals in classes with greater misclassifications.

978-1-6654-9707-7/22/$31.00 ©2022 IEEE 537

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on May 22,2024 at 09:45:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Inventive Research in Computing Applications (ICIRCA 2022)
IEEE Xplore Part Number: CFP22N67-ART; ISBN: 978-1-6654-9707-7

The following level involves detection and categorization, is used to compensate for this flaw. The YOLO -v5 method
which is then examined. The second stage is animal enriches the training data wh ile transmitting each batch train
recognition, in which the animal's class is identified and the data through data loader [20]. Scaling, colo r space
relevant actions are taken. SVM classifies the object where correction, and mosaic enhancement are three types of data
we use hyperplane to separate data points with respect to improvements that the data loader can execute. It also gives
class labels then we supply a test input. We substitute these the YOLO-v 5 algorith m a lot mo re flexib ility when it co mes
inputs in hyperplane’s equation and based on result we to picture sizes. This system can detect and recognizing tiny
identify which region the point belongs to. We optimize the things in remote sensing photos quickly and accurately.
hyperplane in such a way and maximize the distance to the
closest datapoints from each class . YOLO takes an entirely different approach. YOLO is a
smart convolutional neural network (CNN) that detects
A. Stage 1: Segmentation
objects in real time. The approach applies a single neural
Animals are generally sighted around woodlands. In network to the entire picture, then separates it into regions
images, animals are frequently accompanied by flora. As a and
result, just animals must be ext racted while the backdrop is
removed. The image must be split to prevent considering
background elements instead of the targeted animal section.
Segmentation process are currently semi-automatic,
requiring the operator to identify background and
foreground zones [22]. In this system, the user must mark
locations to distinguish between regions; nevertheless, once
regions have been picked, the system performs all further
processes automatically without requiring any more
informat ion. Using a segmented animal picture instead of
the original image imp roves the system's performance. To
distinguish the animal p ictures fro m background, we employ
the Iterated Graph Cuts Algorith m. A collection of
foreground or background seeds fro m user labels the first
labelling f0 of graph cuts. Foreground seed areas are defined
by pixels indicated as foreground, whereas background seed
regions are defined by pixels labelled as background [19].
Iteratively, the selected areas to the flagged area are watched
out and modernize subgraph, starting with the original sub -
graph. Looking through the updated subgraph all the area
are covered, and these nodes are tagged for yield
optimization segmentation wh ich will be classified either
foreground or background.

B. Stage 2: Animal Detection

YOLO v5: The intention of an object detection

algorith m is to predict the coordinates of the bounding box. Fig 1: YOLO Architecture
An image supplied to the yolo model will be div ided into n
grids. Yolo-V model utilizes these grids to predict the predicts bounding boxes and probabilit ies for each region.
bounding box. YOLO stands for " You On ly Look Once". These bounding boxes are weighted by the expected
The model is classified into 3 parts. They are the backbone, probability. It delivers remarkable precision while running in
neck, and head. The backbone is a feature extract ion layer. real time. In the sense that it only needs one forward
This Layer acts as the key feature of the image. The neck propagation run through the neural network to create
layer's functionality is to generate feature pyramids by predictions, the algorithm "only looks once" at the picture.
taking the key features identified with the backbone layer. Following non-max suppression (which ensures that the
Feature pyramids are utilized to detect images at different object detection algorithm identifies each object only once),
scales. Then comes the last layer which is the Head layer. it produces identified objects together with bounding boxes.
This layer collects the feature pyramids fro m neck layer and
then uses it to predict the bounding box coordinates, With YOLO, a single CNN predicts mult iple bounding
probability of the predicted class and object score. boxes and class probabilities for those boxes at the same
time. YOLO trains on entire pictures and immediately
The YOLO algorith m is capable of detecting targets enhances detection performance. This approach offers
quickly; however, it is incapable of detecting tiny targets or various advantages over conventional object detection
has a poor detection effect. The reason for this is because methods:
without proper grid d ivision, many targets are likely to • YOLO is quick.
appear in the same grid. As a result, the YOLO -v5 method

978-1-6654-9707-7/22/$31.00 ©2022 IEEE 538

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on May 22,2024 at 09:45:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Inventive Research in Computing Applications (ICIRCA 2022)
IEEE Xplore Part Number: CFP22N67-ART; ISBN: 978-1-6654-9707-7

• YOLO sees the entire image during training and testing, so IV. RESULT AND DISCUSSION
it imp licitly encodes contextual information about classes as We used the training data to train the model and then
well as their appearance. test its performance on both the training and validation sets
• YOLO learns generalizable representations of objects, so (evaluation metric is accuracy). The validation accuracy is
when trained on natural images and tested on artwork, the 60 percent, whereas the training accuracy is 92.5% percent.
algorithm outperforms other top detection methods. The train ing loss is higher because you've made it
artificially harder for the network to give the right
predictions. However, during validation all the units are
available, so the network has its full computational power.

T ABLE 1: PERFORMANCE EVALUAT ION

S. Train Data Test Accuracy Loss

No Data
1 20% 80% 77.4 0.5764

2 40% 60% 80.1 0.5702

3 50% 50% 85.4 0.5061

4 60% 40% 89.7 0.2962

5 80% 20% 93.3 0.2014

Increasing the training data always provides

Fig 2: Alex Net Architecture informat ion and should improve fit. It reduces
generalization error by train ing your model on more
Alexnet: The Alexnet is divided into eight tiers which samples. In this scenario, when the training data is 20% and
consists of five convolutional layers and three fully linked testing data is 80%, the accuracy and loss are 77.4 and
layers, one with its own set of learnable parameters. Except 0.5764 respectively. Whereas when the training data is 40%
for the output layer, wh ich utilizes a mix of pooling layer and testing data is 60%, accuracy of 80.1 and loss of 0.5702
and three fully - connected, the model consists of five are observed. As we have tested the model with various
layers, in which each uses Relu activation. They observed training and testing data, we observed that we get mo re
that by using the relu as an activation function, researchers accuracy when training data is 80% and testing data is 20%.
were able to accelerate the training phase by roughly six
times. Dropout strata were used to reduce dimensionality in
their model. The model is also developed using the
ImageNet dataset. There are around 14 million photographs
in the ImageNet database, that are split into a thousand
categories [23].

C. Comparison of Simple Model and Hybrid Model

In normal case we use Convolutional neural network
and then we will extract features will be seeded to dense
layers for classification we can modify this solution problem
by first extracting the features using Convolutional Neural
Network [CNN] and then using machine learning
classification models [SVM] in our case. By taking the Fig 3: Training vs Validation Accuracy
extracted features as the input of the machine learning
classification models and it can pred ict the corresponding
class labels using the extracted features. it was observed that
more accuracy is given by hybrid model wh ich we have
discussed above than the normal model. So, in hybrid model
we classify the final class label with the help of the machine
learning model and the ext racted features since we are using
Convolutional neural network to ext ract features and
Support Vector Machine to classify the ext racted features,
we call this a hybrid model.

978-1-6654-9707-7/22/$31.00 ©2022 IEEE 539

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on May 22,2024 at 09:45:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Inventive Research in Computing Applications (ICIRCA 2022)
IEEE Xplore Part Number: CFP22N67-ART; ISBN: 978-1-6654-9707-7

and the animals included in the paper are cows, Billy goats,
and Domestic sheep are among the livestock. Wild sheep,
Mountain sheep, and wild goat are all caprine an imals. Wild
Horse, Mare, Mounts, Stallion, African elephants, and
Indian elephants are example.

A. Dataset

We used ImageNet and Google photos for the tests. A

FLIR thermal imaging camera has also been used to take a
few photographs (both visible and thermal). The visib le
pictures collected with the FLIR thermal camera were used
to create the test cases. ImageNet [47], is a collection of
image data created wh ich undergoes through testing and
training. The data is set up to the Word-Net pecking order.
Fig 4: Training vs Validation Accuracy Each Synset comprises an average of over 1000 photos and
is annotated by humans. For our model, we collected 35992
The suggested approach discovers potential artworks by photos (a sample is d isplayed in Fig. 11) fro m 27 different
extracting deep properties. CNN considers depth to be one species in seven categories (listed in Table 2). Work
of the most essential elements. The deeper the network, the animals, Livestock, do mesticated animals, caprine animals
more characteristics may be ext racted. We ran a few tests on and are the seven groups of animals. Milch co ws, Billy
both datasets to see how the depth of the network affected goats, Nanny goats, and Domestic sheep are among the
the results. The variety of features were set to 100, and we livestock. Wild sheep, Mountain sheep, and wild goat are all
tested about five variety of CNN depths. On two datasets, caprine animals. W ild Ho rse, Mare, Mounts, Stallion,
shows the CNN depth effect. Deeper layer characteristics African elephants and Indian elephants are examp les of
are more robust and have greater representational capacity. work animals.

Fig.6. Sample images from ImageNet

B. System Requirements

Fig.5. Animal Detection Results The tests were conducted on a single Dell Inspiron 15
Laptop,11th gen Intel Core ,1 6GB of RAM, core(TM) i7
To have a better understanding of our model, we'll look Processor. We utilized a one-terabyte Seagate hard disc
at the confusion matrix. characterize a classificat ion model's drive for storage. Anaconda, a Windows -based
performance on test data considering the real values are environment, was used to execute. The neural network and
familiar. Here the classifier pred icted more than 2000 Hybrid Classifier of YOLOV5 and CNN were ran using
images in each set and being tested according to their Jupyter Notebook (6.4.8).
species. We can generate five distinct metrics to assess the C. Experiments
correctness of our model using our confusion matrix.
Around 15 types of animals are classified and trained to the
model and their accuracy after the tested model was 92%

978-1-6654-9707-7/22/$31.00 ©2022 IEEE 540

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on May 22,2024 at 09:45:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Inventive Research in Computing Applications (ICIRCA 2022)
IEEE Xplore Part Number: CFP22N67-ART; ISBN: 978-1-6654-9707-7

The entire dataset was extensively investigated using

various tasks, and the impacts of our suggested approach The Table.2 shows the performance of proposed
were documented. The tasks completed were as follows: approach against exiting system. In task1, we have
implemented images containing single an imal which were
 What can be the accuracy when different classes observed by the neural network and executed at accuracy
have images containing single animal to detect? between 95% to 100%. In this study, we have used about
The maximu m that we attain of 92% accuracy when the 2000 t rain ing images for the t rain ing and testing of different
dataset image contains single animal. It detects the image classes of animals. In task2, we have implemented images
and pre-process through and predicts with good accuracy. containing multip le classes of animals which were observed
 How CNN architecture is implemented with Alex by the neural network executed at accuracy between 80% to
net and YOLOv5? 85%. We have used 500 training images for imp lementing
mu ltip le classes of animals. In task3, Alexnet(CNN Neural
The advantage of using an object detection algorith m Network) incorporates with 8 layers. YOLOv 5 is best for
(YOLOv 5 in our case) is that it could isolate the animal object detection and Alexnet Architecture is the leading
image fro m the g iven image first and then use only that part architecture for classification in which we can load the pre-
of the image for feature extraction. Alexnet is one o f the trained version of a network wh ich was trained on more than
standard architectures used for image classification. So, the millions of images, wh ich is an unsupervised learning
images in the dataset could first be fed to the YOLOv 5 approach
model, and then just that area could be g iven to the image
classification model. This way the architecture won’t be V. CONCLUSION AND FUTURE SCOPE
required to learn useless features from the image because it
focuses more the region of interest. In co mparison to current TensorFlow is an open-source machine learning
neural network designs, how well does the suggested model framework that may be used to create machine learning and
perform. Using the image classification model alone gave a deep learning applications. It co mes with a pre-t rained and
training accuracy of 85 %. But after subjecting the dataset to optimized photo classificat ion algorith m that can handle
YOLOv5 model and then using the image classification many things. It can recognize hundreds of different objects,
model gave an accuracy of 92 %..The perfo rmance of including people, act ivities, an imals, plants, and locations. It
proposed system is measured by using (1)-(4) will be pre-trained with numerous image classes initially.
Then it'll be mostly responsible for p redicting what an
(1) image sy mbolizes. We used 35,992 train ing photos to
categorize 27 an imal types in this article. In conclusion,
(2) using Tensor Flo w, we were ab le to categorize different
(3) types of animals with an accuracy of 92.5 percent. We also
contributed on lowering the misclassification rate using
(4)
Hybrid classifier to t rain dataset of misclassified attributes,
which was created by co mbin ing the decisions of Yo lov5
CNN uses convolution layers to extract features fro m
and CNN classifiers. This method has boosted the total
input training examp les. Convolution and pooling layers
model's training accuracy to roughly 92.5 percent. We use
perform feature extraction. For examp le, g iven an image,
YOLOV5, which is designed specifically for object
the convolution layer recognizes features such as two eyes,
identification, for fine-grained animal breed categorization.
long ears, four legs, and a short tail. A fully connected layer
acts as a classifier for these features, assigning
Therefore, the study uses Hybrid classifier to detect and
the probability that the input image is a dog.
identify Wild animals. We're particu larly inclined in
The first layer recognizes the edges, and the second layer
collaborating with an animal incursion detection system to
combines them to recognize the shape. The next layer
monitor the animal's presence in a residential area and
combines this information and concludes that it is the nose.
inform homeo wners. To expand the stiffness of the train
For clarity, CNN doesn't know what the nose is. By looking
dataset, we want to include photographs of animal in places
at many of them in photographs, he learns to recognize
other than residential areas in public p laces under varying
this as a feature. Fu lly connected layers learn how
lighting conditions as part of our future work.
to properly classify images using these convolution
generation features. REFERENCES
[1] G. Fang, Q.Song, ”Legislation advancement of one health in
T ABLE.2 PERFORMANCE COMPARISON BET WEEN PROPOSED
China in the context of the COVID-19 pandemic: From the
APPROACH AND EXIT ING SYST EM
perspective of the wild animal conservation law”. One
Approach Accuracy Health, 12, 100195.,2021
[2] R.Sabeenian, N.Deivanai, & B.Mythili,,”Wild Animals
WaterShed[9] 54.2% Intrusion Detection using Deep Learning Techniques”, Int. J.
KDD Cup 99 92% Pharm. Res, 12, 2020,1054-1058.
[3] D. Sarala, “Machine learning based wild animal intrusion
[5] detection system in field”, Journal of Research
CNN with 90.2% Proceedings, 1(1), 73-85,2021.
SVM [14] [4] S. Taheri, & O.Toygar, “ Animal classification using facial
Proposed 92.5% images with score-level fusion”, IET computer vision, 12(5),
679-685,2018.
Approach

978-1-6654-9707-7/22/$31.00 ©2022 IEEE 541

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on May 22,2024 at 09:45:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Inventive Research in Computing Applications (ICIRCA 2022)
IEEE Xplore Part Number: CFP22N67-ART; ISBN: 978-1-6654-9707-7

[5] N.Khare,P.Devan,C.L.Chowdhary,,S.Bhattacharya, G.Singh, [25] Machado, N. L., Abbott, S. B., Resch, J. M., Zhu, L., Arrigoni,
S. Singh, & B.Yoon, B,”SMO-DNN: spider monkey E., Lowell, B. B., ... & Saper, C. B. (2018). A glutamatergic
optimization and deep neural network hybrid classifier model hypothalamomedullary circuit mediates thermogenesis, but not
for intrusion detection”, Electronics, 9(4), 692,2020 heat conservation, during stress-induced hyperthermia. Current
[6] N.Datta, S.Sarkar,,A.Malik,& D.Punetha,“ Automatic tracking Biology, 28(14), 2291-2301.
and alarm system for eradication of wild life injury and [26] Kim, S. H., & Lim, S. C. (2018). Intelligent intrusion detection
mortality”,. In 2016 2nd International Conference on Advances system featuring a virtual fence, active intruder detection,
in Computing, Communication, & Automation classification, tracking, and action recognition. Annals of
(ICACCA)(Fall) (pp. 1-4). IEEE.2016 Nuclear Energy, 112, 845-855.
[7] Trnovszky, T., Kamencay, P., Orjesek, R., Benco, M., & [27] Siddhant Kumar, Gourav Chowdhary, Venkanna Udutalapally,
Sykora, P. (2017). Animal recognition system based on Debanjan Das and Saraju P. Mohanty, "gCrop: Internet -of-Leaf-
convolutional neural network. Advances in Electrical and Things (IoLT) for Monitoring of the Growth of Crops in Smart
Electronic Engineering, 15(3), 517-525. Agriculture", IEEE International Symposium on Smart
[8] Sheela, Shivaram, Chaitra, Kshama, Sneha. K.G5 , Supriya. Electronic Systems (iSES) (Formerly iNiS), 2019.
K.S.-(2016) Low-Cost Alert System for Monitoring the Wildlife [28] KBalakrishna, FazilMohammed C.R.Ullas, C.M.Hema,
from Entering the Human Populated Areas Using IOT Devices. S.K.Sonakshi, Application of IOT and machine learning in crop
Department of TCE protection against animal intrusion IEEE International
[9] S.Radhakrishnan, R.Ramanathan, “ A support vector machine Symposium on Smart Electronic Systems (iSES) (Formerly
with Gabor features for animal intrusion detection in agriculture iNiS), 2021.
field”,. Procedia computer science, 143, 493-501,2018. [29] Pankaj Bhambri, Sachin Bagga , Dhanuka Priya , Harnoor Singh
[10] S.Jeevitha, & S.V.Kumar, “ A study on sensor based animal , Harleen Kaur Dhiman, Suspicious Human Activity Detection
intrusion alert system using image processing techniques. System, Journal of ISMAC (2020).
In 2019 Third International conference on I-SMAC (IoT in [30] Chen, Joy Iong Zong. "Smart Security System for Suspicious
Social, Mobile, Analytics and Cloud)(I-SMAC) (pp. 20-23). Activity Detection in Volatile Areas." Journal of Information
IEEE,2019. T echnology 2, no. 01 (2020): 64-72
[11] W.Xue, T.Jiang, and J.Shi, J, “Animal intrusion detection
based on convolutional neural network”.,In 2017 17th
International Symposium on Communications and Information
T echnologies (ISCIT) (pp. 1-5). IEEE.
[12] S.R.Bhagyashree, “IJERT-A Review on the Method of
Diagnosing Alzheimer's Disease using Data
Mining”, International Journal of Engineering
Research, 3(3),2014.
[13] J.A.Eikelboom, J.Wind, J, E.van de Ven, L.M. Kenana,
Schroder, H.J.de Knegt,. & H.H.Prins, “Improving the precision
and accuracy of animal population estimates with aerial image
object detection”,.Methods in Ecology and Evolution, 10(11),
1875-1887,2019.
[14] N.Manohar, Y.H.Kumar, R.Rani, &G.H. Kumar,
“ Convolutional neural network with SVM for classification of
animal images”, In Emerging Research in Electronics,
Computer Science and T echnology (pp. 527-537). Springer,
Singapore,2019.
[15] S.D.Meena, & L.Agilandeeswari, L, “An efficient framework
for animal breeds classification using semi-supervised learning
and multi-part convolutional neural network (MP-CNN)”, IEEE
Access, 7, 151783-151802.
[16] Jeevitha, S., & Kumar, V. (2020). A review of animal intrusion
detection system. Int. J. Eng. Res. T echnol, 9(5), 1219-1221.
[17] Onwuka, I. E., Afolabi, M. O., John, I. O., & Idowu, A. (2018).
Design and implementation of farm monitoring and security
system. Int. J. Comput. Appl, 181(9), 10-15.
[18] Banupriya, N., Saranya, S., Swaminathan, R., Harikumar, S., &
Palanisamy, S. (2020). Animal detection using deep learning
algorithm. J Crit Rev, 7(1), 434-439.
[19] Yasir, M., & Ali, Y. (2021). Review on Real Time Background
Extraction: Models, Applications, Environments, Challenges
and Evaluation Approaches.
[20] Su, H., Kong, W., Jiang, K., Liu, D., Gong, X., Lin, B., ... &
Xu, C. (2021, October). Gold crucian carp identification based
on Siamese network. In International Conference on Image
Processing and Intelligent Control (IPIC 2021) (Vol. 11928, pp.
191-194). SPIE.
[21] Manohar, N., Sharath Kumar, Y. H., Kumar, G. H., & Rani, R.
(2019). Deep learning approach for classification of animal
videos. In Data Analytics and Learning (pp. 421-431). Springer,
Singapore.
[22] Manohar, N., Kumar, Y. S., & Kumar, G. H. (2018). An
Approach for the Development of Animal Tracking
System. International Journal of Computer Vision and Image
Processing (IJCVIP), 8(1), 15-31.
[23] Choudhury, S., Saikia, N., & Das, A. J. Recent Trends in
Learning Based T echniques for Human and Animal Detection.
[24] Sharma, P., Sirisha, C. K., Gururaj, S., & Padmavathi, C.
(2020). Neural Network Based Image Classification for Animal
Intrusion Detection System. International Journal of Progressive
Research in Science and Engineering, 1(4), 1-7.

978-1-6654-9707-7/22/$31.00 ©2022 IEEE 542

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on May 22,2024 at 09:45:17 UTC from IEEE Xplore. Restrictions apply.