Chapter 1

Introduction
1.1 Introduction

Classification of underwater objects such as mines, floating objects, submarines and

controlling of vessel traffic in all major ports along the sea coast has become a crucial factor

and essential for all developing countries.

The crucial equipment that provides navigational and surveillance capability to the

Autonomous Underwater Vehicle (AUV) is the sonar. SONAR (Sound Navigation and

Ranging) is a sensor that uses sound propagation, mounted onto the AUV to navigate,

communicate and detect underwater mines. Sonar is the underwater equivalent of Radar and

its functions are many and varied. These include detection, identification, location and speed

indication of the 'targets of interest'. Radar operates on the principle of transmitting and/or

receiving electromagnetic waves, travelling at approximately the speed of light (3x10 8

meters/sec) through space or the atmosphere. Sonar, on the other hand, relies on acoustic

energy in seawater which travels at 1,500 meter/sec, depending upon conditions such as

temperature, depth, and salinity. There are basically two modes of operation of sonar:

'Passive' and 'Active'. A passive sonar system emits no signals. Active sonar employs a

transmitter to send out sound pulses and a receiver to record returning echoes. In

conventional active sonars, when the separation of the targets is less than the range

resolution, then it displays a single large combined target rather than multiple smaller targets.

Compressed High Intensity Radar Pulse (CHIRP) sonars are invented to overcome the

limitations (i.e., the low coverage area and low range resolution) of conventional monotonic

sonars. In this thesis, the images from the chirp sonar fitted on AUV are used for underwater

object detection and tracking analysis.

In underwater navigation, timely scanning of area in front of the AUV and detection of

objects is very important to avoid collision of AUV with vessels or underwater objects. With

the advancement of technology, there are Imaging sonars which scan areas upto a range of

100 to 300 meters in front of the AUV and provide images as output. But the information that

is available from these imaging sonars is limited to range and bearing only. Therefore, there

is a clear need for extraction of features (range, bearing, size, shape and speed) of all

underwater objects (that are coming in the way of an AUV) by making use of the information

available in the images.

In addition, the sonar information collected while searching for, or identifying, underwater

mines is often presented to the operator in the form of a two dimensional image. This 2D

information provides only range and bearing but not depth of the target. These 2D sonars

cannot distinguish between the safe sea floor and the dangerous under water obstacles

through depth measurement. They must rely on visual cues to guess whether or not the

obstacles are located in the water column or on the sea floor. In order to overcome the

problem, a new surface construction algorithm is developed, in which the complete 3D model

of the object of interest can be extracted from the 2D images obtained from the sonar.

Object identification in the 2D images obtained from sonar is very important for extracting

the 3D features. These sonar images contain noise which is undesired information that

contaminates an image. In order to identify the objects present in the images received from

the imaging sonar fitted onboard an AUV, a process called segmentation is employed. The

segmentation methods include Edge detection, Canny Edge detection, Adaptive

Thresholding, Fuzzy C means Thresholding and Adaptive Histogram Equalization. Even after

segmentation of the real-time sonar images, images contain noise. So to further remove the

noise from the sonar images and to identify the objects clearly, a new Image synthesis

algorithm is developed. The proposed algorithm combines substantial information from

several segmentation methods of original sonar images using mathematical techniques in

order to create a single composite image. The resultant comprehensive image retains all the

important features of the individual images. In order to get the complete 3D information of

the object i.e., to know the depth information, the 2D images of the object are taken at

different elevation angles of the sonar and are used to reconstruct the 3D shape of the object.

From the available data, in order to reconstruct the 3D object a new surface construction

algorithm is developed. The algorithm consists of (i) acquiring data points on the surface of

the solid using sonar, (ii) reconstructing the surface from these points. The developed

algorithms are validated with the real time data corresponding to the underwater images of a

sector scan sonar fitted onboard an AUV. These underwater images are collected for several

days at three different places (Tow tank (NSTL, DRDO), Visakhapatnam Steel Plant

Reservoir and Bay of Bengal sea water) located in India.

1.2 Aim of the Thesis

The main objective of the thesis is to develop a new underwater object detection algorithm

and to develop a new 3 dimensional surface construction algorithm using triangulation to

extract the object depth information from the images obtained from the chirp technology

based Sector Scan sonar fitted on the AUV.

To achieve the above objectives, the following tasks are performed.

1. Investigation of various underwater and terrestrial navigational aids and analysis of

the propagation characteristics of sonar signals under the influence of underwater

noise.

2. Performance analysis of chirp technology and Monotonic sonars in terms of range

resolution with which separation of two objects can be identified clearly in the image

obtained from the sector scan sonar.

 3. Performance analysis of the existing image segmentation schemes and their

applicability to the sonar images.

4. Development of a new Image Synthesis algorithm to identify the objects in the noisy

underwater sector scan sonar images and comparison of its performance with existing

segmentation methods.

5. Development of a surface construction algorithm to construct the 3D shape of an

underwater object from the 2D images obtained from the sector scan sonar.

1.3 Applications of the Thesis

i) The proposed image synthesis algorithm based on image processing technique is

useful for underwater object detection on the sea-bed in the coastal regions where

physical, chemical and biological activities with immediate impact on humans take

place.

ii) The proposed image synthesis algorithm is useful for surveillance of ports, coastal

regions and also for underwater object classification such as identification of mines,

floating objects and submarines, etc.,

iii) The proposed surface construction algorithm is particularly useful in obstacle

avoidance and for aiding navigation.

iv) The developed 3D surface reconstruction algorithm is also useful to survey a

particular area or location in the sea for the oil and gas industry to make detailed maps

of the seafloor before they start building subsea infrastructure.

1.4 Literature Survey

The first task towards achieving the objectives of the thesis is the analysis of

characteristics of sonar and propagation of sound underwater (Qihu Li., 2012). This

analysis plays an important role in understanding the formation of images in Imaging

sonars. In this thesis the chirp technology based sonar is used to improve the range

resolution. The chirp technology is analyzed in comparison with the monotonic sonar.

The basic step to be undertaken in any image processing technique is noise filtering. For

this purpose, the characteristics of the noise present in the sea water and in the sonar

images are comprehensively understood (Urick R. J., 1983). The noise present in the sea

can be categorized in two ways: ambient noise and site specific noise. Ambient noise is

the open sea noise and includes turbulence (1-10Hz), shipping (10-100Hz), surface

(100Hz-100kHz) and thermal (above 100kHz) components. Whereas site specific noise

includes manmade, biological, ice cracking, rain and seismic event components (Fig. 1.1).

Majority of the ambient noise sources exhibit continuous power spectral density and

Gaussian statistics (Waite A.D., 2002). The noise power is characterized as

N(f) = K fb

(1.1)

where, f is the frequency, b is the noise power spectral density and K is the

spreading factor.

K = 2 for spherical spreading

= 1.5 for practical spreading

= 1 for cylindrical spreading

 The underwater noise power spectral density decays at b = 18 dB/dec.

Fig. 1.1 Underwater noise power spectral density

A comprehensive study is carried out on linear and non linear spatial filtering techniques

such as mean filter, Gaussian filter, median filter and frequency filters such as lowpass,

highpass and bandpass filters (Gonzalez R. C., 2005). These filters are implemented on

simulated and real time sonar images. Noise removal is the first step in any image

segmentation technique. Various spatial and frequency domain filtering techniques were

studied for various types of noise removal (Quidu I. et al, 2000; Yu S.C. et al, 2008). The

goal of segmentation is to take an input image and clearly define the objects in it. Image

segmentation is normally the next step in image analysis after noise removal (Zerr, 2005).

Segmentation can also be regarded as a process of grouping together pixels that have

similar attributes (Mignotte, 1998). All subsequent tasks such as feature extraction, object

recognition and 3D surface construction rely heavily on the quality of the segmentation

process (Bell, 2002). Several popular methods like adaptive thresholding, K-means
 clustering, region growing and edge detection methods are studied and some of them are

implemented on sonar images but they failed to identify the objects from the foreground

noise (Martinez G., 2005). Since there is no general solution to the image segmentation

problem, these techniques often have to be combined with domain knolwdge in order to

effectively solve segmentation problem for a specific application (Fox et al, 2004; Reed et

al, 2003). Therefore a new method which takes into account the advantages of the

existing segmentation methods is proposed for sonar object detection. In this process all

these segmented images are combined into a single image, which retains the important

features of the individual images (H B Mitchell., 2010). The developed image synthesis

algorithm is compared with principle component analysis method (Lindsay I Smith.,

2002; Yi Zhou., 2010). The proposed image synthesis algorithm is succeeded in detecting

the objects from sonar images with high PSNR.

The objects detected from the sonar images are two dimensional (2D). So for better

understanding of the features of the underwater object, its three dimensional features are

required (Satyan L. Devadoss et al., 2011). In order to construct the 3D model from the

2D data available from the sonar images a new surface construction algorithm is

developed. The developed 3D surface construction algorithm is compared with the ball

pivoting algorithm (Fausto Bernardini., 2010). The 3D surface reconstructed using

proposed algorithm has better resolution than ball pivoting method.

1.5 Research Methodology

The first step in achieving the research goal is the acquisition and processing of images

generated by the Imaging sonar. The detailed block diagram of the sonar object detection

and its 3D surface construction is shown in Fig. 1.2. The images that are received from
the imaging sonar fitted on the AUV provide only the range and bearing of the objects.

These objects are required to be detected in the image frames received from the sonar.

This involves the processing of the images in such a way that the detection of real objects

is not missed out and no false alarm is detected. The sonar images are required to be pre-

processed and image synthesis algorithm is to be developed which combines the features

of segmentation methods. Segmentation is required to distinguish objects from the

background noise. Several existing segmentation methods have been implemented on

sonar images and it is observed that, with a single method alone, it is not possible to

segment the sonar images properly. Therefore further research was undertaken and a new

image synthesis algorithm that suits the underwater real sonar images over the south east

coast of India is designed and developed.

Once the image is segmented and the objects are detected, their features are extracted.

The main aim of feature extraction is to extract the size and shape of the objects. The

object detected from the sonar is a 2D which provides only range and bearing of the

object. So there is need for the extraction of the 3D shape from the 2D images obtained.

In this thesis, first the different slices of the same object at different elevations are

obtained and the object is formed. Initially a 3D model is developed based on the slice

centroid. In this method centroids of each slice are calculated and all the points available

from the 2D images are joined based on spherical co-ordinates. The surface is

constructed using the method ball pivoting. Though it constructs the surface but it has

poor resolution. So a new algorithm is developed based on triangulation which provides

high resolution.

Clear
Video Identification
Images Removal of
of the object
Data noise and
Conversion of using
Acquisition object
raw video into developed
from the sonar detection using
images image
fitted on AUV segmentation synthesis
methods algorithm

Constructi
 Fig. 1.2 Block diagram of object detection and 3D surface construction algorithm

1.6 Technical Approach

Sonar is a sensor that uses sound propagation, which mounted onto the AUVs to navigate,

communicate and detect underwater mines. In conventional active sonars (monotonic

sonars), when the separation of the targets is less than the range resolution, then it

displays a single large combined target rather than multiple smaller targets. The

conventional active sonars cannot be used in detecting the underwater objects due to its

low range resolution.. The range resolution of conventional active sonar is given by the

following equation.

pulse length speed of sound

Range resolution = (1.2)
2

The minimum range resolution that can be obtained for monotonic sonar is 37.5 mm.

From the above equation, the ability of monotonic acoustic systems to resolve targets is

better, if the pulse duration is short. Ideally, long transmit pulses are needed to get

enough acoustic energy into the water for good identification of the furthermost targets,

but due to the velocity of sound (VOS) through water (1500 metres / second), each pulse

will occupy an equivalent 'distance' related to its pulse duration. The range resolution can
be improved in chirp technology sonar and is given by

velocity of sound
Range resolution =
(bandwidth 2)
(1.3)

The bandwidth of a typical chirp system is 100 kHz, and using the typical VOS of 1500

meters/second, the range resolution obtained is 7.5 mm which is better than conventional

sonar. The monotonic sonar cannot distinguish two objects if the distance between them

is less than 37.5 mm but using chirp technology sonar the objects can be distinguish if

they are 7.5 mm apart.

In this thesis, the images from the chirp sonar are used for underwater object detection

and tracking analysis. The images obtained are processed to remove the overall noise

present in the images. The next step in the research work is object detection. Several

segmentation methods are analyzed and implemented for object detection purpose. Since

the suitability of existing segmentation methods varies from application to application, a

new segmentation algorithm is developed for underwater image processing. In this

method all the features of the existing segmentation methods are combined. The best

pixels from all the segmentation methods are selected and combined into a single image.

The criteria used for selection of the best pixels are the eigen values. In order to assess

the best pixels, given image is partitioned into a lattice of blocks of size b b. For each

x y
block B, let gs= ( g s , g s ) denote the gradient of the gray level intensity. The covariance

matrix J of the gradient vectors for all b2 sites in this block is given by

1 j11 j12
J g sT g s
b2 s B j 21 j 22

(1.4)

From the above matrix the Eigen values (λ1, λ2) are calculated using the following
 equations

1
1 trace( J ) trace J 4 det J (1.5)
2

1
2 trace( J ) trace J 4 det J
2

(1.6)

Where trace (J) =j11+j22, det (J) = j11 j22 -j212 and λ1≥ λ2. The best quality pixels are

determined by using 1 2 . If λij is the quality assessment value of a block Iij, the

ordering of assessment values is given by

λ (1)> λ (2)>……..> λ (5)

and this implies the same ordering to the corresponding blocks

I (1) > I (2) > … > I (5)

where the subscripts are the ranks of the image blocks. Based on these ranks the better

quality blocks of pixels are formed into a new image. The block diagram of the

developed new image processing algorithm is shown in the Fig. 1.3.

Once the objects are detected from the noise free images, the next objective is to

construct the 3D surface of the object. Sonar information collected while searching for, or

identifying, underwater mines is often presented to the operator in the form of a two

dimensional image. This 2D information provides only range and bearing of the target

only. As the depth information is not available, this 2D information cannot distinguish

between the safe sea floor and the dangerous underwater obstacle through depth

measurement.

Edge detection

Chirp
Technology Canny Edge detection
based Sonar Image Noise
Data(Under
Filtering Synthesis Free
water objects Adaptive thresholding
Algorithm Under
images)
water
Image
Fuzzy C Means thresholding
 Fig. 1.3 Block diagram of image synthesis algorithm

In order to overcome the problem, surface construction algorithm is developed in which a

complete 3D model of the object of interest can be extracted from the 2D images

obtained from the sonar. In order to get the complete 3D information of the object i.e., to

know the depth information, the 2D images of the object are taken at different elevation

angles of the sonar and are used to construct the 3D shape of the object. Firstly a 3D

model is developed based on the slice centroid based method. In this method, centroids of

each slice are calculated and all the points available from the 2D images are joined based

on spherical co-ordinates. Using this method the shape is obtained but failed to construct

the complete surface of the object. So in order to construct the surface ball pivoting is

implemented. Even though the slice centroid in combination with ball pivoting

constructs the surface of an object but it has poor resolution. So a new algorithm based on

triangulation and quick hull algorithm is developed in this thesis for surface

reconstruction of underwater objects. This proposed algorithm provides high resolution.

The steps involved in triangulation are,

(i) A triangle of a planar point set should be formed with minimal internal angle

Let us consider a triangle pqr with sides a, b, c with internal angle, γ. The internal angle

can be calculated as

a2 b2 c2
cos
2ab
 (1.7)

The minimizing angle is given as

c(a 2 b2 c2 )
R cos
8A

(1.8)

abc
Where, R , and A is the area of triangle and A can be expressed by the three
4A

point determinant as

1 (1.9)
A
2 | D ( p; q; r ) |

Where,

xp yp 1
D( p; q; r ) xq yq 1 x p yq xr y p xq y r x p yr xq y p xr y q
xr yr 1

Now the third point is determined in such a way that it should minimize the Rcosγ. Let r

be a first third point detected r S be the first point for which D (p; q; r) ≠ 0 and define

one constant K (p; q; r) as

a 2 b2 c 2 (1.10)
K ( p; q; r )
D( p; q; r )

If D (p; q; r) > 0, identify the third point which minimizes K (p; q; r). Thus a new triangle

is formed with identified point.

(ii) This Triangle must satisfy empty sphere condition, that is, no point should lie inside

the circum circle of the triangle.

 To determine whether a point t lies inside the ∆PQR a incircle test should be performed.

The point t lies inside the ∆PQR if the following determinant is negative. The point lies

on the circum circle of the ∆PQR if the determinant is zero. The point lies outside the

triangle if the determinant is greater than zero.

px p y px2 p y2 1
qx q y qx2 q y2 1
incircle( p; q; r ; t ) 0 (1.11)
rx ry rx2 ry2 1
t x t y t x2 t y2 1

This triangulation method in combination with quick hull algorithm is developed in this

thesis. The developed algorithm consists of two steps: (i) acquisition of data points on the

surface of the solid using sonar (ii) reconstruction of the surface from these points. The

steps involved in the proposed surface construction algorithm are as follows:

1. Separate the z coordinates from all the sampled points so as to project them on to the

x-y plane.

2. Now raise all the 2D points to 3D points by taking the z coordinates of the new points

as x2+y2 .

3. Construct the convex hull for all these points using Quickhull algorithm.

4. Once the convex hull is obtained the points are brought down to 2D by removing the

added new z coordinate. By doing this, the triangulation of all the points can be

obtained.

5. Now the original z-coordinates are added to the points and obtained triangulation is

plotted to get the reconstructed surface.

1.7 Organization of the Thesis

The thesis is organized into seven chapters including the Introduction and the
Conclusions. Chapter 2 describes various underwater navigational aids, how the sound

waves propagate and the underwater noise effects on the sound signals. It also provides

analysis of the sonar characteristics in underwater. Chapter 3 introduces the chirp

technology sonar in comparison with monotonic sonar. This chapter also presents the

performance analysis of chirp technology in terms of range resolution with which

separation of two objects can be identified clearly in the image obtained from the sector

scan sonar. The performance analysis of various image segmentation methods that are

used for underwater object detection is presented in Chapter 4. The proposed underwater

object detection algorithm and its performance analysis are presented in Chapter 5.

Chapter 6 presents the results of two 3D models that are implemented based on the slice

centroid and 3D surface construction algorithm. This chapter also includes the

performance analysis of the proposed 3D surface reconstruction algorithm for underwater

objects. The overall conclusions of the thesis along with the future scope of the work are

presented in Chapter 7.