International Journal of Scientific Research in Science and Technology

Print ISSN: 2395-6011 | Online ISSN: 2395-602X (www.ijsrst.com)

doi : https://doi.org/10.32628/IJSRST

Wild Animal Detection System Using Deep Convolutional Neural Networks

Meenatchi K1, Thibishini V1, Vaisnavi K1, Mrs. R. Ahila2
1UG Scholar, Department of Computer Science and Engineering, School of Engineering, Avinashilingam

Institute for Home Science and Higher Education for Women, Coimbatore, Tamil Nadu, India
2Assistant Professor, Department of Computer Science and Engineering, School of Engineering, Avinashilingam
Institute for Home Science and Higher Education for Women, Coimbatore, Tamil Nadu, India

ABSTRACT

Article Info Animal detecting and monitoring has always been a challenging in research
Volume 8, Issue 3 area. Most of the animal detecting and monitoring processes rely on
Page Number : 989-996 commercial wild camera trap to take wild animal pictures which are triggered
by some sort of sensor techniques. The taken images still need human to collect
Publication Issue and get analysed with tremendous amount of effort. In a wild environment, the
May-June-2021 cost for deploying, collecting, analyzing is quite significant. In progress of AI
technique, there are mature tools that can be used to analyse the collected
Article History images. It can be utilized to solve the wild animal detecting and monitoring
Accepted : 18 June 2021 problem using Deep Convolutional Neural Networks. The idea is simple to run
Published : 27 June 2021 AI on Raspberry Pi locally to detect a wild animal and then it verifies the
images. Then it sends a message through GSM module with no need of internet
connection and gives an ultrasonic buzzer sound to divert a wild animal. It is
trying to propose an end –to-end solution which could potentially reduce the
loss of humans, animals and capitals using animal detecting system using Deep
convolutional Neural Networks.
Keywords: Artificial Intelligence, Deep Convolutional Neural Networks,
Raspberry PI, Retinex filtering, Softmax.

I. INTRODUCTION hinder the detection of animal resulting in low-

detection rates and high false discovery rates. To
Animal monitoring and analysis an active research handle this problem using a camera-trap database that
field [1] since last many decades. It focuses on animal has candidate animal proposals using multilevel graph
monitoring and analysis through animal detection cut in the spatio-temporal domain. These proposals
from natural scenes acquired by camera-trap are used to create a verification phase that identifies
networks. The image sequences obtained from whether a given patch is animal or background. It is
camera-trap consist of highly cluttered images that designed animal detection model using self-learned

Copyright: © the author(s), publisher and licensee Technoscience Academy. This is an open-access article distributed under the 989
terms of the Creative Commons Attribution Non-Commercial License, which permits unrestricted non-commercial use,
distribution, and reproduction in any medium, provided the original work is properly cited
 Meenatchi K et al Int J Sci Res Sci & Technol. May-June-2021, 8 (3) : 989-996

Deep Convolution Neural Network (DCNN) [3] features of positive images that have animals present
features. This efficient feature set is then used for in positive class, in addition of negative images that
classification using state-of-the-art machine learning do not have animals present in that class. But, this
algorithms, namely support vector machine, k-nearest approach has low performance for detection of visual
neighbour, and ensemble tree. The intensive results wild animals. Hence, in this paper, Multi-Cluster
show that a detection model using DCNN features Feature Selection (MCFS) is proposed for
provides accuracy of 91.4% on standard camera-trap unsupervised feature selection and wild animal
dataset. detection. Those features are chosen, which the
Visual recognition has been gaining popularity multi-cluster structure of the data is well preserved.
in biodiversity preservation and management. Since Based on spectral analysis approaches, the proposed
launching AI for Good initiative, this have been method suggests a principled manner for calculating
working with biodiversity researchers and the correlations among various features without label
practitioners [2] to deliver Animal image recognition information. Thus, the proposed technique handles
machine learning models and tools. In first foray, this the data with multiple cluster structure. The
area for Wild Detection which aligned with one of experimental results show that the proposed approach
the goals to use data science consulting to aid in the provides the better results.
preservation and management of the planet’s Animal Efficient and reliable monitoring of wild animals in
and environment. The goal was to build a model for the natural habitats is essential to inform conservation
visual recognition of specific kinds of animals. and management decisions. Auto convert cameras or
“camera traps” are being an increasingly popular tool
II. RELATED WORK for wild life monitoring due to unobtrusively,
continuously and in large volume. However
Computer vision techniques [3] are applied to processing such a large volume of images and videos
perform automatic wildlife surveying and animal captured from camera traps manually is extremely
monitoring. Animal detection in aerial videos is expensive, time consuming and also monotonous.
challenging because of the complexity of wild This presence a major obstacle to scientists and
environments. The method for moving animal ecologists to monitor wild life in an open
detection is proposed by taking advantage of global environment. Leveraging on recent advances in deep
patterns of pixel motion. In the video dataset, where learning techniques in computer vision, to build
animals make obvious movement against the automated animal recognition in the wild, aiming at
background, motion vectors of each pixel are an automated wild life monitoring system.
estimated by applying optical flow methods [4]. A Now a days, world has made computers [6] an
coarse segmentation then removes most parts of the inseparable part of their life as computers are used for
background by applying a pixel velocity threshold. performing the entire work of humans with better
Based on the segmented regions, another threshold accuracy and efficiency. Visual scene analysis is a
was employed to filter out negative candidates that high-level tasks that acquire knowledge from videos
could belong to the background. The pros and cons of or digital images that comes under the domain of
this method are discussed. computer vision. Object Detection is a field of
Wild animal detection helps wildlife researchers to computer vision and image processing which involves
analyze and study wild animal habitat and behaviour. detecting objects of varying class (animal, humans or
Discriminative Feature-oriented Dictionary Learning cars) present in images and videos. Some well
(DFDL) [5] was utilized for learning discriminative researched applications of object detection are in the

International Journal of Scientific Research in Science and Technology (www.ijsrst.com) | Volume 8 | Issue 3 990
 Meenatchi K et al Int J Sci Res Sci & Technol. May-June-2021, 8 (3) : 989-996

domain of car detection, face detection, image this a new approaches for study and a variety of
retrieval and video surveillance. This survey technologies/algorithms implemented in the past are
especially focuses on to examine the different images identified and appropriate ways for solving the
and videos based object detection methods to support research gaps are suggested to fill the gap.
various environments. The main objective of this Monitoring animals in the wild without disturbing
research is to study about different images and videos them is possible using camera trapping framework,
based object detection methods used for detecting and which is a technique to study wildlife using
solving images and videos based object detection automatically triggered cameras [9] and produces
problems. It provides detailed information about the great volumes of data. However, camera trapping
different object detection techniques in various collects images often result in low image quality and
environments. Finally, comparisons are made for includes a lot of false positives (images without
different object detection methods used in different animals), which must be detection before the post
images and videos environments. processing step. It presents a two-channelled
A novel method for object recognition based on perceiving residual pyramid networks (TPRPN) for
hybrid local descriptors is presented in [7]. This camera trap images objection. The TPRPN model [10]
method utilizes a combination of a few approaches attends to generating high-resolution and high-
(SIFT - Scale-invariant feature transform, SURF - quality results. In order to provide enough local
Speed up Robust Features) and consists of second information this extract depth cue from the original
parts. The applicability of the presented hybrid images and use two-channelled perceiving model as
methods are demonstrated on a few images from input for training the networks. Finally, the proposed
dataset. Dataset classes represent big animals situated three-layer residual blocks learn to merge all the
in Slovak country, namely wolf, fox, brown bear, deer information and generate full size detection results.
and wild boar. The presented method may be also Besides, it constructs a new high-quality dataset with
used in other areas of image classification and feature the help of Wildlife Thailand’s Community and
extraction. The experimental results show, that the enamel Organization. Experimental results on dataset
combination of local descriptors has a positive effect demonstrate the method is superior to the existing
for object recognition. object detection methods.
Animal detection-based study [8] is useful in many Recent studies in computer vision [11] have provided
real-life applications. Techniques involved in animal new solutions to real-world problems. This focus on
detection are useful in observing the locomotive using computer vision methods to assist in the study
behaviour of the engaged animal and in result it of kangaroos in the wild. In order to investigate the
prevent harmful interruption of animals in residential feasibility, to build a kangaroo image dataset from
areas. There are some branches of research in animal collected data from several national parks across the
detection. Some of these branches will therefore be State of Queensland. To achieve reasonable detection
discussed in this journal. Humans have developed accuracy, we explored a multipurpose approach and
many algorithms and techniques to gain a better proposed a framework based on the state-of-the-art
understanding of animal behaviour. For early Deformable Part Model (DPM). Experiments show
preventive measures, these technologies can also that the proposed framework outperformed the state-
serve as a warning system for humans from of-the-art methods on the proposed dataset. Also, the
encroachment of dangerous wild animals. Such tasks proposed vision tools are able to help our field
can be reduced to three main branches, namely biologists in studying kangaroo related problems such
animal detection, tracking and recognition. Through as population tracking for activity analysis.

International Journal of Scientific Research in Science and Technology (www.ijsrst.com) | Volume 8 | Issue 3 991
 Meenatchi K et al Int J Sci Res Sci & Technol. May-June-2021, 8 (3) : 989-996

III. SYSTEM MODEL monitor the position of more detailed animals by

installing more types of sensors as well as increasing
3.1 EXISTING MODEL the number of sensors.

This is design to build a wild animal pest repellent

3.2 PROPOSED MODEL
device with combination of passive infrared (PIR)
sensor and ultrasonic signal based on microcontroller In this work, it has designed a model that verifies the
as system controller. The PIR sensor is used to detect animal and background patches from the camera-trap
the presence of wild animal objects and ultrasonic images. The challenges associated with the model are
signals to interfere with the hearing. The design of the huge variations in background such as dynamic
the system is built based on microcontroller as the texture of background, change of position of
system controller. The system as a whole includes irrelevant objects (like leaf, branch), illumination
hardware and software. The design of hardware differences due to weather, season and shadows.
consists of the system design on the transmitter side Therefore, features must be invariant to all above
and the system design on the receiver side, the changes. Also, the model has to work with candidate
software in the system are algorithms using C animal patches that are of variable sizes and ratios
language programming. Findings – The resulting since they are obtained through ensemble graph cuts.
repellent device can detect animals approaching up to To handle above challenges, it present the following
a distance of 5 m and may interfere with its hearing scheme f+ or animal– background verification model.
with a 40 kHz ultrasonic frequency up to a distance of 3Our scheme has three steps: (1) pre-processing, (2)
20 m. fine-tuned DCNN features, and (3) classification
through learning algorithms.

Existing literature has shown that DCNN is very

efficient descriptors for object recognition,
classification and retrieval, etc. There are multiple
convolution layers and at least one fully connected
layer in a DCNN. For translation invariant features,
DCNN has pooling layer. Using the VGG-F pre
Figure 1 Existing Model trained model. The pretrained model has been
learned on huge auxiliary ILSVRC 2012 dataset. The
The system also uses remote monitoring devices using
pretrained model has an image size for input of 224 ×
433 MHz radio frequency up to a distance of 60 m.
224 hence, it resize the images to 224 × 224, without
Research Limitations/Implications – Each animal has
considering its actual size and ratio. The image resize
different hearing frequencies, as well as some wild
incurs image distortions which can be neglected due
animals, but the hearing frequencies of wild animals
to the fact that all the images go through the same
are generally at ultrasonic frequencies. The frequency
distortions, and the effect of resizing is negligible. The
of animal hearing may vary from audio frequency to
DCNN provides a feature vector of 1000 dimensions.
ultrasonic frequency, so ultrasonic wave emission
This use DCNN features as they are self-learned
testing with varying frequencies is required. Practical
features that enhance the performance of the system,
Implications – This research combines systems on
and these features contain information that describes
transmitters and receivers, with real-time monitoring
components of an image like edge, shape. The
of wild animal positions, and it can be possible to
architecture of the VGG-F model is described in

International Journal of Scientific Research in Science and Technology (www.ijsrst.com) | Volume 8 | Issue 3 992
 Meenatchi K et al Int J Sci Res Sci & Technol. May-June-2021, 8 (3) : 989-996

detail .The parameters of each layer are given as Deep Convolutional Neural Network (CNN) is a
convolution layer; number of filters with their size; special type of Neural Networks, which has shown
stride value; spatial padding and down-sampling exemplary performance on several competitions
factor of max-pooling. Stride tells the allocation of related to Computer Vision and Image Processing.
spatial dimensions over the input while padding tells Some of the exciting application areas of CNN include
the size of the padding along the borders of the input Image Classification and Segmentation, Object
in a convolution layer. Also, the pooling layer along Detection, Video Processing, Natural Language
with a convolution layer helps in reducing the size of Processing, and Speech Recognition. The powerful
the representations. Moreover, pooling aids in learning ability of deep CNN is primarily due to the
overcoming the problem of over fitting. Similarly for use of multiple feature extraction stages that can
fully connected layers, the dimensionality of each automatically learn representations from the data.
layer along with the method used for regularisation is The availability of a large amount of data and
given, and in last layer, soft-max classifier is used that improvement in the hardware technology has
evaluates the deviation of output to the target. accelerated the research in CNNs, and recently
interesting deep CNN architectures have been
IV. SYSTEM IMPLEMENTATION reported. Several inspiring ideas to bring
advancements in CNNs have been explored, such as
Wild animal detection is reliable and robust method the use of different activation and loss functions,
for animal detection in highly cluttered images using parameter optimization, regularization, and
DCNN. The cluttered images are obtained using architectural innovations. However, the significant
camera-trap networks. The images in camera-trap improvement in the representational capacity of the
image sequences also provide the candidate animal deep CNN is achieved through architectural
region proposals done by multilevel graph cut. It is innovations. Notably, the ideas of exploiting spatial
introduced by a verification step in which the and channel information, depth and width of
proposed region is classified into animal or architecture, and multi-path information processing
background classes, Thus, determining whether the have gained substantial attention. Similarly, the idea
proposed region is truly animal or not. We applied of using a block of layers as a structural unit is also
DCNN features to machine learning algorithm to gaining popularity. This survey thus focuses on the
achieve better performance. The experimental results intrinsic taxonomy present in the recently reported
shows that proposed system is efficient and robust deep CNN architectures and, consequently, classifies
wild animal detection system for both daytime and the recent innovations in CNN architectures into
night time. seven different categories. These seven categories are
based on spatial exploitation, depth, multi-path,
width, feature-map exploitation, channel boosting,
and attention. Additionally, the elementary
understanding of CNN components, current
challenges, and applications of CNN are also provided.

4.1 CNN COMPONENTS

Nowadays, CNN is considered as one of the most

widely used ML technique, especially in vision-
Figure 2 Block Diagram
related applications. CNN can learn representations

International Journal of Scientific Research in Science and Technology (www.ijsrst.com) | Volume 8 | Issue 3 993
 Meenatchi K et al Int J Sci Res Sci & Technol. May-June-2021, 8 (3) : 989-996

from the grid-like data, and recently it has shown 4.3 POOLING LAYER
substantial performance improvement in various ML
Feature motifs, which result as an output of
applications. Since CNN possesses both good feature
convolution operation, can occur at different
generation and discrimination ability, therefore in a
locations in the image. Once features are extracted, its
typical ML system, CNN capabilities are exploited for
exact location becomes less important as long as its
feature generation and classification. A typical CNN
approximate position relative to others is preserved.
architecture generally comprises alternate layers of
Pooling or down-sampling is an interesting local
convolution and pooling followed by one or more
operation. It sums up similar information in the
fully connected layers at the end. In some cases, a
neighborhood of the receptive field and outputs the
fully connected layer is replaced with a global average
dominant response within this local region. The use
pooling layer. In addition to different mapping
of pooling operation helps to extract a combination of
functions, different regulatory units such as batch
features, which are invariant to translational shifts
normalization and dropout are also incorporated to
and small distortions. Reduction in the size but also
optimize CNN performance. The arrangement of
helps in increasing the generalization by reducing
CNN components plays a fundamental role in
overfitting. Different types of pooling formulations
designing new architectures and thus achieving
such as max, average, L2, overlapping, spatial pyramid
enhanced performance. This section briefly discusses
pooling, etc. are used in CNN.
the role of these components in a CNN architecture.

4.4 TRAINING OF CNN

4.2 CONVOLUTION LAYER
An SGD optimizer was used to optimize CNN training.
The convolutional layer is composed of a set of
In the SGD optimizer, optimization is performed
convolutional kernels where each neuron acts as a
using a step policy that multiplies the gamma value
kernel. However, if the kernel is symmetric, the
for each fixed iteration so that the training accuracy
convolution operation becomes a correlation
and loss converge quickly. Training, a function of
operation. Convolutional kernel works by dividing
SGD, was performed in mini-batch size units. The
the image into small slices, commonly known as
number of iterations is calculated as ‘‘number of
receptive fields. The division of an image into small
training data / mini-batch size,’’ defined as 1 epoch. In
blocks helps in extracting feature motifs. Kernel
this experiment, learning rate is 0.0005, momentum is
convolves with the images using a specific set of
0.9, and gamma is 0.1, whereas the minibatch size in
weights by multiplying its elements with the
ResNet-50 is 16, 6 in ResNet-101, 3 in ResNet-152,
corresponding elements of the receptive field. Due to
and 20 in visual geometry group (VGG)-16, with a
weight sharing ability of convolutional operation,
maximum epoch of 10. Because fine-tuning was
different sets of features within an image can be
performed using the existing pre-trained weights, we
extracted by sliding kernel with the same set of
used small learning-rate values. After seven epochs,
weights on the image and thus makes CNN parameter
the learning rate is reduced. One epoch indicates that
efficient as compared to the fully connected networks.
training is performed as many times as the total
Convolution operation may further be categorized
number of iterations.
into different types based on the type and size of
Therefore, the total number of trainings is equal to
filters, type of padding, and the direction of
the number of iterations × the number of epochs. To
convolution.
calculate the training loss, the softmax function was
used to calculate multinomial logistic loss. When

International Journal of Scientific Research in Science and Technology (www.ijsrst.com) | Volume 8 | Issue 3 994
 Meenatchi K et al Int J Sci Res Sci & Technol. May-June-2021, 8 (3) : 989-996

training is performed, the accuracy converges to 100

and the loss converges to 0. This shows that the
training of the CNN model used in this study was
successful. We made the self-collected DMFW-DB1
and trained ResNet model available to other
researchers through for fair comparisons.

4.5 COMPARISON OF FINGER-WRINKLE

RECOGNITION PERFORMANCE ACCORDING TO Figure 4 Detection Output
COLOR SPACE

For the first experiment, we compared the

recognition performance of input images of gray and
various color spaces. In the following experiment, we
compared the recognition performance of Retinex
filtering and the original image. but it also increases
the distinctiveness of finger-wrinkle texture.
Moreover, the deep ResNet used in this study can Figure 5 Input-2 for animal identification
obtain sufficiently robust features to address the
illumination variation. In the next experiment, we
measured the recognition accuracy according to the
processing method of the background region.

V. RESULT AND ANALYSIS

DCNN Algorithm is processed and the animal image

is detected and classified by comparing the input Figure 6 Detection Output
image with existing dataset. To check the working of
the exact output the code is done and the sample VI. CONCLUSION
input is given as chair.
Deep Convolution Neural Network (DCNN)
algorithm is used to detect wild animals. DCNN
algorithm classifies animals efficiently with a good
number of accuracy and also the image of the
detected animal is displayed for a better result so that
it can be used for other purposes such as detecting
wild animals entering into human habitat and to
prevent wildlife poaching then gives an alarm sounds
and send message to farmers. Also, the future scope of
this system can be extended to various purposes like
roadside safety of traveller from wild animals in forest
Figure 3 Input-1 for animal identification
bound roads. In future along with identification of
wild animals, control measures like sedating based on

International Journal of Scientific Research in Science and Technology (www.ijsrst.com) | Volume 8 | Issue 3 995
 Meenatchi K et al Int J Sci Res Sci & Technol. May-June-2021, 8 (3) : 989-996

animal type can be implemented with more real time Identification of animal species in camera-trap
data and advanced sensors and systems which would images using very deep convolutional neural
make this technique completely automated without networks. Ecological informatics, 41, 24-32.
any manual interference. [8]. Xue, C., Wang, P., Zhao, J., Xu, A., & Guan, F.
(2017). Development and validation of a
VII. REFERENCES universal primer pair for the simultaneous
detection of eight animal species. Food
[1]. Fang, Y., Du, S., Abdoola, R., Djouani, K., & chemistry, 221, 790-796.
Richards, C. (2016). Motion based animal [9]. Xue, W., Jiang, T., & Shi, J. (2017, September).
detection in aerial videos.Procedia Computer Animal intrusion detection based on
Science, 92, 13-17. convolutional neural network.
[2]. Jasko, G., Giosan, I., &Nedevschi, S. (2017, InCommunications and Information
September). Animal detection from traffic Technologies (ISCIT), 2017 17th International
scenarios based on monocular color vision. In Symposium on (pp. 1-5). IEEE.
Intelligent Computer Communication and [10]. Zhang, T., Wiliem, A., Hemsony, G., & Lovell,
Processing (ICCP), 2017 13th IEEE B. C. (2015, April). Detecting kangaroos in the
International Conference on (pp. 363-368). wild: the first step towards automated animal
IEEE. surveillance. In ICASSP (pp. 1961-1965).
[3]. Nguyen, H., Maclagan, S. J., Nguyen, T. D., [11]. Zhu, C., Li, T. H., & Li, G. (2017, October).
Nguyen, T., Flemons, P., Andrews, K., ... & Towards automatic wild animal detection in
Phung, D. (2017, October). Animal recognition low quality camera-trap images using two-
and identification with deep convolution neural channeled perceiving residual pyramid
networks for automated Animal monitoring. In networks. In Computer Vision Workshop
Data Science and Advanced Analytics (DSAA), (ICCVW), 2017 IEEE International Conference
2017 IEEE International Conference on (pp. 40- on (pp. 2860-2864). IEEE.
49). IEEE.
[4]. Kumar. S., & Singh, S. K. (2016). Monitoring of
pet animal in smart cities using animal Cite this article as :
biometrics. Future Generation Computer
Systems. Meenatchi K, Thibishini V, Vaisnavi K, Mrs. R. Ahila,
[5]. Parham, J., Stewart, C., Crall, J., Rubenstein, D., "Wild Animal Detection System Using Deep
Holmberg, J., & Berger-Wolf, T. (2018, March). Convolutional Neural Networks ", International
An Animal Detection Pipeline for Journal of Scientific Research in Science and
Identification. In 2018 IEEE Winter Conference Technology (IJSRST), Online ISSN : 2395-602X, Print
on Applications of Computer Vision (WACV) ISSN : 2395-6011, Volume 8 Issue 3, pp. 989-996,
(pp. 1075- 1083). IEEE. May-June 2021.
[6]. Matuska, S., Hudec, R., Kamencay, P., Benco, Journal URL: https://ijsrst.com/IJSRST2183215
M., &Zachariasova, M. (2014). Classification of
wild animals based on SVM and local
descriptors. AASRI Procedia, 9, 25-30.
[7]. Villa, A. G., Salazar, A., & Vargas, F. (2017).
Towards automatic wild animal monitoring:

International Journal of Scientific Research in Science and Technology (www.ijsrst.com) | Volume 8 | Issue 3 996