A Major Project Final Report on

ALERT SYSTEM USING TENSORFLOW

‘WATCHMAN’
Submitted in Partial Fulfillment of the Requirements for
the Degree of Bachelors of Engineering in Computer Engineering
under Pokhara University

Submitted by:
Amrit Sunar,14302
Binita Rai,14711
Prakash Yadav,14331
Rabin Rai,14333
Suraj Suwal,14345

Under the supervision of

Mr. Chandan Gupta

Date:
06/12/2018

Department of Computer
NEPAL COLLEGE OF
INFORMATION TECHNOLOGY
Balkumari, Lalitpur, Nepal

i
 ACKNOWLEDGEMENT

First and foremost, we would like to thank our supervisor Er. Chandan Gupta for his guidance
throughout the project. We would also like to thank Er. Shiva Hari Acharya, Er. Saroj Bista
and Dr. Roshan Chitrakar for their proper guidance and support for counseling throughout our
project. Without their invaluable advice and assistance, it would have not been possible for us to
complete this part of our project on time.
Moreover, we would like to owe our special thanks to all subject teachers who provided us with
necessary information during the course of studies which somehow helped us in various ways for
the development of project. Similarly, we truly admire the effort of all persons who knowingly or
unknowingly provided us with the concepts about the project that assisted us for the development
of the system.

i
 ABSTRACT
Here, in this project we tried to implement the object detection and recognition to detect and
recognize the wild animals through the application of the TensorFlow library. Here we made a
device that can detect and recognize what kind of animal is in its frame and can generate the sound
to derive out them and also the device can generate alert sound to alert the peoples about these
wild animals.

The propose for developing “Alert system using TensorFlow ‘Watchman’” is to monitor the
possible threats of the wild animals on human beings to those villages that are close to the jungles
through the video surveillance. It facilitates the identification of possible threats/danger from
different animals and notify the humans about that threat. It also provides the facility to track down
what kind of possible loss may arise by animals on humans and their farming’s.

We proposed to implementation of the CNN based object detection algorithm on Respberry PI.
The model architecture is based on TensorFlow library to get image feature maps and to detect the
wild animals.

Keywords: Video surveillance, Object detection, Recognize, Respberry PI.

ii
TABLE OF CONTENTS

ACKNOWLEDGEMENT ............................................................................................................... i
ABSTRACT .................................................................................................................................... ii
1. INTRODUCTION ................................................................................................................... 1
2. PROBLEM STATEMENT...................................................................................................... 4
3. JUSTIFICATION .................................................................................................................... 5
4. SCOPE OF THE PROJECT AND LIMITATIONS................................................................ 6
5. PROJECT OBJECTIVE .......................................................................................................... 7
6. SIGNIFICANCE OF THE STUDY ........................................................................................ 8
7. FEASYBILITY STUDY ......................................................................................................... 9
8. LITERATURE STUDY ........................................................................................................ 10
8.1 EXISTING SYSTEM ..................................................................................................... 10
8.2 OBJECT DETECTION .................................................................................................. 13
8.3 BASICS KNOWLEDGE OF DEEP LEARNING ......................................................... 14
8.4 TENSORFLOW ............................................................................................................. 15
8.5 FASTER R-CNN ........................................................................................................... 16
8.6 CREATING A SET OF TRAINING IMAGES WITH TENSORFLOW...................... 18
9. METHODOLOGY ................................................................................................................ 19
9.1 SOFTWARE DEVELOPMENT PROCESS MODEL .................................................. 19
9.2 TOOLS AND TECHNOLOGIES USED ...................................................................... 23
9.3 SYSTEM DEVELOPMENT.......................................................................................... 25
9.3.1 INCREMENT-1: object detection .......................................................................... 32
9.3.2 INCREMENT-2: image training............................................................................. 38
9.3.3 INCREMENT-3: hardware implementation ........................................................... 43
9.3.4 INCREMENT-4: integration of the project ............................................................ 47
10. TASK DISTRIBUTION .................................................................................................... 49
11. DELIVERABLES .............................................................................................................. 52
12. PROJECT TASK AND TIME SCHEDULE ..................................................................... 53
13. CONCLUSION AND FUTURE SCOPE .......................................................................... 57

i
 13.1 CONCLUSION .............................................................................................................. 57
13.2 FUTURE SCOPE ........................................................................................................... 57
14. BIBLIOGRAPHY .............................................................................................................. 58

LIST OF TABLES
Table 1. 1: Patterns of human death and injury due to large-mammal attacks ............................... 2

Table 8. 1: showing animal detection systems at different places. ............................................... 12

Table 9. 1: showing hardware required......................................................................................... 23

Table 9. 2: showing software required .......................................................................................... 24
Table 9. 3: showing components required .................................................................................... 24
Table 9. 4: showing planning for increment-1 .............................................................................. 32
Table 9. 5: showing test results ..................................................................................................... 37
Table 9. 6: showing planning for increment-2 .............................................................................. 38
Table 9. 8: showing planning for increment-3 .............................................................................. 43
Table 9. 9: showing hardware cost estimations ............................................................................ 44

Table 10. 1: showing task division ............................................................................................... 49

ii
LIST OF FIGURES
Figure 1. 1: Frequency of attacks .................................................................................................... 3
Figure 1. 2: Incident prevalence and kill prevalence ...................................................................... 3

Figure 8. 1: showing object(animal) detection ............................................................................. 13

Figure 8. 2: Neural network showing its i/p, o/p and hidden layers ............................................. 14
Figure 8. 3: TensorFlow logo........................................................................................................ 15

Figure 9. 1: Incremental Model .................................................................................................... 20

Figure 9. 2: Block diagram for the project .................................................................................... 25
Figure 9. 3: flow chart for the system ........................................................................................... 26
Figure 9. 4: use case diagram for the system ................................................................................ 27
Figure 9. 5: showing activity diagram for the project ................................................................... 28
Figure 9. 6: showing class diagram............................................................................................... 29
Figure 9. 7: context level DFD ..................................................................................................... 30
Figure 9. 8: level-1 DFD for the system ....................................................................................... 31
Figure 9. 9: flowchart for animal detection ................................................................................. 34
Figure 9. 10: use case diagram for animal detection .................................................................... 35
Figure 9. 11: level-1 DFD for animal detection ............................................................................ 36
Figure 9. 12: flowchart for training animals/objects ..................................................................... 39
Figure 9. 13: use case diagram for training animals ..................................................................... 40
Figure 9. 14: level-1 DFD for training animal .............................................................................. 40
Figure 9. 15: training Log ............................................................................................................. 41
Figure 9. 16: Loss graph ............................................................................................................... 42
Figure 9. 17: loss vs steps graph ................................................................................................... 42
Figure 9. 18: Pi model 3 B+ .......................................................................................................... 44
Figure 9. 19: Pi camera module v.2 .............................................................................................. 45
Figure 9. 20: Arduino YUN .......................................................................................................... 45

Figure 12. 1: Showing planned duration for tasks of increment-1 ............................................... 53

Figure 12. 2: showing Gantt chart for increment-1....................................................................... 53
Figure 12. 3: Showing planned duration for tasks of increment-2 ............................................... 54
Figure 12. 4: showing Gantt chart for increment-2....................................................................... 54
Figure 12. 5: showing task schedule for increment-3 ................................................................... 55
Figure 12. 6: showing Gantt chart for increment-3....................................................................... 55
Figure 12. 7: showing task schedule for increment-4 ................................................................... 56
Figure 12. 8: showing Gantt chart for increment-4....................................................................... 56

iii
 1. INTRODUCTION
Wildlife is one of the best attraction of the Nepal. Many peoples come here to see these animals
and are still growing. Wildlife diversity is a notable feature of Nepal. Nepal has a large variety of
plants and animals because of the variance in climate, from tropical to arctic. Wildlife tourism is
one of the major source of tourism in Nepal. In Nepal a large number of rhododendron species can
be found. They are no natural habitats in Nepal. Nepal has established numerous National Parks
and reserves in order to protect its diverse fauna [1]. But there is also some negative part of these
wildlife animals. People are being attacked by large mammal species such as tigers, common
leopards, rhinoceros, elephants, bears, etc.

Rural area near the forest are venerable to wild animals. The danger might be for the human beings
or for their occupation. The wild animals like tiger, elephant, rhino, etc. are threat to human society.
Other animals like monkey, Boar, Deer, etc. are harmful and destroy Crops. So, we need to monitor
and protect ourselves and the farms from these animals. For this, we have to watch and monitor
the possible danger and threats by ourselves.

Here, we proposed the system that can continuously watch and monitor the possible dangers and
give some notifications to the peoples on the village with minimum involvement of humans. The
main aim of our project is to reduce the possible loss on human beings and their crops by
monitoring the wildlife near the village and giving the result information to the peoples. Here our
system exists in between the possible incoming path of the wild animals towards the village. The
system detects the motion first and then detects the type of animal through camera surveillances.
After that the system generate some sound to drive out these animals and it also alert the peoples
on the village.

Deep learning-based object detection has been very successful in recent years. Especially the CNN
(convolutional neural network) model has significantly improved the recognition accuracy on large
data-sets. For the ImageNet benchmark data set, the CNN based model has been dominating the
leader-board since it’s introduced by Krizhevsky in 2012 for the first time. [2]

1
The main motivation for us to go for this project was the continuous and inefficient traditional
manual monitoring system for threats from wild animals. This made us to think why not make it
automated, fast and mush efficient.

Injury and death from wildlife attacks may results in people feeling violent resentment and hostility
against the wildlife involved and, therefore, support for wildlife conservation may not be topic of
interest for public. Though Nepal, with rich biodiversity, is doing good in its conservation efforts,
human-wildlife conflicts have been a major challenge in recent years. We analyze patterns of
human injury and death caused by large mammals using data from attack events and their
spatiotemporal dimensions collected from a national survey of data available in Nepal over five
years (2010–2014). [3]

Table 1. 1: Patterns of human death and injury due to large-mammal attacks

Wildlife Contribution Avg. Avg. Avg. Avg.

(%) attacks/yra fatalities/ attacks/seasonb fatalities/seasonb
yra

Elephant 30 27.4±7.7 18±4.6 34.2±16.5 22.5±11.7

Leopard 21 19.4±11.6 8±5.4 24.2±3.8 10±6.6

Rhinoceros 18 17±4.3 3±1.2 21.2±16 3.7±3.5

Bear 12 11±4.3 1±1.2 13.7±2.6 1.2±1.2

Tiger 10 8.8±5.4 4.8±3.3 11±4.8 6±1.4

a
Observation Period= 5years

b
Number of seasons per year= 4 seasons

2
 frequency of attacks
100
90
80
70
60
50
40
30
20
10
0
Bear Elephant Leopard Rhinoceros Tiger Other

Injury Fatality

Figure 1. 1: Frequency of attacks

Incident and Kill prevalance

60

50

40

30

20

10

0
Bear Elephant Leopard Rhinoceros Tiger Other

Incident Prevalence Kill prevalence

Figure 1. 2: Incident prevalence and kill prevalence

3
 2. PROBLEM STATEMENT
The traditional manual methods of monitoring and protecting the human beings and their farming
from possible danger and threats from wild animals is hard and tedious as the continuous
monitoring is needed which have to fulfil by peoples themselves. This is hard, tedious, time
consuming and prone to inaccuracies as people may not have time to monitor and track the possible
threat continuously.

In Nepal, people are mainly attacked by large mammals such as tigers, common leopards,
rhinoceros, elephants, bears and others. Other than these animals like Deer, Boar, Monkey, etc.
harms the corps. So, we need continuous monitoring from these possible danger and threats.

Use of the object detection and recognition system in lieu of the traditional methods will provide
a fast and effective method of monitoring the possible threats continuously while offering an easy,
stable and robust service for tracking the possible danger and threat.

While developing the product, we have been faced various problems. The difficulty arose during
gathering the large image data and train those images as the main objective of the project is to
detect the animals from those images. Another problem that we have been faced is to integrate the
project. Here we also have difficulties to gather requirement for sound, from which animal can
drove out. Also, the system fails to differentiate the real object from the dummy object.

CNN based model can achieve higher accuracy, but they have following disadvantages:

High computation cost. In the CNN based model, each layer takes a lot of computation since are
usually very deep with tens or hundreds of layers.

large memory demand. In the CNN based model, it has a lot of parameters that usually take
hundreds of Megabytes of memory space.

Low efficiency. Most CNN based model are designed without efficiency improvement.

4
 3. JUSTIFICATION
This project serves to automate the prevalent traditional tedious and time-wasting methods of
monitoring and tracking wild animals. The use of “Alert system using TensorFlow ‘Watchman’”
will increase the effectiveness of monitoring & tracking the wild animals and alerting the peoples
about these animals.

This method could also be extended for use in similar fields such as animal tracking and counting’s
for the purpose of wild animal conservations. Applications of object detection and recognition are
widely spreading in different areas such as vehicle tracking and identification, face detection and
recognitions, OCR, etc.

5
 4. SCOPE OF THE PROJECT AND LIMITATIONS
The intention for developing this project is to computerize the traditional way of tracking and
monitoring the possible threats from wild animals. Another purpose of this project is to generate
the desired sounds to drive out the wild animals and alert peoples about those wilds.

The project is developed as a desktop application, and it will work for a particular area at a time.

This project serves to automate the prevalent traditional hard, non-efficient and continuous
methods of monitoring and alerting peoples from the danger from wild animals to humans and
their farms. The use of automatic watching and tracking the threats through object detection and
recognition will increase the effectiveness of threat monitoring, management and alert system.

This method could also be extended for use in detecting the wild animals that may harm the corps
and farming. Applications of object detection are widely spreading in areas such as autonomous
driving, face detection, video surveillance, security systems, image and film processing, etc.

Although this product has various advantaged but also have some limitations. Here, after the
product has been fully developed, we faced various problems. The system does not exactly detect
those wild animals which may harm. The leading cause for this problem are heavy humidity, heavy
rainfall, wildfire, distance, light sensitivity of the camera device etc. which are being problematic
for device to capture the image. The other problem we faced is the system does not able to drive
out / rout those harmful wild animals as expected by the system since it is not always possible to
predict the behavior of wild animals.

6
 5. PROJECT OBJECTIVE
The overall objective is to develop an automated Alert system using TensorFlow “watchman”
comprising of a desktop application working to perform the following tasks:

• To identify motion of wild animals near the village.

• To detect & recognize the wild animals through video surveillance using CNN based
algorithm.
• To generate the sound specific for the detected animals that drive out these animals and
also to alert the people.
• To analyze the architecture and algorithm that can help in other similar fields.

7
6. SIGNIFICANCE OF THE STUDY
• First and foremost, the system reduces the threat level from the wild animals faced by
people specially those living in the area near to the forest.
• This system also eliminates the traditional way of driving away the wild animals which
may harms the crops as well other property along with human damages.
• Wild animals are also very known for their aggressive behavior towards people. So, this
system provides security to the people by alerting them of the danger helping them to
maintain sound mental health.
• Due to the various damages done by wild animal, there have been many conflictions about
destroying them. So, this system helps to reduce these conflictions as it helps to control the
destructions of wild animals.
• As there have been many illegal hunting as well as poaching of animals, this system helps
to control these illegal activities.
• There has been great increment in the efficiency in the task of driving out animals as system
works much more efficiently than that is done manually which saves the time of people.
• In future we can implement this type of animal detection systems in between the roads
within the jungles, national parks etc. to avoid possible threats in human beings by these
animals and vice versa.

8
 7. FEASYBILITY STUDY

Economic Feasibility:

The system developed is economic with respect to National park, conservation authorities & office
and the government point of view. It is cost effective in the sense that has eliminated the manual
monitoring and tracking completely. Since it needs to be installed only one time and the system
will work until the system crash or power failure.

Time period:

The system is time effective because the continuous manual monitoring is eliminated. The system
minimizes the manual monitoring of the wild animals.

Technical feasibility:

The technical requirement for the system is economic & easily available since all the development
software are open source and the system uses Hardware that can be easily available.

Behavioral Feasibility:

The system working is quite easy to use and does not take any longer time since once the system
installed, it will work till any serious failure. Also, the operator requires no special training for
operating the system.

9
 8. LITERATURE STUDY
Many systems related to wild animal has been developed so far. These system works for the
conservation as well as promotion of wild animals. Wild animal protection has been one of the
main concern in today’s world since many illegal hunting and poaching of an animal are being
carried out. Although many systems to protect the wild animal has been developed we haven’t
found the system for protecting people who faces economic as well as physical damages.

8.1 EXISTING SYSTEM

Here existing system [6] consists various animal detection and tracking systems that are
implemented to detect the animals on roads to prevent from possible threats on humans from these
animals and vice-versa.

Kootenay, British Columbia, Canada

In June 2002 an animal detection system was installed along Hwy 93, in Kootenay National Park
in British Columbia, Canada, about 60km north of Radium, immediately north of the Dolly Varden
Day-Use. The system was designed to detect large animals, specifically white-tailed deer
(Odocoileus virginianus). Two infrared cameras that detect heat and additional equipment were
installed in the right of way. The software uses a combination of motion, speed and size to
determine whether the warning system should indeed be triggered. The system, especially the
cooling system of the cameras, experienced technical difficulties during the first year (June
through October 2002).

A modified system with different infrared cameras was installed in May 2003 (Pers. com. Hillary
Page, Sage Consulting). The road length covered by the system was cut in half (from 2,000m to
1,000m) because of the different cameras. The system has standard black-on-yellow deer warning
signs with amber flashing lights on top to warn drivers. The system became operational in
September 2003 (Pers. com. Nancy Newhouse, Sylvan Consulting). The system is currently only
active from dusk to dawn. The system may eventually be operational 24 hours a day.

10
Seven Locations, Switzerland

Kistler (1998) and Tschudin (1998) reported on a study that covered seven locations in
Switzerland. The systems were supplied by Calonder Energy AG in Dietikon, Switzerland. Each
system consisted of a series of passive infrared sensors. The sites, their installation date, the width
of the crossing area and number of sensors installed, are listed in table below. The passive sensors
were designed to detect ungulates such as roe deer (Capreolus capreolus) and red deer (Cervus
elaphus) within a 30-100m radius.

Rosvik, Sweden

In 1999 an animal detection system was installed along highway E4 near Rosvik in northern
Sweden (between Piteå and Luleå). The system was designed by PIK AB, Karlskrona, Sweden,
and it was installed by the manufacturer and the Road Administration. The system operates on a
break-the-beam principle with infrared light. The system was installed in a 100-m-wide opening
in a fence, and it was designed to detect moose. When an animal is detected lights are turned on
that illuminate the highway at the crossing area. This should allow drivers to see the animal better.
In addition, red warning lamps in the right-of-way are activated. A standard moose crossing sign
with the text “wildlife passage” is located just before the crossing area. The electricity supply was
a major problem, but that issue was solved in winter 2001/2002.

The other similar animal detection systems were implemented at different locations as shown in
table below,

11
 Table 8. 1: showing animal detection systems at different places.

ID Location Target Distance Fence Cost Cost Evaluation Operational

# species covered system Install. Installed

a 7 loc., Roe & 50-200m No $11,500 ? OVC ‘93- ‘93/’96-

Switzerland red ‘96 present1
deer
b Box, Moose 220m Yes $60,000 $40,000 OV Sep Dec ’96-
Uusimaa, ‘96 present
Finland
c Mikkeli, Moose 90m Yes $40,000 $30,000 O ‘99 ‘99-
Finland present
d Nugget Mainly 92m Yes $200,0002 ? OV 1 Dec 8 Dec ’00-
Cany., mule ‘00 21 May ‘01
WY, USA deer
e KootenayNP, White- 1,000m No ? ? O Jun Sep ’03-
BC, Canada tailed ‘02 present
deer
f 4 loc. CH; Roe & ? ±$20,0003 ? ‘98- ?
2 loc. D red ‘01/
deer ‘02?
h Rosvik, S Moose 100m Yes ±30,0004 ? O ‘99 ’00-present

i Colville, Deer, 402m No $9,0005 $3,000 O 20 Jun Taken down

WA, USA elk ‘00 spring ‘02
j Marshall, White- 1,609m No $50,000 $7,0007 O Jun Turned off
MN, USA tailed ‘01 Nov ‘01
deer
k Wenatchee, Deer 213m No <$40,0008 ? O Oct Oct ‘02-
WA, USA ‘02 present
l Yellowstone Elk 1,609m No $350,0009 $60,000 O Oct/ Not
NP, Nov operational
MT, USA ‘02
m South Bend, White- 9,654m10 No ? ? O Apr Not
IN, USA tailed ‘02 operational
deer
n Sequim, Elk 4,827m No $60,00011, ? O Apr Apr ‘00-
WA, USA $13,00012 ‘00 present
o Harris, Mostly 5,000m No $36,00013 ? O Apr Apr ’02-
SK, Canada Mule ‘02 present
deer

1
All in operation except Marcau site (road work Aug ‘97) 8
Incl. research, design, installation
2 9
Incl. operat. & maint., research, excl. WYDOT salaries Including research and development
3 10
In Switzerland Divided over 6 sections (1 mile each)
4 12
Excl. ± $70,000 for electricity For herding and collaring
5
Excl. signage, batteries
13 7
Excl. in kind contributions Excl. salariesTable 2. Main characteristics of the seven
systems located in Switzerland (Kistler 1998).

12
 8.2 OBJECT DETECTION
Object detection involves in detecting instances of objects from a particular class in an image
captured. The goal of object detection is to detect all instances/features of objects from a known
class, such as people, cars or faces in an image. Usually, only a small number of instances of the
object are present in image, but there are various possible locations and scales at which they can
occur and that need to somehow be expressed.

Images of objects from a specific class are highly stochastic. One source of randomness is the
actual imaging processing. Variations in illumination, variations in camera position as well as
digitization artifacts, all produce significant variations in image appearance, even in a stable scene.
Another source of variation is because of the intrinsic appearance variability of objects within a
class, even assuming no variation in the imaging process. For example, people have variable
shapes and wear a variety of clothes, while the handwritten digit 7 can be written with or without
a line through the middle, with different slants, stroke widths, etc. The hard part is to develop
detection algorithms that are invariant with respect to these variations and are computationally
efficient.

Figure 8. 1: showing object(animal) detection

13
 8.3 BASICS KNOWLEDGE OF DEEP LEARNING
The complete name is Machine Deep Learning; it’s a branch of the learning based on a set of
algorithms, that attempts to model high level abstractions in data, using multiple processing layer.
These models are based on two fundamental concepts, the neural network and the back-
propagation algorithm.

The first is a system that tries to approximate the human brain structure with programs and data
structure combination. The network initially is trained on a large amount of data and rules about
their relationship and then tested on the learned function, like object detection. The structure of
the network is usually composed by three fixed layer typology: Input Layer, Hidden Layer, Output
Layer. The ones that really learn are the hidden ones, that can be 1+n, in any combination and
functionality, where their total number defines the length of the model; from here the adjective
deep.

Figure 8. 2: Neural network showing its i/p, o/p and hidden layers

Then it’s the Back-Propagation Algorithm that makes possible training the model respect to an
optimization method. The system calculates the gradient of a loss function, respect to all the
weights in the network and then it’s passed to the optimization method that updates all the network
weights trying to minimize the loss function. [5]

14
 8.4 TENSORFLOW
Released on the 15th of November 2015 by Google, TensorFlow is the newest open source library
written in Python and c++ for numerical computation.

It has immediately a great success in the Machine Learning community and in less than one year
it also had a lot of support and development by Google itself, more over by many other projects,
developed in any area of Deep Learning.

The peculiarity of TensorFlow is its work flux, made by data flow graphs. Where Nodes represent
mathematical operations, edges represent the multidimensional data arrays communicated between
them; the latter can be considered, as in electronics, a Tensor, from here its name. In May 2016,
Google has revealed that it has used TensorFlow in AlphaGo project, with a special hardware
dedicated to boost library’s performances.

To reach rapidly this goal, Google dedicated a special attention to the user experience of
TensorFlow, which arrives with a great basic support and a wellgrown GitHub community, the
key of this quick improvement.

All these are the peculiarities that pushed us to choose TensorFlow, over more developed and
bigger communities. The irruption made in the Deep Learning environment, shows up its great
future potentialities, that are rolling out day by day. [5]

Figure 8. 3: TensorFlow logo

15
 8.5 FASTER R-CNN
Faster R-CNN is built upon Fast R-CNN. It unifies proposal and classification parts into a single
neural network which provides nearly cost-free region proposals and allows end-to-end training
and testing. Its architecture consists of multiple convolutional layers, Region Proposal Network
(RPN), a few fully-connected layers and two output layers – one for classification and one for
bounding box regression. Either ZF model, which has 5 shareable convolutional layers, or VGG16
model, which has 13 shareable convolutional layers, is used. The convolutional layers are
interspersed with ReLU and max-pooling layers.

RPN generates object proposals by sliding a small window over the feature maps. Each proposal
is given by a bounding-box and objectness score. The bounding boxes are regressed from 9 anchor
boxes that improve ability to detect objects of various scales and aspect ratios.

16
The feature maps and 300 proposals with the highest objectness score are passed to a region of
interest (RoI) pooling layer. The RoI pooling layer uses max-pooling to crop and reshape the
regions of feature maps defined by the proposals to achieve the same spacial size, typically 6 by
6. The resized regions are then processed by two fully-connected layers and the resulting feature
vector continues to two sibling fully-connected layers. One is followed by softmax and generates
probabilities of each class (including background). The other fully-connected layer outputs
regression of proposed bounding-boxes. There are two proposed strategies to train the Faster R-
CNN network, both of them use stochastic gradient descent. One alternates between training RPN
alone and the rest of network, the other is end-to-end but the gradients for RPN are only
approximate.

17
 8.6 CREATING A SET OF TRAINING IMAGES WITH TENSORFLOW

The first place to start is by looking at the images you've collected, since the most common issues
we see with training come from the data that's being fed in.

For training to work as we expect, you should collect at least a hundred photos in every possible
surrounding condition of an object you want to recognize. The more you collect, the better the
accuracy of your trained model is likely to be. You also need to make sure that the photos are of
good representation of what your application will actually encounter. For example, you take all
your photos indoors and your users are trying to recognize objects outdoors, you probably won't
get good results.

Another risk to avoid is that the learning process will pick up on anything that the labeled images
have in common with each other, and if you're not careful that might be something that's not useful.
For example, if you capture kind of object in a blue room, and another in a green one, then the
model will prediction on basis of the background color, not the characters of the object you actually
care about. To avoid these problems, take pictures in a variety of situations as you can, at different
times, and with different devices.

You may also want to think about the classes you use. It might be worth categorize big verities
that cover a lot of different forms into smaller ones that are more visually distinct. For example,
for ‘vehicle' we might use 'car', 'motorbike', and 'truck' etc. you also have to think about whether
you have a 'closed world' or an 'open world' problem. In a closed world problem, things to
categorize are the classes of object you know about. On contrasting, a roaming robot may see all
types of different things through its camera as it wanders around the world. In those situations,
you'd want the classifier to report if it was not sure about what it was seeing. This cannot be easy
to do well, but if you collect a large number of specific 'background' photos with no relevant
objects in them, you can do well. It is also good to make sure that all of your images are labeled
correctly. Often user-generated tags are unreliable for our purposes. For example: pictures tagged
#daisy might also include people and characters named Daisy. [7]

18
 9. METHODOLOGY
9.1 SOFTWARE DEVELOPMENT PROCESS MODEL
The framework we used in development of this software is “Incremental model”, i.e. a method for
software development where the product is designed, implemented, and tested incrementally. This
model combines the elements of waterfall model with iterative philosophy of prototyping i.e.
multiple development cycles are present here, that makes the life cycle a multi-waterfall cycle. In
Incremental model, the whole task is divided into various increments.

In incremental model, the first increment is a core product. That means, basic requirements are
incorporated, but many supplementary features (some known, others unknown) remain un-
delivered. The core product is used by the end user or undergoes detailed review. As a result of
use and/or evaluation of an increment, plans are developed for the next increments. The plan tries
to incorporate the modification of the core product to meet the needs of the customer and the
delivery of additional features and functionality. This process is iterative through the delivery of
each increment, until the complete product is produced.

Here we used Incremental model since we gathered and analyzed all the requirements for the
project i.e. requirement for the system are clearly defined and understood. Major feature of the
project is also known. So, we planned and divided our whole task into four increments.

19
 Figure 9. 1: Incremental Model

Analysis Phase: in this phase, gathering requirements and analyzing were done to find out the
actual requirement of the system. The result of this phase is a SRS (System Requirement
Specification). Here are some requirements for the project as,

Functional Requirements

System’s functional requirement describes about the activities and services that must be provided
i.e. the user/customer requirements.

• tracking wild animals in-front of device by object detection and recognition in specific
time.
• Generate specific sound or any noise to drive out those animals.
• Generate sound or any alert type to alert peoples about detected wilds.

Here, the requirements are fulfilled as expected.

20
Non-Functional Requirements

The Non-functional Requirements are the characteristics or attributes of the system/project that
can judge its operation. The following points clarify them:

• Accuracy and Precision: the system performed its process in accuracy and Precision as
expected to avoid problems.
• Security: the system is secure since there are no external interferences that can be arose
once the system installed.
• Usability: the system is easy to deal with and simple to understand since it does not need
much interactions.
• Speed and Responsiveness: Execution of operations should be fast but does not respond
as fast as we expected due to the processing power of Respberry Pi.

Design Phase: In this phase the SRS is translated into the systems is design. Context Diagram,
DFD, use case, activity diagram and class diagram have been developed. The modeling diagrams
are presented at system development section.

Coding Phase: In this phase code needed to map designs were done as per the design and a
working system is developed by the end of this process.

Testing Phase: In this phase, the system is tested through different testing methods. With each
testing a list of changes to the system were developed, suggested and the changes were applied to
the software and the software is delivered as a successive increment until a satisfying system is
achieved. The test results are shown below in each increment.

21
Advantages of Incremental model:

• gives core software quickly and early during the software life cycle.

• is less costly, more flexible and easy to change scope and requirements.

• easier to test and debug for a smaller iteration.

• customer can respond to each iteration.

• Lowers initial delivery cost.

• Easy to manage risks because risky pieces are identified and handled during iterations.

Disadvantages of Incremental model:

• Needs good planning and design.

• Needed a clear and complete specifications of the whole project before it can be broken
down and built incrementally.
• Total cost is higher than waterfall.

When to use the Incremental model:

• This model can be used if the requirements of the system are clearly defined &
understood.

• Major features must be defined; however, some details can evolve with time.

• If there is a need to get a product to the market in less time.

• If resources with needed skills sets are not available.

22
 9.2 TOOLS AND TECHNOLOGIES USED
To realize project successfully, it is necessary to select the right tools and technologies, Designs
should be made with the end-product in mind and keep in mind that which tools and applications
should be used to build the applications and how to make the environment scalable enough for the
final product. Tools and Technologies that are used in this project/system are as folloews.

Development Hardware required:

Table 9. 1: showing hardware required

Section requirements

Display Minimum resolution 800 x 600 pixel

1024 x 768 pixel recommended

RAM 4 GB or more recommended especially

for Microsoft Windows 7, 8 and 10.

CPU 1.7 GHz processor speed or higher

GPU GTX 960 or higher

GTX 1050 Ti or higher recommended

Webcam Minimum 360p webcam required

(recommended 720p, 5MP or higher)

23
Development Software required:
Table 9. 2: showing software required

Section requirements

Supported Operating System Resbian OS

Additional Software Library Requirements Lxml, anaconda, Cython, jupyter, matplotlib,

pandas, cuda, cudnn

Development software requirements Python 3+, OpenCV library, TensorFlow lib.

Component Used
Table 9. 3: showing components required

Section requirements

Respberry pi Pi model 3 B+

Webcam Pi camera module / webcam (Logitech c310)

Motion sensor Ultrasonic Sensor (HC-SR04)

Arduino Arduino Yun

Speaker Jy-25

24
 9.3 SYSTEM DEVELOPMENT
As we proposed the system that can continuously watch and monitor the possible dangers and give
some notifications to the peoples on the village with minimum involvement of humans, it has been
developed as expected. The main aim of our project is to reduce the possible loss on human beings
and their crops by monitoring the wildlife near the village and giving the result information to the
peoples. Here our system exists in between the possible incoming path of the wild animals towards
the village. The system detects the motion first and then detects & recognize the type of animal
through camera surveillances. After that the system generate some sound to drive out these animals
and also alert the peoples on the village.

Block diagram

Figure 9. 2: Block diagram for the project

25
Flowchart

Figure 9. 3: flow chart for the system

26
Use case diagram

Figure 9. 4: use case diagram for the system

27
Activity diagram

Figure 9. 5: showing activity diagram for the project

28
Class Diagram

Figure 9. 6: showing class diagram

29
Context diagram

Figure 9. 7: context level DFD

30
Level 1 DFD

Figure 9. 8: level-1 DFD for the system

31
 9.3.1 INCREMENT-1: object detection
Object detection involves in detecting instances of objects from a particular class in an image
captured. The goal of object detection is to detect all instances/features of objects from a known
class, such as people, cars or faces in an image. Usually, only a small number of instances of the
object are present in image, but there are various possible locations and scales at which they can
occur and that need to somehow be expressed.

The goal was to produce a working system for tracking objects in 3-dimensional space. The aim
of this iteration was to begin from this spec and design a solution to the problem. After approximate
models were designed, the task came to implement the models. Throughout the iteration, many
problems arose. These problems were varied from performance issues related to the code, and from
implementation issues related to the limitations of software technologies used in the project. All
attempts at overcoming these problems are discussed in this section. TensorFlow library is used to
detect objects in the image.

A. Analysis

In this phase, all the requirements for detecting objects were gathered. The required environments
were installed on PC. The requirements were analyzed whether it is possible to implement or not.
After analysis, planning for the increment was done and tasks for the iteration were divided.
Finally, the design concept for the increment was figured out.

Table 9. 4: showing planning for increment-1

Activity Period Comment

Requirement Gathering 39d Requirement gathering has took placed

through searching on internet & by taking the
information from various animal conservation
related organizations and taking the ideas,
sharing the views among group members.

Planning 37d Planning has done by Reviewing of literatures,

existing systems and other similar systems on
internet, by taking the walkthrough.

32
 Design 36d Designing has done by creating UML diagram
and by creating Charts

Coding & Implementation 34d Implementation has done by creating module

by module.

Testing 39d Testing has done by perfuming unit testing,

black-box testing and system testing.

B. Design

Design is the first step in the software development phase for any tools, techniques and principles
for the purpose of defining a device, a process or a system. Once the software requirements and
planning’s have been analyzed and specified, the software development involves three major
activities design, coding, implementation and testing that are required to build & verify the
software. Here are some model diagrams as followings,

33
Flowchart

Figure 9. 9: flowchart for animal detection

34
Use case diagram

Figure 9. 10: use case diagram for animal detection

35
Level-1 DFD

Figure 9. 11: level-1 DFD for animal detection

36
 C. Coding
In this phase coding was done as per the design and a working system was developed by the end
of this process.

D. Testing

At this section of work the testing strategies are applied and discussed the result. The unit testing
approach was applied to debug the coding errors, the white-box testing approach was done to test
whether the system gives output or not and finally system testing was done to test the The result
is shown below,

Table 9. 5: showing test results

Data sets iterated No of successfully % correct result

object detected object

Centered light 10 10 100%

Left light 15 15 100%

Right light 12 12 100%

Covered object 10 3 30%

Unveiled/uncovered object 10 9 90%

Fast moving object 20 8 40%

Similar objects 20 16 80%

White background 20 19 95%

Random background 20 15 75%

Multiple objects 10 8 80%

37
 9.3.2 INCREMENT-2: image training

For training to work as we expect, we should collect at least a hundred photos in every possible
surrounding condition of an object we want to recognize. The more we collect, the better the
accuracy of our trained model is likely to be. we also need to make sure that the photos are of good
representation of what our application will actually encounter. For example, we take all our photos
indoors and our users are trying to recognize objects outdoors, we probably won't get good results.
So, by keeping this in mind we have done the trainings & the discussions and results were
discussed in this section.

A. Analysis

In this phase, all the requirements for training the objects were gathered. The required
environments were installed on PC. The requirements were analyzed whether it is possible to
implement or not. After analysis, planning for the increment was done and tasks for the iteration
were divided and concept for the design are gathered.

Table 9. 6: showing planning for increment-2

Activity Period Comment

Requirement Gathering 13d Requirement gathering for training has took

placed through &sharing the views among
group members.

Planning 16d done by Reviewing of literatures, existing

systems and other similar systems.

Design 16d done by creating UML diagram and by

creating Charts

Coding & Implementation 30d Done by creating module by module.

Testing 63d Testing has done by perfuming various tests

strategies.

38
 B. Design

Once the software requirements and planning have been analyzed and specified, the software
development involves three technical activities design, coding, implementation and testing that
are required to build and verify the software. Here are some model diagrams as followings,

Flowchart

Figure 9. 12: flowchart for training animals/objects

39
Use case diagram

Figure 9. 13: use case diagram for training animals

Level-1 DFD

Figure 9. 14: level-1 DFD for training animal

40
 C. Coding

In this phase coding was done as per the design and a working system was developed by the end
of this process.

D. Testing

At this section of work, the different testing strategies are applied and the result are discussed.
Here we have done testing on the training sets in different modes so that the accuracy ca be
increased.

Figure 9. 15: training Log

41
 Figure 9. 16: Loss graph

Figure 9. 17: loss vs steps graph

42
 9.3.3 INCREMENT-3: hardware implementation
Hardware implementation involves various planning, task, cost & time estimations, coding and
testing. Here in this increment of the project, we have gathered various requirements about the
hardware devices and their interconnections. After gathering of the requirements, we designed the
device and implement the code that have already prepared in pc console. Finally, we tested the
system device whether it works or not.

A. Analysis

Here, all the requirements for h/w implementations were gathered. The requirements were
analyzed whether it is possible to implement or not. After analysis, planning for the increment was
done and tasks for the iteration were divided. Finally, the design concept for the increment was
gathered.

Table 9. 7: showing planning for increment-3

Activity Period Comment

Requirement Gathering 14d Requirements were gathered, estimated and

the views were shared among group members.

Planning 15d Planning has done by Reviewing of literatures,

existing systems and other similar systems on
internet, by taking the walkthrough.

Design 12d Designing has done by creating diagrams and

Charts

Coding & Implementation 12d Implementation has done by integrating the

hardware components.

Testing 36d Testing has done by perfuming system testing.

43
 Table 9. 8: showing hardware cost estimations

SN Components Quantity Rate Price (Rs)

1 Pi model 3 B+ 1 3920 3920
2 Arduino YUN 1 8400 8400
3 HC-SR04 2 440 880
4 Pi camera 1 3575 3575
module
5 Jy-25 1 1050 1050

Hardware Description
Pi model 3 B+

The Raspberry Pi 3 Model B+ contains a wide range of improvements and features that will
benefit the designers, developers, and even engineers who are looking to integrate Pi systems
into their products. Here are some of the Pi model 3 B+’s specs:

• Quad core 64-bit processor clocked at 1.4GHz

• 1GB LPDDR2 SRAM
• Dual-band 2.4GHz and 5GHz wireless LAN
• Bluetooth 4.2 / BLE
• Higher speed ethernet up to 300Mbps
• Power-over-Ethernet capability (via a separate PoE HAT)

Figure 9. 18: Pi model 3 B+

44
pi camera module v-2

The Raspberry Pi Camera Module v2 is a high quality 8-megapixel custom designed add-on board
for Raspberry Pi, featuring a fixed focus lens. It's capable of 3280 x 2464-pixel static images, and
also supports 1080p30, 720p60 and 640x480p60/90 video. Pi camera attaches to Pi by way of one
of the small sockets on the board (upper surface) and uses the dedicated C Si interface, designed
especially for interfacing to pi cameras.

Figure 9. 19: Pi camera module v.2

Arduino YUN

AVR Arduino Microcontroller

o -ATmega32u
o -Input Voltage: 5V
o -20 Digital I/O Pins
o -7 PWM Channels
o -12 ADCs
o -16MHz Clock Speed
o -32 KB Flash Memory

Figure 9. 20: Arduino YUN

45
 B. Design
Designing is the first step in the development phase for any techniques and principles for the
purpose of defining a device, a process or system in sufficient detail to permit its physical
realization. Once the hardware requirements have been analyzed and specified the software design
involves three technical activities design, coding, implementation and testing that are required to
build and verify the hardware.

C. Coding
In this phase coding was done as per the design and a working system was developed by the end
of this process.

D. Testing

At this section of work the testing strategies are applied and discussed the result. The unit testing
approach was applied to debug the coding & connection errors, the white-box testing approach
was done to test whether the system gives output or not and finally system testing was done to test
the system. Hence the system device tested successfully.

46
 9.3.4INCREMENT-4: integration of the project
After working on the various parts of the project separately we integrated them as one to give out
the final product. We have done the various task such as object detection (recognizing the object
specification), image training for the purpose of detecting the object as well as hardware
implementation of the project.

The foremost work of our project is to identify the object for the detection of an object to carry out
the further task. Object detection involves the detecting instances of an object from a particular
class in an image that has been captured in real time. The main objective of an object detection is
to detect the features of an object and recognize it. Image training is another important part of the
project.

To detect the object firstly some input should be given to detect which type of object is present.
Images of animals are given as input for our project. For the image training we have collected at
least hundred pictures of the specific object from every possible angle and surrounding so that they
can be recognized in any random conditions. The large number of pictures of a specific object
from different perspective has helped to increase the object detection accuracy of our system in a
great amount.

As we proposed we need a device. So, we have to implement our works in hardware components
to achieve the project goal.

All these task that were carried out separately are integrated at final stage to provide the final
system. After achieving the final product, we ran the test for the system to check whether it
performs as expected or not. We performed various test for different possible situation to find out
the results. The results obtained from all these tests were discussed among the team members and
supervisor.

A. Analysis

Here, all the requirements for increment integration were gathered. The requirements were
analyzed whether it is possible to implement or not. After analysis, planning for the final increment
was done and tasks for the iteration were divided.

47
 B. Integration Testing

We integrated all the task that were done separately to give out the final product. The integrated
product was monitored and tested. The result achieved through the test were discussed and the
conclusion was drawn from it. We obtained the output from the system and found the system is
working. Though the system is working but process is slow which is limiting the scope of project.

Due to the limited processing power of Raspberry Pi, real time object detection process couldn’t
be achieved as expected. Not enough memory space also contributed to the slow processing
resulting to the disappointment in the expected outcome of object detection. High resolution
picture led to the decrease in the frame rate of the system and vice versa. But when the resolution
was decreased the detail of the picture could not be reached. So, another limitation was
encountered when the expected frame rate per second was not achieved by the system.

Sound generation process encountered the error causing incompleteness to the system. It was also
difficult to integrate the motion sensor with our product. The main theme of our project to alert the
people through sound generation was achieved in the desktop but we found it very difficult to
implement in the hardware and couldn’t achieve it as a final product.

48
 10. TASK DISTRIBUTION

Table 10. 1: showing task division

Member Activity Comment

M1-M5 Requirement M2, M1 and M4 has performed the searching for project requirement
on the internet by reviewing the related literature and by analyzing
Gathering
the related project which is already available in the market. Regularly
inform to the other member of team.

M1-M5 Analyzing Whole team done the requirement analyzing of project by sharing
Of the the ideas, and by discussing on related information which is gather
requirement by the M1, And M2 and M3. M1 and M3 has created the list of
requirements after every meeting

M1-M5 Finalizing Whole team finalize the requirement. M1, M2 and M3 has created a
requirements list of finalize requirement.

M3, M4 Installation M3 and M4 installed the all the require tools which is use for object
of tools and detection. And M1 installed required UML and diagrammatic tools.
technology for Also informed to M1.
object
detection

M1, M4 Implementing M1 performs various modeling diagrams and M4 implements the

TensorFlow TensorFlow library to detect object (i.e. wild animals) and informed
Library to detect to M1 & M3
object

M4 Testing of object M3 done various test techniques and gathered different test cases.
detection Also informed to M1.

49
 M3 Installation M3 installed all the require tools which is use for training objects.
of tools and And M1 installed UML and other diagrammatic tools.
technology for
training

M1, M3 Implement M1 design model diagrams and M3 implemented the training

training module algorithm (i.e. TensorFlow library) and discuss on it with other team
members.

M3 Implement of M3 implemented the dataset module and discussed to the other

Dataset & team members.
trainer module

Testing and M3 tested the implemented module in various cases and discussed
verifications of with other members. Also informed to M1.
trained dataset

M5, M1 Gathering M5 and M1 gathered all requirements to implement hardware for the
requirements project and informed to other members.
for hardware
implementation

M5, M1 Integrating M5 and M1 integrated all the modules and discuss on it with M3 and
to each other regularly.
Hardware
modules

M5 Integrated M5 performed the unit testing and noted down results and discuss with
hardware testing M1 and other members of team.

50
 M3, M4, Requirement M3, M4 and M5 gathered requirements for project integration and
M5 gathering discussed with other members.

M1- M5 Implementing Whole members were gathered for integration and system integration
system was done successfully
integration

M1-M5 Unit testing All team members performed the functional testing and noted down
results and discussed the result of testing with each other.

M1-M5 Black box All team members performed black box testing and discussed the
result
testing

M1-M5 System testing Performed system testing and result was discussed with each other.

M1, M2 documentation M1 perform documentation portion of the project and M2 helped M1

at each phase.

M1= Amrit Sunar

M2= Binita Rai
M3= Prakash Yadav
M4= Rabin Rai
M5= Suraj Suwal

51
 11. DELIVERABLES
• A device that can detect and recognize wild animals.
• The device that can generate sounds to drive out those animals.
• The device that can generate the alert sounds to alert the peoples.
• project document and related demos. Furthermore, answers for the queries.
• review, feedback and overall judgement on the project.

52
 12. PROJECT TASK AND TIME SCHEDULE
The time schedule has been designed as per the requirements. Requirement analysis has been given
more emphasis. System Design is to be done first. Testing and Debugging is to be done along with
the development of the project. Finally, documentation is done throughout the project.

Figure 12. 1: Showing planned duration for tasks of increment-1

Figure 12. 2: showing Gantt chart for increment-1

53
Figure 12. 3: Showing planned duration for tasks of increment-2

Figure 12. 4: showing Gantt chart for increment-2

54
Figure 12. 5: showing task schedule for increment-3

Figure 12. 6: showing Gantt chart for increment-3

55
Figure 12. 7: showing task schedule for increment-4

Figure 12. 8: showing Gantt chart for increment-4

56
 13. CONCLUSION AND FUTURE SCOPE
13.1 CONCLUSION
In order to monitor and alert the peoples this system has been proposed. It replaces the manual
system with an automated system which is fast, efficient, cost and time saving as replaces the
continuous, tedious and manual monitoring and tracking of wild animals.

Although we had tried our best to achieve the expected result, we were still bounded by some
limitations. Due to the limited processing power of Raspberry Pi, real time object detection process
couldn't be achieved as expected. Not enough memory space also contributed to the slow
processing resulting to the disappointment in the expected outcome of object detection. High
resolution picture led to the decrease in the frame rate of the system and vice versa. But when the
resolution was decreased the detail of the picture could not be reached. So, another limitation was
encountered when the expected frame rate per second was not achieved by the pi.

Hence this system is expected to give desired results and in future could be implemented in other
various fields. Also, the efficiency could be improved by integrating other techniques with it in
near future.

13.2 FUTURE SCOPE

It can be easily implemented at any places where there is danger from wild animals. A method
could be proposed to illustrate robustness against the variations that is, in near future we could
build a system which would be robust and would work in undesirable conditions too. Here it is
proposed for an area where there is some danger from wild animals but in future it can be used to
do the same work at various areas that has same problem. Authors are working to improve the
object detection and recognition effectiveness to build more efficient systems in near future.

In the future if the processing power as well as memory of Raspberry Pi is increased then we can
obtain the expected outcomes. Also, if the TensorFlow is made compatible with the Raspberry Pi,
we can carry this project to have better performance.

57
14. BIBLIOGRAPHY
[1]. https://en.wikipedia.org/wiki/Wildlife_of_Nepal [ accessed on 28-06-2018]
[2]. Zhongjie Li and Rao Zhang, “Object Detection and Its Implementation on Android
Devices”, feb 2017.
[3]. Krishna Prasad Acharya, Prakash Kumar Paudel, Prem Raj Neupane, Michael Köhl
“Human-Wildlife Conflicts in Nepal: Patterns of Human Fatalities and Injuries Caused by
Large Mammals”, Sept 9-2016.
[4]. https://en.wikipedia.org/wiki/Object_detection [accessed on 29th jun 2018]
[5]. Andrea Ferri, “Object Tracking in Video with TensorFlow” October 2016.
[6]. Marcel P. Huijser and Patrick T. McGowen, “Overview of Animal Detection and Animal
Warning Systems in North America And Europe” 2003
[7]. https://www.tensorflow.org/tutorials/image_retraining [accessed on 30-06-2018]

58