1

CHAPTER 1

INTRODUCTION

1.1 VIDEO PROCESSING

Video is a sequence of images displayed with high frequency to

enable perception of human eyes to observe the continuity in visual content.
The two consequent images (called as frames) are closely related with each
other based on information content. Visual content in the frame can be
interpreted as a hierarchy of abstractions. Raw pixels are at first level of
hierarchy which provides color or brightness information. By processing
further, features like edges, lines, curves and color intensity regions can be
obtained. These features are combined to form attributes at higher layer of
abstraction (called as objects). Further multiple objects and their relationships
are interpreted. If hierarchical structure is defined on the basis of temporal
interpretation frame, shot and scene are the steps of a video. Frame is the
lowest level of this hierarchical structure. In certain situations, video must be
partitioned into shots. A shot is an image sequence that portrays a continuous
action. These continuous shots are merged together to form a scene.

Digital videos are taken in real world environment containing 3D

objects in motion. These objects in motion are projected as continuous
sequence of images meaning that the image intensities and colors are in
motion. Indeed, the first step in most of the video processing algorithms is
motion detection, where the movement of intensities or colors of object is
estimated. These estimated motion parameters can be used for further
 2

processing and analysis. But, the temporal dimension of video indicates

significant increases in required storage, bandwidth, and processing
resources. There is a need for efficient algorithms that exploit some of the
properties of video like temporal redundancy to eliminate complex
computation and higher storage requirements.

Computer vision and Human vision works in similar passion but

does not exhibit exactly same functions. Human vision totally relies on
perceptional view of the observer. The results of perception vary from person
to person. Computer vision relies on algorithms and their constituent
declarations. Thus the results depend on computational structure of the
defined algorithm. The application of computer vision falls in one of three
categories - surveillance, control and analysis. Object detection and tracking
is the basic item in surveillance application. It is a non invasive approach of
observing and understanding the motion patterns of objects in video for
information extraction. In control applications, some parameters are
considered to control the motion and thereby controlling the visual system.
Analysis applications and they are generally automatic and used to diagnose
and optimize the performance of visual system.

VISUAL SURVEILLANCE

TARGET DETECTION TARGET TRACKING

Figure 1.1 Components of visual surveillance

 3

In visual surveillance systems as in Figure 1.1, moving object

detection and tracking are vital steps in context analysis, recognition and
indexing processes. Choosing a more suitable algorithm for varying
environmental condition is a big challenge.

1.2 UNMANNED AERIAL VEHICLES

The UAV is an acronym for Unmanned Aerial Vehicle, which is an

aircraft with no onboard pilot. UAVs can be remote controlled aircraft or can
fly autonomously based on pre-programmed flight plans or dynamic
automation systems. UAS denotes Unmanned Aircraft System which reflects
the complex systems including ground stations and other elements with the
actual air vehicles. Along with intelligence, surveillance, and reconnaissance
(ISR) missions, the role of UAV has expanded to areas like network node or
communications relay, electronic attack (EA), suppression of enemy air
defense (SEAD), destruction of enemy air defense (DEAD), strike missions
and combat search and rescue (CSAR).

1.2.1 Types of UAV

There are different types of UAVs, designed for different purposes.

Some UAVs use a blimp design, well suitable for carrying large amounts of
payload cargo. Some UAVs called as drones are not autonomous, because
they require constant control input from a remote human pilot.

1.2.2 Applications of UAV

UAVs are being used in military, and in some cases, civil

applications. Currently, UAVs are used for tasks like aerial reconnaissance,
scientific research, logistics and transportation.
 4

 Aerial Reconnaissance – UAVs are often used to obtain aerial

video of a remote location, especially where there would be
unexpected risk to the pilot of the aircraft. UAVs can be
loaded with high resolution video and thermal cameras. The
video and other information obtained by the UAV can be
streamed back to the ground control station in real time.

 Scientific Research – In many cases, scientific research

requires obtaining data from hazardous or remote locations
like nuclear power plants, high altitude towers and multi
storey buildings etc. Here, UAV can be used to receive data
without risk to human pilot.

 Logistics and Transportation – UAVs can be used to carry and

deliver a variety of payloads. Here, helicopter type UAVs are
used because payloads can be suspended from the bottom of
the airframe.

1.2.3 Multirotor UAV

A multirotor or multicopter is a rotorcraft with two or more than

two rotors. The benefit of multirotor aircraft is its simple rotor mechanics for
flight control. Multirotor often uses fixed-pitch blades and thereby the control
of vehicle motion is achieved by varying the relative speed of each rotor to
change the thrust and torque produced by each.

As the construction and control is easier, multirotor aircrafts are

frequently used in model and radio control aircraft projects in which quad
copter, hex copter and octocopter are used to name 4, 6 and 8 rotor copters
respectively. Multicopters often use fixed-pitch propellers, so that the control
of vehicle motion is achieved by varying the relative speed of each motor.
 5

Low budget Radio controlled multicopters are increasingly popular for aerial
photography and videos of landscape and buildings. The number of motors
and configuration of each type of multicopters bring some advantages and
disadvantages to their performance. For instance multicopters possess high
lifting capacity, can carry more payloads and can fly for long time in the air
and the disadvantage is that it more expensive for the additional motors and
batteries and it consumes more power.

A bicopter has two motors that can be moved by servos. It is the

cheapest multicopter because it uses only two motors, but it is least stable and
hard to tune. A tricopter has 3 motors in a “Y” shape apart or “T” shape. Two
propellers on the front arms pointing to the side forward, and one arm
backward. It has low lifting power and flight time because the motors have to
run faster to hold it all in the air. A quadcopter has 4 motors mounted on 4
arms of a frame 90 degree apart. There are two sets of clockwise (CW) and
counter clockwise (CCW) propellers mounted on the motors. The quadcopter
configuration is the most popular type. There are 2 configurations: ‘X’ or ‘+’.
Sometimes “H” configuration is used for aerial photography or first person
view (FPV). The hexarotor has 6 motors mounted on 6 arms 120 degree apart
on a symmetric frame. It is similar to the quad copter, but provides high
lifting capacity. It is also possible that if one motor fails, the aircraft can still
land safely. Octocopter has 8 motors on the same level with four sets of CW
and CCW propellers. It is an upgraded version of the hexarotor.

1.3 UAV BASED AERIAL SURVEILLANCE

Aerial surveillance with unmanned aerial vehicles has a wide range

of potential applications like traffic surveillance in cities and highways,
security guard for important venues and buildings and detection of military
targets etc. It renders access control in restricted areas like military base so
 6

that the unknown and unwanted movements can be restricted. A security

survey can be done by deploying unmanned aerial Vehicles and ensuring the
safety can be done. The movement pattern of vehicles in the highway or in
the traffic congested areas can be monitored. The behavior of the people can
be observed non-invasively thereby a close watch on the actions ongoing in
public meeting places with crowd can done. The following difficulties would
arise while detecting and tracking a target with unmanned aerial vehicle.

 A 3D object is depicted in 2D. Thus there is significant loss of

information.

 The objects may undergo complex motion, variable velocity

and acceleration.

 The limited image resolution caused by compression may

provide low quality data.

The work of the thesis is shown in Figure 1.2, focuses on

processing the video data obtained from UAV to extract some information
about the target like size, heading and its description. This is done through
target detection and target tracking algorithms. Video detection and tracking
is the process of estimating the timely location of one or multiple objects of
interest in a video stream captured by a visual camera. The aircraft will be
controlled by an operator from ground control station (GCS). Along with
telemetric data, video will be transmitted from UAV to GCS. Tracking is
initialized by selecting the object of interest by the user.

The target detection and tracking algorithm should consider the

requirements and purpose it is designed for. There is no universal algorithm
that handles all scenarios magically. The main challenge a video processing
algorithm for aerial platform would face is detection of objects, recognition,
 7

clusters, variation in pose, equipment constraints and changes in external

environment. While designing an algorithm certain aspects regarding UAV
are to be considered. The primary aspects are the altitude and environment.
The UAV can operate at altitude ranging from 50 to 1000 meters with
average speed of 20km/hr. The UAV can be operated in day time or night
time under varying weather conditions. Also the challenging scenarios
including small objects, obstacles and radical changes in terrain should be
considered. The target may enter or leave from the field of view many times.
If the interference between UAV and GCS occurs, noise will be introduced in
the video stream, thus making it harder to detect, distinguish and track the
objects.

Telemetry Data & Video stream

Ground Control Station PC- Video Processing Algorithms

Figure 1.2 Experimental setup

The works of Mattyus et al (2010), Ali & Shah (2006) and Yu et al

(2008) render discussions on video tracking problems like motion platform
compensation, target detection and tracking in videos captured by UAV for
aerial surveillance. To deal with the problems, a few assumptions are to be
considered. If the area of camera is not wider than the relative altitude, it can
be exempted from same parallax that rises to project 3D object in 2D plane
 8

(Kang et al 2005). Thus traditional motion alone shall be considered to

simplify the processing complexity.

Camera mounted on UAV is not stationary and produces ego

motion. This affects motion detection in targets. Thus motion compensation is
to be done. This can be done by image registration or image stabilization
which aligns two or more objects geometrically taken of same scene at
different angles. It may be done by warping based direct method used by
Szeliski (2010). It can be feature based registration. These features are done
by Harris corner detector discussed by Yu et al (2008) and Ali & Shad
(2006). Another popular feature is SIFT discussed by Lowe (2004). There are
three distinct motives in motion detection and tracking namely surveillance
(Moeslund et al 2006), control (Pavlovic et al 1997) and analysis (Balcones et
al 2009).

1.3.1 Requirements of an Aerial Surveillance System

Aerial surveillance is a very challenging domain due to loss of

information while projecting a three dimensional real world data in two
dimensional images. Noise in image, clutters in background of object,
illumination changes also interfere the performance of a surveillance system.
The algorithm designed for target detection and tracking problem must be
capable of adjusting and adapting to abrupt changes in motion and drifts. A
perfect surveillance system should possess three important properties

 Robustness: A surveillance system should be capable of

detecting and tracking a target even under very complex
circumstances like complex movement, cluttered environment
and abrupt illumination changes.
 9

 Adaptivity: A surveillance system may be used in different

conditions with varying backgrounds of varying properties.
The target may vary its property like size, scale and
orientation in displacement. The surveillance system should
have adaptive mechanism to overwhelm the changes.

 Ability for real time processing: The surveillance system

should be capable of coping with the real time environment.
Thus the design of algorithms should be focused on less time,
cost and memory requirements.

The properties and criteria of aerial surveillance is different from

that of civil surveillance platform. Motion detection and tracking system for
aerial surveillance should outfit the efficiency of an algorithm designed for
general visual surveillance application. The algorithms that are suitable for
major surveillance applications may not be suitable for the aerial surveillance
applications. A few parameters are to be considered based on which motion
detection and tracking systems can be designed and compared.

1.3.1.1 Multiple motion detection

In aerial videos there may be more than one object under motion.
The designed algorithm should be capable of detecting the change in position
of one or more objects in the scene.

1.3.1.2 Immunity to changes in background distraction

Aerial video contains complex background structure. The

background will not be smooth in all frames. A car moving in a high way may
produce smooth background but when the white lane mark interrupts, the
background changes (black road to white lane mark and vice versa). This
 10

increases the complexity in detection and tracking and an algorithm designed

for aerial surveillance should be capable of overcoming these issues.

1.3.1.3 Ability to distinguish other close by motions

There may be multiple objects closer to the target of interest. An

algorithm should be able to distinguish the target object from them. There
arises complexity in detection and tracking when two or more objects
including the targets of same features. The algorithm should be capable of
tracking the exact target.

1.3.1.4 Shadows

In aerial surveillance videos, since the application is in outdoor

environment shadows are the major source of noise that affects the efficiency
in detection and tracking. A well defined tracking algorithm should be able to
differentiate the shadow from target and thereby track the exact target.

1.3.1.5 Presence of complex and cluttered background

The targets may move in various environments with varying

background. The algorithm should be capable of adapting to the changes in
changing and cluttered background.

1.3.1.6 Variations in velocity

The target to be tracked cannot be expected to maintain regular and

standard speed. The moving speed and the direction pattern of the target may
vary randomly. The detection and tracking algorithm should be able to cope
up with the changes in the velocity.
 11

1.3.1.7 Day night operation

Aerial surveillance is done in day time with higher illumination or

at night with lesser illumination. The UAV may be deployed in day or night
with ordinary visual camera or thermal camera as payload. The algorithm
should be capable of detecting and tracking in varying mode of illumination

1.3.1.8 Camera positioning

The data from unmanned aerial vehicle is obtained from the camera
mounted at the bottom. The video taken may be perfect orthogonal or a little
angular. In perfect orthogonal mode, the size (length and breadth) is relatively
constant but the altitude information (height) is almost lost. In angular videos
the height information can be obtained but the length and breadth information
is varying from frame to frame. The designed algorithm should be capable of
processing in spite of varying data parameters.

1.3.1.9 Object too small in Field of View

Aerial surveillance is the environment of changing altitude from

low to higher values. In general, detection and tracking efficiency is inversely
proportional to the altitude. If the height is lesser, more feature information
about the target is obtained thereby increasing the efficiency. When the
altitude is high, the size of the target is very smaller (and sometimes
negligible) than the field of view. Thus only a few features can be extracted
from the target of interest. The designed algorithm should be capable of
performing well in this criterion.
 12

1.3.2 Unmanned Aerial Vehicle DHAKSHA

DHAKSHA is an unmanned aerial vehicle designed and developed

by the faculty, students and scholars of Division of Avionics, Department of
Aerospace engineering, Madras Institute of technology campus of Anna
University Chennai. Many models like quadcopter, hexacopter are developed
and experimented for assisting in surveillance. These models are very
successful at different research level deployments (in national and
international level) and civilian and military deployments.

Figure 1.3 Unmanned aerial vehicle - DHAKSHA

A hexacopter shown in Figure 1.3 is equipped with a high

definition camera as payload is used for video database generation. The
model is designed for robustness in high altitude, endurance, high wind and
varying weather conditions. The specifications are listed in Table 1.1. The
algorithms in this thesis make use of the video data generated by the model
DHAKSHA.
 13

Table 1.1 UAV specifications

Parameter Specification
Propeller 14 inch radius
Arm length 40 inches
Ground clearance 15 inch
Endurance 1 hour
Altitude 300 -1000 feet above ground level
Range coverage 10-15 kms
Speed 55-60 km/hr

1.3.3 Camera Setup

The Camera GoPro product with HERO3+ Silver Edition is

mounted on UAV as payload. It captures professional-quality video at
resolutions up to 1080p, 60 FPS and 10MP photos at up to 10 frames per
second with 4x faster Wi-Fi, enhanced low-light performance. The properties
are listed in the Table 1.2

Table 1.2 Video camera specifications

Video Resolution 1080p 960p 720p

Frames per Second 60/50/30/25 60/50/30/25 120/100/
(fps) NTSC/PAL 60/50/30/25
Field of View (FOV) Ultra Wide/ Ultra Wide/
Ultra Wide
Medium/Narrow Medium/Narrow
Screen Resolution 1920x1080 1280x960 1280x720
Aspect Ratio 4:3 4:3 4:3
 14

1.4 TARGET DETECTION

Detecting the target in video data sequence is the primary step in

aerial surveillance for subsequent processing. The efficiency of surveillance
systems relies on detecting the motion of any object in consecutive video
frames, thereby assisting in understanding the movement patterns of persons
and objects to uncover suspicious events. The designed algorithm should be
accurate in presence of cloud, dynamic clutter and variations in scene. The
objective of detection is to detect the parts or pixels in an image that are
significantly different from the background. It evolves choosing the best
models, initializing model parameters, distinguishing objects from shadows
and other artifacts and managing or adapting to random illumination and
scene changes. Since the videos from UAV are dynamic, the target of interest
and its background remains moving. Thus the algorithm should be capable of
adapting to the dynamic scene and background changes. A basic approach to
detect target is to perform background subtraction, which detects moving
objects that differs from the background model in each frame. It is a region
based technique. A background subtraction algorithm can be evaluated by
how effectively it finds all the target pixels when the pixels of the target first
appear in the scene. There are three stages in designing a background model
namely model representation, feature representation and model adaptation.

A good background subtraction algorithm must be robust against

illumination changes. It should avoid detecting other objects like rain, snow,
leaves and shadows. Also, the background model should react or adapt
instantaneously for rapid changes in scene. A good algorithm should handle
the relocation in background objects like waving trees and image changes due
to motion of camera. A background subtraction system should adapt to the
illumination changes, shadows and reflections. During initialization phase a
common problem faced is the existence of foreground objects occluding with
 15

the background. Some of these issues can be handled by computationally

expensive methods but in real time, a short processing time is required.

1.4.1 Model Representation

Target models are represented by their shapes and appearances.

The different ways of representing a target is shown in Figure 1.4. The model
may be represented by a point i.e. centroid or set of points. This
representation is suitable for tracking objects/targets that occupy small
regions in an image. The target model may be described as shapes by a
rectangle or by an ellipse. The motion of this pattern is usually modeled by
translation, affine or projective transformations. They are suitable for tracking
rigid objects. The model may be represented as a contour or a boundary. The
region inside the boundary is called Silhouette. They are suitable for complex
and non rigid objects. Articulated shapes are designed with body parts held
together with joints. The skeletal models can be obtained by deriving median
transform to the Silhouette.

POINTS

SHAPES

TARGET
CONTOUR
REPRESENTATION

SILHOUTTE

ARTICULATED
SHAPES
SKELETON

Figure 1.4 Model representation

 16

1.4.2 Feature Initialization

The desirable property of a feature is its uniqueness to distinguish

the targets easily in feature space. Color is considered as a feature in
histogram based representations. Color feature extraction is done in RGB
space in general. But the RGB space is not uniform and they are highly
correlated. So HSI color space is preferred due to its uniformity. Contour
based representation prefers edge as an important feature. The edges are less
sensitive to illumination changes. Optical flow based features have dense
field of displacement vectors which defines the translation in each pixel in a
region. It is computed by brightness constraint. Texture feature can also be
extracted which is a measure in intensity variation of the surface that defines
smoothness and regularity.

1.4.3 Model Adaptation

If a camera is fixed and if the background stays relatively constant,

the background can be modeled as single static image which can be easily
identified and ignored. But, if the background is not constant it is to be
modeled with mean intensity at pixel and the corresponding variance. These
models will design an adaptive tolerance for the changes in background. More
tolerant models are needed for the cases of background having motion pixels
which should be considered as background. It can be done by introducing
average measure at each frame. But the single average model fails to classify
all pixels with high degree of accuracy due to change in mean at each frame.
Adaptive mean technique helps in eliminating the issues faced by single
average model with the cost of increased complexity. In such cases Gaussians
can be used which incorporates the adaptive tuning measure and adaptive
Gaussian function.
 17

There exist various approaches to background subtraction varying

from simpler techniques like frame differencing, adaptive mean filtering to
sophisticated complex techniques like probabilistic modeling. The higher
level techniques produce superior performance but the simpler techniques
produce good results with less computational cost.

1.4.3.1 Frame differencing

Frame differencing is also known as temporal frame differencing

technique. It is the basic background subtraction technique. It uses the video
frame at time‘t-1’ as background model for the incoming frame at time ‘t’ and
subtracts them. It faces aperture problem for targets of larger size.

1.4.3.2 Mean filter

Mean filter finds the average of pixel intensity in each frame of the
video stream thereby creates background approximation which is similar to
current static scene other than the cases with motion. The technique is not
robust to the scene changes with slow moving objects. It recovers slowly with
uncovered background and requires predefined threshold for the entire scene.

1.4.3.3 Median filter

Median filter assumes background to be the median at each pixel

location of all frames in the buffer. It makes assumption that the pixel stays in
the background for more than half of the frame is buffer.

1.4.3.4 Min-Max filter

In min-max filter, three values (min, max, difference) are defined in

training phase for foreground detection. These values are estimated for all
frames and updated periodically.
 18

1.4.3.5 Linear predictive filter

Linear predictive filter (LPF) computes the current background

estimate by applying a filter in the buffer using wiener filter to predict a pixel
current location from k- linear estimates of previous values.

1.4.3.6 Principal component Analysis

Principal Component analysis (PCA) is a linear transformation of

variables that retains for multiple iterations into the largest amount of
variation in training data. Estimation of basic components for larger vectors is
obtained from Singular value decomposition (SVD).

1.4.3.7 Kernel density estimation

In Kernel density estimation (KDE), an approximation of

background probability density function is defined by histogram of most
recent background values. To model a kernel based background for ‘n’ frames
‘n’ buffers are required.

1.5 TARGET TRACKING

Target tracking is the problem of following moving elements across

a video sequence. The aim of a tracker is to generate a time based trajectory
of the target by locating its position in each frame. This is an essential
building block for vision systems addressing tasks like visual servoing,
metrology, surveying and mapping. Tracking systems must address two basic
problems namely motion and matching.

 Motion: Prediction of the location of target being tracked in

next frame in limited search region where the element is
expected to be found with high probability.
 19

 Matching: Identifying the image element in the next frame

within the designated search region.

While tracking a target, at initial level the descriptive feature of the

target is to be interpreted. Tracking a moving object can be done in three
ways – point, kernel, and silhouette as in Figure 1.5.

Deterministic

Point
Probabilistic

Template
Target
Kernel
Tracking
Multiview

Silhouette
Shape Match

Contour

Figure 1.5 Types of tracking

 Point Tracking – Targets detected in consecutive frames are

represented as points. The association of points is based on
previous state of the target. They can be done by deterministic
or statistical approach.

 Kernel Tracking – Kernel refers to a target shape or

appearance. Targets are tracked by computing the motion of
kernel in subsequent frames. Example – rectangular or
elliptical kernel. These kernel models are defined by template
based, density based and multi view based appearance models.
 20

Mean Shift Tracking, Lucas Kanade Tracking and Eigen

Tracking are of this category.

 Silhouette Tracking – Here, tracking is done by estimating the

region of the target in each frame. The object information in
each frame is encoded. This information is in form of
appearance density and shape models. They perform by
contour evolution and matching shapes. State space models,
variational methods and heuristic methods fall in this catagory.

To track a target, primarily the description of the target is required.

It may be shape, color, description or texture. Accuracy in description of the
target leads to higher accuracy in tracking.

1.5.1 Kernel Based Tracking

Mean shift tracking is the main element in kernel based tracking.

The concept is based on shifting from one position to another position in
consecutive frames based on some similarity measures. The comparison is
done by comparing the target model created and stored in kernel with the
candidate model in incoming frames. There are many approaches in shifting a
model from frame to frame. A brief discussion on some of the techniques is
given below.

1.5.1.1 CAMShift

The continuously adaptive mean shift tracking is based on the

adaptation based on mode of distribution of probability density function of
the target window. Unlike traditional mean shift technique, CAMshift uses
continuously adaptive probability density function frame by frame. This is
 21

one of the simplest techniques to provide robust results for the cases if
background and target are of significantly different color.

1.5.1.2 ABCShift

Adaptive Background CAMshift tracking is the updated version of

CAMshift technique specially designed for close range tracking from a static
camera. This suits for cases where color of the target and background are
similar. It achieves robustness for stationary mode camera setup only.

1.5.1.3 PAMShift

Path assigned mean shift is a fast method to track objects. In this

technique, all points along the mode points are assigned to the final mode
value. Points with already assigned nodes are eliminated and not traversed in
the next iteration. Convergence to the target is faster due to reduction of
feature space in each iteration.

1.5.1.4 FCMShift

Fuzzy Clustering mean shift technique uses color information to

characterize the object. Here unlike other techniques which use fixed bin size
for describing color model, a fuzzy approach is used. The fixed bin size
would be unfeasible for classification of target and introduces quantization
errors. Training of the data set is done in FCM shift model. In training phase
only one pattern is considered at a time. The similarity between training
pattern and input test pattern are measured. As the difference is measured, a
membership function is also measured. Based on the membership function a
cluster is formed.
 22

1.5.1.5 SOAMShift

Scale and orientation adaptive mean shift tracking technique uses

weight image derived from the target model and candidate model. Based on
the weights the scale and orientation of the target in the next frame is
determined and the target tracking window is shifted.

1.6 OBJECTIVES OF THE THESIS

 To study and discuss various issues and solutions for

surveillance videos captured by Unmanned aerial vehicles

 To evaluate the feasibility of implementing various video

processing algorithms for aerial videos to extract some useful
information.

 To design and implement a target detection algorithm and

target tracking algorithm suitable for dynamic environment
of aerial platform.

 To implement target detection and tracking algorithms for

different data set of varying conditions.

 To evaluate the performance of the derived algorithm by

comparing it with the existing algorithms.

1.7 ORGANISATION OF THE THESIS

The thesis consists of 6 chapters. Chapter 1 provides an overview

on aerial surveillance, unmanned aerial vehicles, target detection and
tracking techniques. Chapter 2 gives a comprehensive review of the
literature on assessment of target detection and tracking techniques for
 23

surveillance applications. Chapter 3 describes the methods and protocols

implemented to detect and track targets. Chapter 4 focuses on the results of
the designed algorithms compared with the existing algorithms. Chapter 5
deals with the significant conclusions. Chapter 6 discusses the scope for
future work.