2023 International Conference on Recent Trends in Electronics and Communication (ICRTEC)

A Review on Animal Detection and Classification

using Computer Vision Techniques: Scope for
Future Enhancement to Application
Balakrishna K Dhanushree V
Dept of ECE Dept of ECE
Maharaja Institute of Technology Mysore Maharaja Institute of Technology Mysore
Karnataka, India Karnataka, India
[email protected] [email protected]
ORCID - 0000-0002-5562-0106
Abstract – As the human population increases day by day and very much needed to study their behavior, to analyze
because of the leading lavish lifestyle, demand for space and the population demography and counting the population
resources grows between humans and wildlife animals. The of the different species was a very tedious task which is
interaction of humans and wildlife animals results in conflict, time-consuming with high manpower cost. But
which leads to injuries, permanent disabilities and loss of life
in humans and animals. In many cases, the result ends with
nowadays advance in technical aspects leads to giving
the threat of declining or extinction of abundant species in the prominent solution to detect and classify wild animals
animals. Computer vision technology-related solution in using computer vision technology [3].
animal detection and classification plays a critical role with
Computer vision is an interdisciplinary field that deals
the advantage of safety, inexpensive and convince adopting of
the advanced algorithm. But still in the infancy of achieving
with gaining high-level knowledge in understanding
the expected level of efficiency and performance of the and automating tasks in digital images, videos and
model, also to adopt for real-time application. So here the visual inputs through the perspective of engineering.
authors give an immersive way of direction for future work to Here humans teach computers to interpret the
have real-time application by integrating computer vision knowledge of human vision to see around the real world
technology with the Internet of Things (IoT) technology. to detect, track and monitor well-being remotely as
Keywords: Animal Detection, Artificial Intelligence, shown in Fig 1. Normally, computer vision technology
Computer Vision, Neural Network, Internet of Technology deals with the enhancement and understanding of much
(IoT): more complex topics such as object detection, facial
recognition, image segmentation, image classification
I. INTRODUCTION etc., through various algorithms [4]. The algorithms
India records the tenth largest forest cover geographical developed were written in high-level language over
area of 21.71% in the world and ranks third globally in general-purpose computer systems for applications in
the net gain of geographical forest area between the fields like wildlife, agriculture, healthcare, security,
years 2010 to 2020 according to ISFR (Indian State of automotive and everyday life. Algorithms were
Forest Report) 2021 [1]. Forest provides food and developed to extract keypoints and descriptors from the
shelter for more than 1000 species of classes such as images for feature extraction algorithms such as SIFT
predators, mammals, reptiles, amphibians, birds and (Scale In-variant Feature Transform), BRISK (Binary
various insects [2]. Animals living in the forest are Robust In-variant Scalable Keypoints), FREAK (Fast
adapted to their unique structure of forest ecosystem Retina Keypoint), SURF (Speeded-Up Robust Features)
and food habits such as herbivorous, carnivorous and or ORB (Oriented FAST and Rotated BRIEF). Which
omnivorous. To maintain ecological balance and a are often collectively called features in machine
healthy environment around the world protecting learning, whereas in deep learning this distinction
wildlife is most important, which is the precious god between keypoints and descriptors does not even exist
gift to our planet. In past, humans were living harmony [5]. Rather than extracting the typical features from the
with other living creatures balancing the ecosystem. input images, supplanted all these algorithms in many
Over the past few years, the dominance of humans over cases, especially when having enough data called CNN
other creatures has increased tangibly due to an increase (Convolutional Neural Network) algorithm. CNN
in human population, hunting, deforestation etc,. These extract features taking into account the information in
wild animals enter the forests nearby places in search of the training data often called data-driven feature
food and water, which leads to conflict between humans extractors. Depending on the segmentation like
and wild animals. Due to this hundreds of humans semantic, instance etc, the CNN algorithm was
suffer serious injuries or death and also loss of upgraded and called RCNN (Region-based
agricultural productivity in their farm fields. To avoid Convolutional Neural Network), Fast RCNN, Faster
this conflict, many steps have been taken by the RCNN, Mask RCNN, ResNet (Residual Network) etc,.
government like the translocation of animals to secured Nowadays, object detection has dominated by its
places, population size regulation, effect on climate performance and speed with architectures such as RPN
change etc,. Also, monitoring of wildlife animals is

979-8-3503-9619-5/23/$31.00 ©2023 IEEE

(Region Proposal Network), SSD (Single-Shot Detector), YOLO (You Only Look Once) etc, [6,7].

Fig 1: Human-Vision versus Computer-Vision system.

Moreover, the animal detection and classification conflict, which refers to negative interaction. This
approach is a challenging task due to its unlike body, conflict between them leads to injuries, permanent
face, distinct position, a contrast in light intensities and disabilities and loss of life to humans and also threatens
shape in appearance. The current development of neural to decline or extinction of abundant species in the
networks has given eminent scope to researchers in the animal kingdom. For the growth of sustainable
field of animal detection and classification because of development activities, human-animal coexistence
their incredible precision [8]. But still, this field is in its management is most important and plenty of measures
infancy in achieving the expected level with the advent have been taken by authorities but are still not sufficient
of modern technologies, which seems to have a lot of to address it. Even some of the attempts tried adopting
opportunities for researchers and attract young technology so far.
researchers. The motivation for the selection of this
b. Attacks on Farm field by animals
research area is first concerning the long-term viability
conflict between humans and animals. Second, About two-third of the working population are mainly
determining predator and non-predator animals aids of dependent on agriculture to earn their living. As the
significant threat to humans and their cultivation in the population growth is increasing every year the need for
farm field. Third, minimizing the rate of traffic agricultural products is also increasing which in turn
accidents occurring on road crosses, where animals motivates the people to occupy the forest area to expand
collide with automobiles. Fourth, remotely monitor and their agricultural field [10]. Farmer’s face plenty of
count the wildlife species for the survey of wildlife. challenges in the process of farming; apart from this
one of the major challenging tasks for farmers near
II. LITERATURE SURVEY
forest areas is being attacked by wild animals. Some of
a. Conflict between humans and animals
the wild animals like elephants, deer, boar, gaur, bears
As the human population increases and needs of lavish etc, cause serious damage to farm fields by grazing,
life, demand for space continues to grow between running and trampling over the crops. The wild animals
humans and animals [9]. Humans and animals are raid the farmlands and livestock in search of food
interacting and encroach into the natural resources of causing huge economic losses to the farmers. While for
the animal kingdom as result lead to human-animal the protection of their farmland, farmers use to fence
their boundaries with wires, electricity, poles etc which VGG16 (Visual Geometry Group) acts as a backbone
have not reached a satisfactory level. Even some of the for the architecture. The algorithm is trained using SGD
attempts tried adopting modern technology so far. (Stochastic Gradient Descent) having a momentum
value of 0.9, batch size of 16 and learning rate of 1e-3.
c. Animal-Automobile Collision
In the Faster R-CNN algorithm, ResNet101 acts as a
The animal-automobile collision (AAC) is a significant backend for the feature extraction. The algorithm is
threat to road safety, which not only causes a dangerous trained using an SGD optimizer having 15 epochs with
threat to animals but also causes damage to automobiles a batch size of 8, momentum value of 0.9 and learning
and fatalities to riders and passengers [11]. Many rate of 0.001. The result shows that SSD performs
national highways (NH) like NH6 and NH7 pass comparably better than the Faster R-CNN algorithm for
through the wildlife reserve areas, which are meant for bigger objects when the objects are smaller in size
the safety of animals like Tiger reserve, Elephant leading to false detection.
reserve, Beer reserve, Leopard reserve, Deer reserve,
In 2020, S Divya Meena and Agilandeeshwari
etc. According to survey conducted by the Traffic
Loganathan [14] used a Deep Learning-based Sparse
Injury Research Foundation, Government of India
Multi Discriminative-Neural Network (SMD-NN)
justifies that AAC is a major threat to achieving road
algorithm to detect animals and mitigate vehicle
safety measures in India, where the number of cases
collisions. The authors proposed an integrated feature
was higher during night-time compared to day-time due
efficient learning algorithm named as Sparse Network
to low visibility and rider get exhausted. Even plenty
of Winnows (SNOW) to extract the features optimized
measures have been taken by authority for the safety of
with FixResNeXt to detect the animals. The algorithm
wildlife such as restricting the private automobiles,
is tested for three benchmark sets of datasets collected
banning the automobiles during the night-time. Even
on Animals on the Web (AoW), COCO and Animals
some of the computer vision technology based attempts
with Attributes (AwA) of 10 different classes of
have done so far.
animals. The experimentation was carried out based on
d. Wildlife Census 40:40:20 with developed model, where 40% was for
testing and training purpose each and the remaining
Today about 14% of birds and 26% of mammals are on 20% of the datasets for validating it. Performance of the
brink of extinction. One of the reasons for this algorithm analyzed with considering computational
extinction is natural calamities and the major region one
time and complexity, where SNOW accuracy is
is humans. Human activities such as land development,
comparably better to Deep learning algorithm.
pouching and mining which lead to a harmful
environment have already brought endangered species In 2022, Maheswari, M et al., [15] used a customized
at an alarming rate [12]. One of the ways to protect this Deep Learning algorithm for autonomous and efficient
endangered species is by performing a census. The surveillance of wildlife and vehicles in national parks.
enumeration of particular species in particular areas The authors proposed a video object identifying
grouping them by age and sex can be done by direct multiclass (animals and vehicles) object detection
counting or indirect counting methods called wildlife model using a customized CNN algorithm. For
census. With the help of the census, one will come to experimentation, datasets were collected from Open
know how many animals are there in each species, Image Datasets (OID) from Google for the five classes
where and when they can be found. By the outcome of namely Lion, Tiger, Elephant, Car and Motorcycle. The
the census, one can predict which endangered species algorithm is trained using Deep CNN having 2700
and then they can be protected in the national park, epochs with a batch size of 64 and a learning rate of
protected area or wildlife refuge etc., where they can be 0.001. The result shows that accuracy for the developed
continuously monitored and their behavioral change can model is varying from 67% to 82%.
be identified.
In 2020, Anamika Dhillon and Gyanendra K. Verma.,
In 2020, Atri Saxena et al. [13] used a Deep Learning [16] used a Deep Learning algorithm for the application
algorithm to detect animals and avoid collisions causing of wild animal detection, small arms detection and
injuries and deaths to animals inroads. The authors human being detection. The authors examined various
proposed an object detection technique for the detection functioning CNN models such as LeNet-5, AlexNet, ZF
of animals using neural network architecture like Faster Net, GoogleNet, VGG, ResNet, ResNeXt, SENet,
R-CNN and SSD. For experimentation, collected DenseNet, Xception and PNAS/ENAS for object
datasets of 31,774 images (both domestic and wildlife) identification. Each model is studied considering the
considering 25 different classes using PASCAL Visual database and accuracy, which provides a path for future
Object Classes (VOC), Microsoft Common Objects in research work in the area.
Context (COCO) and Open Image Dataset. Here
Some of the recent noticeable works attempted in the
datasets used for training the model are 31,744 and for
field of object recognition or detection are mentioned in
evaluation 6,088, where evaluation metrics used were
Table 1 considering algorithm, datasets and
precision, recall, map (mean average precision) and IoU
performance.
(Intersection Over Union). In the SSD algorithm,
Table 1: Noticeable work
 work (year) Technique Dataset Accuracy

Google image of animals horse, dog, cow

H. Kalla., et al (2022) [17] NASNetMobile model Precise value of nearly 86.5%
& buffalo

DenseNet201, CNN and Two classes of animals of ten image

Naveen., et al (2022) [18] Precision value of nearly 98%
Symbolic classifier each.

CNN like RCNN, Faster

R. Gandhi., et al (2022) [19] Google images of animals Accuracy of around 85% to 94%
RCNN, YOLO & SSD

Publicly available dataset of animal’s Accuracy 99% and mAP (mean

Sayagavi, A.V et al (2020) [20] YOLO
elephant, zebra, giraffe, lion & cheetah. average precision) 48.1

Serengeti & Wisconsin dataset of Achieved accuracy of 89-94% for

Ibraheam, M., et al (2021) [21] RCNN & YOLO animal’s lion, zebra, buffalo, giraffe, fox, the dataset of 70% training, 15%
deer, and elephant. validation & 15% testing.

Kaggle datasets of animals buffalo,

Ukwuoma, C. C., et al (2022) [22] ResNet-50 & YOLOv3 mAP between 0.76 to 0.86.
elephant, rhino & zebra.

Nilgai and Dog images with acquisition

Achieved accuracy of 4.76% to
Radhakrishnan S., et al (2018) 2D Gabor filter & Support of 25 & 50, adding noise such as
100% for varying filters for testing
[23] Vector Machine Gaussian, salt & pepper and speckle, and
and training set.
changing orientation from -45˚ to +45˚.

Faster RCNN, Mask Microsoft Common Objects in Context

Gupta, S., et al (2020) [24] RCNN and vicinity (MSCOCO) dataset and dashcam videos Accuracy of 84.12% to 86.95%.
tracking of animal’s cow and dog.

Genetic algorithm and Extended Complex Scene Saliency

Chandrakar, R., et al (2021) [25] Accuracy of 98.69%
CNN Dataset (ECSSD)

Five classes of animals such as horse,

Balakrishna K., et al (2021) [26] RCNN & SSD zebra, cheetah, elephant & cow of Each mAP of 85.22% & 89.32%
60 images.

The state-of-the-art from above-discussed detection and a) Wildlife animal’s detection and classification for
classification algorithm has been achieved considerably, large classes of species in a scene of sensitive
but still lacks in achieving expectation level and is still parameters for the best and worst-case scenario
in the infancy of implementing the application in real- in real-time (both day and night conditions).
time. Some of the challenges in the view of animal b) Automatically detecting, classifying and
detection and classification adopting computer vision counting wildlife animals from a remote location
technology are partial occlusion, cluttered background, for the best and worst-case scenarios in real-time
pose variation, poor illumination, overexposure, fur (both day and night conditions).
detection and classification etc. Here this creates c) Decimating the information to the required end
interest and attracts eminent and young researchers to to take appropriate action through a
the field. Based on the study of relevant research work, communication protocol, so right measures can
here came up with a few problem statements that need be taken to avoid conflict between humans and
to be addressed in near future with the integration of wildlife animals.
other technologies to avoid conflict between humans
Here to achieve the above-stated problem statement
and wildlife animals.
through Artificial Intelligence-based detection and
III. FUTURE SCOPE classification models such as SSD and YOLO [28]. The
decimation of information can be achieved through IoT
Animal detection and classification in the view of
Application Programming Interfaces (APIs) such as
computer vision technology give an immersing way of
Twilio [27].
directions for future work by integrating other
technologies into it. One of the trending integrated SSD: Single shot multibox detection model is simple,
technologies with computer vision is the Internet of fast and widely used in Computer Vision technology. It
Things (IoT), which is called Artificial Intelligence for consists of a base network to extract features from the
the infrastructure of IoT (AIoT) [27]. Some of the input images followed by several multiscale feature
problem statement needs to be addressed based on the map blocks to generate an anchor box to detect the
AIoT technology are smaller objects. The anchor box predicts the bounding
box with its classes and offsets [5].
YOLO: You Only Look Once model is designed to [5]. Balakrishna, K., and Mahesh Rao. "Tomato plant leaves disease
classification using KNN and PNN." International Journal of
work for real-time applications, where the models like
Computer Vision and Image Processing (IJCVIP) 9.1 (2019):
R-CNN, Faster RCNN etc, take 2 – 5 seconds to predict 51-63.
the objects. But the YOLO network predicts the objects [6]. Balakrishna, K., and N. Rajesh. "Design of Remote Monitored
for only one forward pass through the network. The Solar Powered Grasscutter Robot with Obstacle Avoidance
using IoT." Global Transitions Proceedings (2022).
model uses Leaky Rectified Linear Unit (ReLU) as its
[7]. Balakrishna, K. "WSN-Based Information Dissemination for
activation function and processes the frames at a rate of Optimizing Irrigation Through Prescriptive Farming.
45-150 frames per second, which is better in real time. " International Journal of Agricultural and Environmental
Information Systems (IJAEIS) 11.4 (2020): 41-54.
Twilio: Twilio provides programmable communication [8]. Parham, Jason, et al. "An animal detection pipeline for
APIs to channel communication functions using web identification." 2018 IEEE Winter Conference on Applications
services. IoT devices can be built, manage and of Computer Vision (WACV). IEEE, 2018.
[9]. Prat, Yosef. "Animals have no language, and humans are
connected without having setup and complex animals too." Perspectives on Psychological Science 14.5
infrastructure using the Twilio IoT. Here the secured (2019): 885-893.
and unbreakable connections are established to have [10]. Lamichhane, Babu Ram, et al. "Spatio-temporal patterns of
connectivity to authenticated persons to take attacks on human and economic losses from wildlife in Chitwan
National Park, Nepal." PloS one 13.4 (2018): e0195373.
appropriate actions at the user end. [11]. K. Balakrishna, B. R. Ganesh Prasad, N. D. Dhanyashree, V.
Balaji and N. M. Krishna, "IoT based Class Attendance
CONCLUSION Monitoring System using RFID and GSM," 2021 IEEE
Computer vision technology seems to be a promising International Conference on Mobile Networks and Wireless
Communications (ICMNWC), 2021, pp. 1-5, doi:
technology to maintain ecological balance and a healthy 10.1109/ICMNWC52512.2021.9688482.
environment around the world protecting wildlife, [12]. Balakrishna, K., and N. G. Sandesh. "Design of Dynamic
which is the precious god gift to our planet. Many Induction Charging Vehicle for Glimpse of Future: Cutting
computer vision-related algorithms were developed by Down the Need for High-Capacity Batteries and Charging
Stations." Advances in VLSI, Signal Processing, Power
eminent researchers to better understand animal Electronics, IoT, Communication and Embedded Systems.
detection and classification such as SIFT, BRISK, Springer, Singapore, 2021. 197-204.
FREAK, ORB, CNN, RCNN, Fast RCNN, Faster [13]. Saxena, Atri, Deepak Kumar Gupta, and Samayveer Singh. "An
RCNN, Mask RCNN, ResNet, RPN, SSD and YOLO. animal detection and collision avoidance system using deep
learning." Advances in Communication and Computational
The author’s main agenda is to avoid the conflicts Technology. Springer, Singapore, 2021. 1069-1084.
between humans and animals sharing a common [14]. Meena, S. Divya, and Agilandeeswari Loganathan. "Intelligent
boundary area, protecting the agricultural product and animal detection system using sparse multi discriminative-
counting wildlife animals. The overall study of the neural network (SMD-NN) to mitigate animal-vehicle
collision." Environmental Science and Pollution Research 27.31
paper gives a brief idea about the current development (2020): 39619-39634.
and way for future work to be carried out in the area of [15]. Maheswari, M., M. S. Josephine, and V. Jeyabalaraja.
animal detection for young and eminent researchers. "Customized deep neural network model for autonomous and
The fusion of IoT technology with Artificial efficient surveillance of wildlife in national parks." Computers
and Electrical Engineering 100 (2022): 107913.
Intelligence gives an immersive way of solution to the [16]. Dhillon, Anamika, and Gyanendra K. Verma. "Convolutional
real-time application. neural network: a review of models, methodologies and
applications to object detection." Progress in Artificial
Moreover, it’s a challenging task for eminent Intelligence 9.2 (2020): 85-112.
researchers to work on the normally captured images [17]. H. Kalla, B. Ruthramurthy, S. Mishra, G. Dengia and S. R, "A
dataset, which is usually collected during the day and Practical Animal Detection and Collision Avoidance System
Using Deep Learning Model," 2022 IEEE 7th International
night time with various behaviors, movements and conference for Convergence in Technology (I2CT), 2022, pp. 1-
motions of the animals. 6, doi: 10.1109/I2CT54291.2022.9824594.
[18]. N. M, M. R, N. B and R. R, "DenseNet201 for Animal detection
ACKNOWLEDGEMENT and repellent system," 2022 International Conference on
Electronic Systems and Intelligent Computing (ICESIC), 2022,
We thank each one of the staff of MIT Mysore and our pp. 213-218, doi: 10.1109/ICESIC53714.2022.9783582.
parents for directly and indirectly supporting in one or [19]. R. Gandhi, A. Gupta, A. K. Yadav and S. Rathee, "A Novel
other way for our research work. Approach of Object Detection using Deep Learning for Animal
Safety," 2022 12th International Conference on Cloud
REFERENCES Computing, Data Science & Engineering (Confluence), 2022,
pp. 573-577, doi: 10.1109/Confluence52989.2022.9734225.
[1]. Jayanthi, M., et al. "Spatial and temporal changes in mangrove [20]. Sayagavi, Ashwini V., T. S. B. Sudarshan, and Prashanth C.
cover across the protected and unprotected forests of Ravoor. "Deep Learning Methods for Animal Recognition and
India." Estuarine, Coastal and Shelf Science 213 (2018): 81-91. Tracking to Detect Intrusions." International Conference on
[2]. Atwood, Trisha B., et al. "Herbivores at the highest risk of Information and Communication Technology for Intelligent
extinction among mammals, birds, and reptiles." Science Systems. Springer, Singapore, 2020.
advances 6.32 (2020): eabb8458. [21]. Ibraheam, Mai, et al. "A performance comparison and
[3]. Høye, Toke T., et al. "Deep learning and computer vision will enhancement of animal species detection in images with various
transform entomology." Proceedings of the National Academy R-CNN models." AI 2.4 (2021): 552-577.
of Sciences 118.2 (2021): e2002545117. [22]. Ukwuoma, Chiagoziem C., et al. "Animal species detection and
[4]. Singh, Shilpi, and S. V. A. V. Prasad. "Techniques and classification framework based on modified multi-scale
challenges of face recognition: A critical review." Procedia attention mechanism and feature pyramid network." Scientific
computer science 143 (2018): 536-543. African 16 (2022): e01151.
[23]. Radhakrishnan, Saieshwar, and R. Ramanathan. "A support
vector machine with Gabor features for animal intrusion
detection in agriculture fields." Procedia computer science 143
(2018): 493-501.
[24]. Gupta, Savyasachi, Dhananjai Chand, and Ilaiah Kavati.
"Computer Vision based Animal Collision Avoidance
Framework for Autonomous Vehicles." International
Conference on Computer Vision and Image Processing.
Springer, Singapore, 2021.
[25]. Chandrakar, Ramakant, Rohit Raja, and Rohit Miri. "Animal
detection based on deep convolutional neural networks with
genetic segmentation." Multimedia Tools and
Applications 81.29 (2022): 42149-42162.
[26]. Balakrishna, K., et al. "Application of IOT and machine learning
in crop protection against animal intrusion." Global Transitions
Proceedings 2.2 (2021): 169-174.
[27]. Balakrishna, K. "WSN, APSim, and Communication Model-
Based Irrigation Optimization for Horticulture Crops in Real
Time." Artificial Intelligence and IoT-Based Technologies for
Sustainable Farming and Smart Agriculture. IGI Global, 2021.
243-254.