69

CHAPTER 4

RESULTS AND DISCUSSION

The video data from UAV DHAKSHA is considered for the

analysis of implemented algorithms. Variety of dataset stating different
criteria have been used and the algorithms are implemented and evaluated.
The following criteria are considered.

 Motion – linear and non linear (curved)

 Velocity – Regular and random

 Background – Smooth, cluttered and abrupt changes in

background

 Target – Isolated targets and models similar to target models

 Size – Constant size and varying size and very small size of
target

 Altitude – Low altitude and high altitude

For each dataset the results of target detection and target tracking
algorithms are obtained and their performance is evaluated with a few
existing techniques. Both image based and video based performance measures
are considered for analysis. The proposed Running gaussian background
subtraction technique for target detection is compared with Temporal frame
differencing, Running average, Temporal median filtering techniques and the
proposed Adaptive background mean shift tracking algorithm for target
 70

tracking is compared with the traditional mean shift tracking technique and
continuously adaptive mean shift technique. While implementing CAMShift
algorithm and Traditional mean shift algorithm with the same input data used
for proposed Adaptive background mean shift tracking the two algorithms
result in target loss problem shown in Figure 4.30. This is due to the fact that
the algorithms cannot cope with the frame rate of 30 FPS due to their design
for stable background. Thus to make the algorithms work for dynamic platform
of aerial videos, the input frame rate is reduced to 20 FPS to the two algorithms
to make a pseudo arrangement that the background is reasonably stable to
detect and track the target. Thus the input frame rate for the algorithms used in
target tracking is as follows

 Traditional Mean shift tracking – 20 FPS

 Continuously Adaptive Mean shift tracking – 20 FPS

 Adaptive background mean shift tracking – 30 FPS

The sample frames for the results of implementation of algorithms

for each dataset is shown separately. Their respective performance measures
are depicted both numerically and graphically. The Table 4.1 and 4.2 shows the
description and the properties of the video data considered for implementation.

This chapter is organized as follows. Section 4.1 deals with the

description and relevant equations of the performance metrics employed to
assess the efficiency of target detection and tracking. Section 4.2 includes the
results of target detection algorithms for 10 different data set and their
respective object based performance assessment and discussions with Tables
and plots. Section 4.3 gives the results of target tracking algorithms for 10
different data set and corresponding frame based performance assessment and
discussions with Tables and plots.
 71

4.1 PERFORMANCE EVALUATION

Performance evaluation has become a vital feature for researchers

to assess the efficiency and reliability of the design or algorithms. The
performance analysis is done for the following reasons

 To demonstrate the correctness and robustness of the proposed

technique.

 To allow comparison with alternative algorithms of

remarkably same passion.

 To assess the improvements in performance of the developed

algorithm.

In video processing, performance evaluation can be done in each

frame and also frame by frame. In object based performance metrics each
frame is considered and the measures are evaluated. In frame based metrics
the measures are evaluated for consequent frames. As in Figure 4.1, object
based performance evaluation is done for target detection techniques and
frame based performance evaluation is done for target tracking techniques.

Variables / Parameters / Attributes

Ground Truth
IN
P Object Detection
U EVALUATION
T METHODOLOGY
D
A
Object Tracking
T
A Metrics

Variables / Parameters / Attributes

Variables / Parameters / Attributes

Figure 4.1 Performance Evaluation flow diagram

 Table 4.1 Properties of Input video data

Video Description Altitude Duration Size FPS No of

Data (Seconds) frames
Data 1 Random movement of target Low 8 856 x 480 30 241
Data 2 Cluttered background, regular speed, isolated target in High 6 856 x 480 30 180
curved motion
Data 3 Isolated target with regular speed High 9 720 x 480 30 286
Data 4 Cluttered background, regular speed, multiple objects, partial Low 7 856 x 480 30 212
intrusion to the tracker, merge and split
Data 5 Regular speed, changing size of target Low 12 432 x 240 30 360
Data 6 Regular speed, multiple objects similar to target High 8 720 x 480 30 242
Data 7 Regular speed, varying size of target, change in background Low 15 856 x 480 30 451
Data 8 Stationary for a few seconds and then moving, regular size Low 4 720 x 480 30 105
and smooth background
Data 9 Smooth background, to and fro motion of target Low 18 856 x 480 30 541
Data 10 Isolated target, regular speed, very small in size High 14 856 x 480 30 461

72
 Table 4.2 Target model and size

Video Target Target Target PDF Window size Initial target Minimum Maximum
Data model Size Target Size target size

3082 Pixels 1581 Pixels 1000 Pixels 2898 Pixels

Data 1 A cow

Data 2 A Car 999 Pixels 299 Pixels 342 Pixels 650 Pixels

Data 3 A Car 810 Pixels 209 Pixels 56 Pixels 273 Pixels

Data 4 A Bus 1008 Pixels 636 Pixels 490 Pixels 1054 Pixels

Data 5 A Car 5180 Pixels 2652 Pixels 1221 Pixels 2862 Pixels

73
 Table 4.2 (Continued)

Data 6 A car 324 Pixels 84 Pixels 77 Pixels 646 Pixels

Two
Data 7 744 Pixels 228 Pixels 180 Pixels 610 Pixels
persons

Data 8 A boat 4029 Pixels 324 Pixels 255 Pixels 399 Pixels

A
Data 9 2520 Pixels 784 Pixels 644 Pixels 936 Pixels
person

Data 10 A Bike 208 Pixels 45 Pixels 21 Pixels 63 Pixels

74
 75

4.1.1 Object based performance metrics

Object based performance evaluation is done for target detection

algorithms. The approach is dependent on the pixel intensity of the
background model in each frame. The background model is compared with
the reference frame (initial input frame). Similarity metric, noise metrics and
detection rate are considered for evaluation of the proposed adaptive Gaussian
background subtraction with a few well established techniques like temporal
frame differencing, running average background subtraction and temporal
median filter.

The Structural Similarity measure (SSIM) is the comparison

between the reference model and the output model. The index consists of 3
factors namely luminance model, contrast model and structural model. These
performance metrics are discussed by Premakumar et al (2011), Aja
Fernandez et al (2006) and Wang & Bovik (2002).

Let two real valued pixel data be x  x1 , x2 ,...xn  , y   y1 , y2 ,... yn  .

‘x’ is the reference data and ‘y’ is the result data. x be the mean,  x2 be the

variance and  x be the standard deviation of data ‘x’, y be the mean,  y2 be

the variance and  y be the standard deviation of data ‘y’,  xy be the

covariance of x and y.

1 n
 
2
x 
n  1 i 1
( xi  x )2 (4.1)

1 n
 y2   ( yi  y )2
n  1 i 1
(4.2)

1 n
 xy   ( xi  x )( yi  y )
n  1 i 1
(4.3)
 76

The luminance model (LM) is defined as

2xy
LM  (4.4)
x  y2
2

The contrast model (CM) is defined as

2 x y
CM  (4.5)
 x2   y2

The structural model (SM) is defined as

 xy
CM  (4.6)
 x y

Based on these three measures SSIM is computed as

 xy 2 x y 2 xy
SSIM  (4.7)
 x y  x2   y2 x 2  y 2

The mean SSIM (MSSIM) for k number of frames is defined as

1 k
MSSIM   SSIM ( x j , y j )
k j 1
(4.8)

The noise based metrics namely Peak signal to noise ratio (PSNR)
and mean square error (MSE) are calculated by the following equations with
M, N as the size of the image, (i,j) as the position of pixel.

M N
1
MSE 
MN
 ( x
i 1 j 1
ij  xij )2 (4.9)
 77

 2552 
PSNR  10log   (4.10)
 MSE 

The Detection rate is the ratio of total number of frames the target
is detected to the total number of frames in video.

(4.11)

The results of the object based performance metrics are calculated

and tabulated. The metrics are depicted as graph for easy observation.

4.1.2 Frame based performance metrics

Frame based performance metrics are based on ground truth.

Ground truth tends to provide an abstract data (location, size, type) that can
be made relative to the observations extracted from the video frames. The
metrics are discussed by Bashir & Porikli (2006). Ground truth shall be,

1. An indication of location of target by circle or bounding box

2. An accurate mark of target boundary

3. Information classification from generated result.

The ground truth based evaluation relies on four basic

terminologies namely True Positive (TP), False Positive (FP), True Negative
(TN) and False Negative (FN).

 True Positive - Both ground truth and result agree that there is
presence of target or bounding box of result and ground truth
coincides.
 78

 True Negative - Both ground truth and result agree the

absence of target.

 False Positive - Tracker detects object but ground truth has no

object or no system object fall within bounding box of ground
truth object

 False Negative - Ground truth detects object but tracker

doesn’t detect any object.

Based on these metrics the parameters namely completeness (CS),

False alarm rate (FAR), Similarity (SS) and Accuracy (ACC) are evaluated
for Traditional Mean shift tracking, Continuously adaptive mean shift
tracking (CAMS) and the proposed Adaptive background Mean shift tracking
in total number of frames (TF). The results are tabulated for each video
dataset and their respective graphs are plotted.

T
Completeness (CS)= (4.12)
T +FN

F
False Alarm Rate (FAR)= (4.13)
T +F

T
imilarity(SS)= (4.14)
T +F +FN

T +TN
Accuracy(ACC)= (4.15)
TF

A parameter namely duration of successful tracking (DOST) is

evaluated based on the ratio between number of frames the target is tracked
successfully and the total number of frames the target is visible.

No.of seconds target is successfully tracked

OST= (4.16)
No.of seconds in video

The tracking rate (TR) is calculated as the number of frames

tracked correctly in the total number of input frames.
 79

No.of frames successfully tracked

TR= (4.17)
No.of frames in video

The results of the frame based performance metrics are calculated

for traditional mean shift technique, continuously adaptive mean shift
technique and proposed Adaptive background mean shift technique and
tabulated. The metrics are depicted as graph for ease of observation.

4.2 IMPLEMENTATION OF TARGET DETECTION

ALGORITHMS

In this section, the results of proposed Running Gaussian

background subtraction technique for target detection for 10 different video
data is shown and discussed as a comparison with temporal median filtering
technique, running average background subtraction and traditional frame
differencing technique.

4.2.1 Target detection results for dataset 1

The video data 1 consists of a target to be tracked (a cow) with

random movement. The movement of the target is non-deterministic. The
target defined is changing in shape, orientation and speed. The video is a low
altitude angular video taken at Kovalam beach, Chennai, India with the
experimental bias. Properties of the video are listed in Table 4.1 and 4.2. The
results are displayed in Figures 4.2 (a, b, c, d, and e) and Table 4.3.

Figure 4.2 (a) is the initial frame of the input video of dataset 1.
There are number of objects (cows) in the scene. The white cow is considered
as the target to be detected and tracked. In Figure 4.2 (b) and (c) a sample
background model and background update is shown respectively. The sample
background subtraction results are shown in Figure 4.2 (d) and (e).
 80

(a)

(b) (c)

(d) (e)

Figure 4.2 (a) Initial frame for input video 1 (b) Background model of
Sample frame 4 (c) Background update of sample frame 4
(d) Background subtraction of sample frame 4 (e)
Background subtraction of sample frame 59.
 81

Table 4.3 Target detection metrics for dataset 1

LM CM SM MSSIM MSE PSNR DR

TFD 0.941 0.945 0.068 0.061 2.735 43.69 62.23
RABS 0.982 0.985 0.072 0.070 2.413 44.32 65.57
TMF 0.985 0.982 0.075 0.072 2.159 44.81 67.21
RGBS 0.999 0.997 0.082 0.082 1.950 45.23 75.41

At initial stage, when the UAV is in hovering state and target also
remains static, the PSNR value for detection techniques is high. After a few
frames the target starts moving and the UAV also starts following the target.
The motion of both sensor and target results in ego noise and thus the
efficiency reduces to an extent. Although the value of metrics goes down it is
higher than other technique. The plot for PSNR in Figure 4.3 shows the
variation in metrics.

80

60
PSNR(%)

TFD
40
RABS
20
TMF
0 AGBS
110

130

150

170

190

210

230
10

30

50

70

90

Number of Frames

Figure 4.3 PSNR for Dataset 1

4.2.2 Target detection results for dataset 2

The video data 2 consists of a vehicle (car) as target to be tracked

in a cluttered environment. The target moves in a bridge in regular speed in
curved direction. The video is a moderately high altitude angular video taken
by UAV at Chennai Kathippara fly over situated at the heart of Chennai city,
Tamilnadu, India. The video is taken for the purpose of traffic monitoring and
surveillance. The result metrics of the video are listed in Table 4.4.
 82

(a)

(b) (c)

(d) (e)

Figure 4.4 (a) Initial frame for input video 2 (b) Background model of
Sample frame 8 (c) Background update of sample frame 9
(d) Background subtraction of sample frame 3
(e) Background subtraction of sample frame 21.
 83

Figure 4.4 (a) is the initial frame of the input video of dataset 2. A car
on the flyover is the target to be tracked. The car is in motion right from the initial
frame till the end. The motion is curvilinear. In Figure 4.4 (b) and (c) a sample
background model and background update is shown respectively. The sample
background subtraction results are shown in Figure 4.4 (d) and (e).

Table 4.4 Target detection metrics for video data set 2

LM CM SM MSSIM MSE PSNR DR

TFD 0.95 0.956 0.0127 0.012 0.048 61.32 74.19
RABS 0.98 0.975 0.0137 0.013 0.044 61.69 74.19
TMF 0.985 0.98 0.0149 0.014 0.039 62.00 80.6
RGBS 1 0.995 0.0164 0.016 0.036 62.6 87.1

Right from the initial frame, the UAV is in regular motion tracking
the car moving in curvilinear motion in regular speed. Since there is no
reasonable noise due to regular motion pattern of both sensor and target, the
PSNR value for detection techniques is high.

80

60
PSNR(%)

TFD
40
RABS
20
TMF
0 AGBS
100
110
120
130
140
150
160
170
180
20
10

30
40
50
60
70
80
90

Number of frames

Figure 4.5 PSNR for Dataset 2

4.2.3 Target detection results for dataset 3

In video data 3, an isolated target (a car) with regular movement is

to be tracked. The movement of the target is regular and defined. The target
 84

moves in a high way road with no other intrusions in a regular speed. The
video is a high altitude angular video taken by UAV at military base Orrisa,
India with the experimental purpose. The result metrics of the video are listed
in Table 4.5.

(a)

(b) (c)

Figure 4.6 (a) Initial frame for input video 3 (b) Background model of
Sample frame 6 (c) Background update of sample frame 6
(d) Background subtraction of sample frame 3
(e) Background subtraction of sample frame 49.
 85

(d) (e)

Figure 4.6 (Continued)

Figure 4.6 (a) is the initial frame of the input video of dataset 3. A
car in a highway is the target to be tracked. The car is in motion right from the
initial frame till the end. The motion is linear and regular. In Figure 4.6 (b)
and (c) a sample background model and background update is shown
respectively. The sample background subtraction results are shown in Figure
4.6 (d) and (e).

Table 4.5 Target detection for video data set 3

LM CM SM MSSIM MSE PSNR DR

TFD 0.98 0.979 0.08 0.077 2.66 43.88 68.33
RABS 0.98 0.998 0.093 0.089 2.44 44.26 71.66
TMF 0.99 0.99 0.099 0.097 2.25 44.60 71.66
RGBS 0.995 0.996 0.113 0.112 1.95 45.23 76.7

Right from the initial frame, the UAV is in regular motion

tracking the car moving in linear motion in regular speed. Even though there
is a regular motion in the pattern of both sensor and target, as the altitude is
very high, the size of target becomes very small. Also other elements in the
video gets included as objects.
 86

The atmospheric mist is observable in the video. Thus the PSNR

value for detection techniques is lesser. Anyhow, the proposed technique
outfits other techniques with quotable margin in noise rejection. The Figure
4.7 and shows the variation in metrics of PSNR.

70

60

50
PSNR (%)

40 TFD
30 RABS

20 TMF

10 RGBG

250
110

130

150

170

190

210

230

270
10

30

50

70

90

Number of frames

Figure 4.7 PSNR for Dataset 3

4.2.4 Target detection results for dataset 4

The video data 4 consists of a vehicle (bus) to be tracked in a

cluttered environment. The target moves in a bridge in regular speed.
Meanwhile a light post interfere the tracker by hiding a part of the target. The
video is a low altitude angular video taken by UAV at Chennai Kathippara fly
over situated at the heart of Chennai city, India. The video is taken for the
purpose of traffic monitoring and surveillance. The resultant metrics of the
video are listed in Table 4.6.
 87

(a)

(b) (c)

(d) (e)
Figure 4.8 (a) Initial frame for input video 4 (b) Background model of
Sample frame 6 (c) Background update of sample frame 9
(d) Background subtraction of sample frame 3 (e)
Background subtraction of sample frame 21.
 88

Table 4.6 Target detection metrics for video data set 4

LM CM SM MSSIM MSE PSNR DR

TFD 0.975 0.97 0.018 0.017 0.20 55.10 65
RABS 0.98 0.985 0.019 0.018 0.18 55.57 68
TMF 0.99 0.985 0.019 0.018 0.17 55.70 70
RGBS 1.00 0.995 0.022 0.021 0.15 55.17 78

Figure 4.8 (a) is the initial frame of the input video of dataset 4. A
bus in a bridge with a cluttered background is the target to be detected. The
bus is in motion right from the initial frame till the end. The motion is linear
and regular with some interference which hides the target partially in few
frames. In Figure 4.8 (b) and (c) a sample background model and background
update is shown respectively. The sample background subtraction results are
shown in Figure 4.8 (d) and (e).

From the initial frame, the UAV is in hovering state and then in
slow motion thereby tracking the bus moving in linear motion in regular
speed. Even though there is a regular motion pattern of both sensor and target,
as the background is cluttered, the appearance of target is occluded. Thus the
PSNR value for detection techniques is lesser. Anyhow, the proposed
technique outfits other techniques with quotable margin in noise rejection.

The cluttered environment and varying color distribution of

different objects results in inclusion of noise parameters. The Figure 4.9
shows the variation in metrics.
 89

70

60

50
PSNR (%)

40 TFD
30 RABS

20 TMF
RGBS
10

100
110
120
130
140
150
160
170
180
190
200
210
10
20
30
40
50
60
70
80
90
Number of frames

Figure 4.9 PSNR for dataset 4

4.2.5 Target detection results for dataset 5

The video data 5 consists of a vehicle (car) to be tracked. The target

moves in regular speed in defined direction. The UAV follows the car by
tracking it. As the target moves, its size varies from large to smaller. The
video is a low altitude angular video taken by UAV at kovalam beach,
Chennai India. The video is taken for the purpose of analysis and
experimentation. The metrics of the video are listed in Table 4.7.

Figure 4.10 (a) is the initial frame of the input video of dataset 5.
A car with a relatively smooth background is the target to be detected. The
car is in motion right from the initial frame till the end. The motion is linear
and regular. In Figure 4.10 (b) and (c) a sample background model and
background update is shown respectively. The sample background subtraction
results are shown in Figure 4.10 (d) and (e).
 90

(a)

(b) (c)

(d) (e)
Figure 4.10 (a) Initial frame for input video 5 (b) Background model of
Sample frame 8 (c) Background update of sample frame 9
(d) Background subtraction of sample frame 3 (e)
Background subtraction of sample frame 21.
 91

Table 4.7 Target detection for video data set 5

LM CM SM MSSIM MSE PSNR DR

TFD 0.975 0.96 0.021 0.021 0.86 48.7 69
RABS 0.98 0.97 0.022 0.021 0.78 49.0 71.4
TMF 0.985 0.97 0.023 0.022 0.71 49.6 73.8
RGBS 0.99 0.98 0.025 0.025 0.53 50.9 80.9

From the initial frame, the UAV is in motion thereby tracking the
car moving in linear motion in regular speed. The movement of the sensor is
irregular producing ego noise. Thus the PSNR value for detection techniques is
lesser. Anyhow, the proposed technique outfits other techniques with quotable
margin in noise rejection. The Figure 4.11 shows the variation in PSNR.

80
60
PSNR (%)

TFD
40
RABS
20
TMF
0
RGBS
170

270
110
130
150

190
210
230
250

290
310
330
350
70
10
30
50

90

Number of frames

Figure 4.11 PSNR for dataset 5

4.2.6 Target detection results for dataset 6

The video 6 data consists of a vehicle (car) to be tracked in a

cluttered environment. The target moves in a bridge in regular speed. There
are multiple vehicles similar to the target possessing the same property as that
of the target. The target is intruded by a light post (merge and split). The
video is a moderately high altitude angular video taken by UAV at Chennai
Kathippara fly over situated at Chennai city, India. The video is taken for the
purpose of traffic monitoring and surveillance. The result metrics of the video
are listed in Table 4.8.
 92

(a)

(b) (c)

(d) (e)

Figure 4.12 (a) Initial frame for input video 6 (b) Background model of
Sample frame 8 (c) Background update of sample frame 6
(d) Background subtraction of sample frame 4 (e)
Background subtraction of sample frame 26.
 93

Table 4.8 Target detection for video data set 6

LM CM SM MSSIM MSE PSNR DR

TFD 0.97 0.96 0.023 0.021 0.24 54.3 68.7
RABS 0.98 0.97 0.025 0.024 0.21 54.9 74.0
TMF 0.99 0.98 0.026 0.025 0.20 55.1 77.0
RGBS 1.00 0.998 0.028 0.028 0.18 55.6 83.9

Figure 4.12 (a) is the initial frame of the input video of dataset 6.
A car is the target to be detected from high altitude. The car is in motion right
from the initial frame till the end. The motion is linear and regular. In Figure
4.12 (b) and (c) a sample background model and background update is shown
respectively. The sample background subtraction results are shown in Figure
4.12 (d) and (e). From the initial frame, the UAV is in motion thereby
tracking the car moving in linear motion in regular speed. The movement of
the sensor is regular but the altitude is very high. The background is cluttered
and multiple targets are present in the scene. Thus the PSNR value for
detection techniques relatively is lesser. The Figure 4.13 shows the variation
of PSNR metrics for different algorithms.

70
60
50
PSNR (%)

40 TFD
30 RABS

20 TMF

10 RGBS

0
10 30 50 70 90 110 130 150 170 190 210 230
Number of frames

Figure 4.13 PSNR for dataset 6

 94

4.2.7 Target detection results for dataset 7

The video data 7 consists of two persons walking in the seashore as

target to be detected and tracked. The target moves in regular speed in defined
direction. As the target moves, its size varies from smaller to larger. Also
since the target is moving in shore, the background varies randomly (sea
water and sand). The video is a low altitude angular video taken by UAV at
Chennai beach, India. The video is taken for the purpose of coastal security
and surveillance. The result metrics of the video are listed in Table 4.9.

(a)

(b) (c)

Figure 4.14 (a) Initial frame for input video 7 (b) Background model of
Sample frame 48 (c) Background update of sample frame 39
(d) Background subtraction of sample frame 46 (e)
Background subtraction of sample frame 63.
 95

(d) (e)

Figure 4.14 (Continued)

Figure 4.14 (a) is the initial frame of the input video of dataset 7.
Two persons walking in sea shore are the targets to be detected. The target is
in motion right from the initial frame till the end. The motion is linear and
regular. In Figure 4.14 (b) and (c) a sample background model and
background update is shown respectively. The sample background subtraction
results are shown in Figure 4.14 (d) and (e).

Table 4.9 Target detection metrics for video data set 7

LM CM SM MSSIM MSE PSNR DR

TFD 0.96 0.96 0.0071 0.0065 0.110 57.2 73.33

RABS 0.985 0.97 0.0074 0.0071 0.101 58.1 76.67

TMF 0.98 0.98 0.0076 0.0073 0.098 58.2 78.3

RGBS 1.0 0.996 0.0098 0.0098 0.088 58.7 83.3

From the initial frame, the UAV is in motion thereby tracking the
target moving in linear motion in regular speed. The movement of the sensor
is regular but the background is changing. The size or the target varies and
 96

there is variation in altitude. The wave also gets included as object due to
their differentiated intensity. Anyhow the PSNR value for detection
techniques relatively is good due to the fact that the target is nearer. The
Figure 4.15 shows the variation of PSNR for different algorithms.

80

70

60

50
PSNR (%)

TFD
40
RABS
30
TMF
20 RGBS
10

0
170
110
130
150

190
210
230
250
270
290
310
330
350
370
390
410
430
450
10
30
50
70
90

Number of frames

Figure 4.15 PSNR for dataset 7

4.2.8 Target detection results for dataset 8

The video data 8 consists of a boat sailing in the sea as the target to
be tracked. The target is stationary for a few frames and later it moves in
regular speed in defined direction. The background is smooth and regular.

The video is a low altitude angular video taken by UAV at Chennai

beach, India. The video is taken for the purpose of coastal security and
surveillance. The result metrics of the video are listed in Table 4.10.
 97

(a)

(b) (c)

(d) (e)
Figure 4.16 (a) Initial frame for input video 8 (b) Background model of
Sample frame 4 (c) Background update of sample frame 3
(d) Background subtraction of sample frame 4 (e)
Background subtraction of sample frame 63.
 98

Table 4.10 Target detection metrics for video data set 8

LM CM SM MSSIM MSE PSNR DR

TFD 0.97 0.96 0.008 0.007 0.05 61.14 83.81
RABS 0.98 0.975 0.009 0.008 0.042 61.89 85.7
TMF 0.99 0.985 0.010 0.009 0.04 62.11 87.62
RGBS 1.00 1.00 0.011 0.011 0.036 62.6 91.4

Figure 4.16 (a) is the initial frame of the input video of dataset 8.
A boat in sea is the target to be detected. The motion is linear and regular. In
Figure 4.16 (b) and (c) a sample background model and background update is
shown respectively. The sample background subtraction results are shown in
Figure 4.16 (d) and (e).

At initial frame, the UAV is in hovering state and the boat is also
static. Thus PSNR is high at initial stages. After few frames as the boat moves
the noises due to motion gets included. But, due to even and smooth
background the boat is detected efficiently with higher PSNR. The Figure
4.17 shows the variation of PSNR for different algorithms.

70
60
50
PSNR (%)

40 TFD
30 RABS
20 TMF
10 RGBG
0
100

105
20
10

30

40

50

60

70

80

90

Number of frames

Figure 4.17 PSNR for dataset 8

 99

4.2.9 Target detection results for dataset 9

The video data 9 consists of a walking person as target to be

tracked. The target moves in a few frames in forward direction, it stays
stationary for few frames and moves in backward direction in the remaining
frames. The background is smooth and regular. The video is a low altitude
angular video taken by UAV at Chengalpattu forest terrain, India. The video
is taken for the purpose of forest security and surveillance. The result metrics
of the video are listed in Table 4.11.

(a)

(b) (c)
Figure 4.18 (a) Initial frame for input video 9 (b) Background model of
Sample frame 48 (c) Background update of sample frame 38
(d) Background subtraction of sample frame 16 (e)
Background subtraction of sample frame 143.
 100

(d) (e)

Figure 4.18 (Continued)

Figure 4.18 (a) is the initial frame of the input video of dataset 9.
A walking person is the target is to be detected. The motion is non linear and
irregular. In Figure 4.18 (b) and (c) a sample background model and
background update is shown respectively. The sample background subtraction
results are shown in Figure 4.18 (d) and (e).

At initial stages, the UAV is in hovering state and the person to be

detected and tracked is also static. Thus PSNR is high at initial stages. After
few frames the person moves forward and later he moves sidewise. The grass
and the soil components are relatively bigger to the target and hence interfere
as noise with the algorithm. The Figure 4.19 shows the variation of PSNR for
different algorithms

Table 4.11 Target detection metrics for video data set 9

LM CM SM MSSIM MSE PSNR DR

TFD 0.965 0.97 0.014 0.013 0.24 54.33 68.8
RABS 0.97 0.975 0.015 0.014 0.23 54.5 71.8
TMF 0.98 0.985 0.016 0.016 0.22 54.7 75.0
RGBS 1.00 0.995 0.018 0.018 0.19 55.4 81.3
 101

.
80
70
60
50
PSNR (%)

TFD
40
RABS
30
TMF
20
RGBS
10
0

310

490
100
130
160
190
220
250
280

340
370
400
430
460

520
10
40
70

Number of frames

Figure 4.19 PSNR for dataset 9

4.2.10 Target detection results for dataset 10

The video data 10 consists of an isolated target to be detected and

tracked (a person in a bike) with regular movement. The target moves in a
bridge with no other intrusions at a regular speed.

The target to be tracked is very small in size to detect and identify.

The video is a high altitude angular video taken by UAV at Chennai beach for
coastal surveillance and security. The result metrics of the video are listed in
Table 4.12. The presence of water makes the background relatively smooth.
More over the altitude makes the information present in the video less visible
which helps in reduction of noise but adversely effects in detection of target.
 102

(a)

(b) (c)

(d) (e)

Figure 4.20 (a) Initial frame for input video 10 (b) Background model of
Sample frame 16 (c) Background update of sample frame 6
(d) Background subtraction of sample frame 3 (e)
Background subtraction of sample frame 49.
 103

Table 4.12 Target detection metrics for video data set 10

LM CM SM MSSIM MSE PSNR DR

TFD 0.975 0.97 0.007 0.007 0.026 63.9 68.75
RABS 0.98 0.985 0.008 0.008 0.024 64.2 71.88
TMF 0.99 0.98 0.008 0.008 0.024 64.2 75.0
RGBS 1.00 1.00 0.010 0.010 0.023 64.5 84.38

Figure 4.20 (a) is the initial frame of the input video of dataset 10.
An isolated bike in the bridge is the target is to be detected. The motion is non
linear and irregular. In Figure 4.20 (b) and (c) a sample background model
and background update is shown respectively. The sample background
subtraction results are shown in Figure 4.20 (d) and (e).

From the starting frame, the target is moving and the UAV is
following the target. Thus PSNR is high at initial stages. The size is very
small and hence detection of target becomes complicated. The Figure 4.21
shows the variation of PSNR metrics for different algorithms.

80
70
60
50
PSNR (%)

TFD
40
RABS
30
TMF
20
RGBS
10
0
10 40 70 100 130 160 190 220 250 280 310 340 370 400 430 460
Number of frames

Figure 4.21 PSNR for dataset 10

 104

4.2.11 Plots for target detection techniques

The plots for different performance metrics for target detection

algorithms are shown from Figures 4.22 to 4.28.

1.01
1
0.99
0.98
0.97 TFD
LM

0.96
RABS
0.95
0.94 TMF
0.93 RGBS
0.92
0.91

Video 10
Video 1

Video 2

Video 3

Video 4

Video 5

Video 6

Video 7

Video 8

Video 9

Figure 4.22 Luminance model for target detection techniques

1.01
1
0.99
0.98
0.97
0.96 TFD
CS

0.95 RABS
0.94 TMF
0.93
RGBS
0.92
Video 10
Video 1

Video 2

Video 3

Video 4

Video 5

Video 6

Video 7

Video 8

Video 9

Figure 4.23 Contrast model for target detection techniques

 105

0.12

0.1

0.08
TFD
SS

0.06
RABS
0.04 TMF
RGBS
0.02

Video 10
Video 1

Video 2

Video 3

Video 4

Video 5

Video 6

Video 7

Video 8

Figure 4.24 Similarity model for target detection techniques Video 9

0.12

0.1

0.08
MSSIM

TFD
0.06
RABS
0.04 TMF
RGBS
0.02

0
Video 10
Video 1

Video 2

Video 3

Video 4

Video 5

Video 6

Video 7

Video 8

Video 9

Figure 4.25 Mean SSIM for target detection techniques

 106

2.5

2
TFD
1.5
MSE

RABS
1 TMF
RGBS
0.5

Video 10
Video 4
Video 1

Video 2

Video 3

Video 5

Video 6

Video 7

Video 8

Figure 4.26 Mean square error for target detection techniques Video 9

70

60

50
PSNR ( % )

40 TFD

30 RABS
TMF
20
RGBS
10

0
Video 10
Video 1

Video 2

Video 3

Video 4

Video 5

Video 6

Video 7

Video 8

Video 9

Figure 4.27 Peak signal to noise ratio for target detection techniques
 107

100
90
80
70
60
DR (%)

TFD
50
RABS
40
30 TMF

20 RGBS
10
0

Video 10
Video 1

Video 2

Video 3

Video 4

Video 5

Video 6

Video 7

Video 8

Video 9
Figure 4.28 Detection rate for target detection techniques

In the above results, the value of PSNR is comparatively good for

running Gaussian background subtraction compared to other techniques. Here
the value of ‘α’ is maintained as 0.6. The threshold of 100 is used in frame
differencing technique. The motion of both the UAV and object causes noises
like occlusion and ego motion. Anyhow, the detection rate is high compared
to other algorithms which are very effective in static background. The
maximum PSNR value ranges around 65 which happen at the initial stages of
background. This is because, at initial stages when the background model is
updated it will be more similar to sample frame. But as the frame count
increases, the new model will be based on previous model and present frame.
Thus there will be a minor degradation in the background model definition.
This will be higher if the value of ‘α’ is used as variable instead of constant.
Thus the optimal value of ‘α’ is determined as 0.6 (In Chapter 3). The
efficiency of detection of moving platform is higher than other algorithms.
 108

The interpretations from the results are as follows

 The detection of target mainly depends on the background of

the frame and background being modeled. The size of the
target affects the detection rate of the object.

 If the background is smooth then the detection rate is higher

and it is lower for cluttered background. Presence of multiple
colors introduces noise elements due to varying intensity.

 Since the background model degrades with the increase in

frame number changing the tuning factor will not help in
efficiency of detection.

 The orientation, change in direction and motion pattern

influences the detection algorithms. Thus these problems are
to be faced by tracking algorithms to enhance tracking
efficiency, which is the ultimate aim of this work.

4.3 IMPLEMENTATION OF TARGET TRACKING

ALGORITHMS

4.3.1 Target tracking results for dataset 1

The results of target tracking algorithm (ABMST) for video data

set 1 is shown in Figures 4.29 (a, b, c, d and e). The Figure 4.29 (a) shows the
sample model of transition of elements between two consecutive frames. This
transition model depicts the change in position of each element in consecutive
frames. Based on the change in position the new background is estimated and
shifted. The green color bounding box tracks the target in consecutive frames.
Four sample frames out of total 241 frames are shown below in Figure 4.29
(b, c, d and e).
 109

(a)

(b) (c)

(d) (e)
Figure 4.29 (a) Transition model between two sample frames (b)
ABMST result for video 1 sample frame 5 (c) ABMST result
for video 1 sample frame 16 (d) ABMST result for video 1
sample frame 82 (e) ABMST result for video 1 sample
frame 170
 110

In Figure 4.29 (b), it can be observed that the target is facing the
camera. At that time the UAV is in hovering state. Later the target changes its
direction (and thus its shape) and starts moving. The UAV also starts
following the target thereby causing transition in every element of the frame
(Figure 4.29 (a)). Anyhow the tracker is able to track the target with change in
orientation, speed and size.

The Figure 4.30 shows a sample of target loss problem faced by

traditional mean shift tracking at frame rate of 30 FPS for change in shape
and size of object with varying velocity. Thus while implementing and
evaluating the TMST and CAMST algorithms 20 FPS is used.

Figure 4.30 Target loss problem in TMST

The results of performance metrics for three target tracking

algorithms (TMST, CAMST and ABMST) for video data 1 are shown in
Table 4.13. These values are obtained for video 1 with 20FPS for CAMST
and TMST and 30 FPS for ABMST.
 111

Table 4.13 Target tracking metrics for video data set 1

CS FAR SS ACC DOST TR

T-MST 73.81 12.43 66.81 68.05 67.90 68.46
CAMST 75.32 09.53 68.47 70.21 68.86 70.00
ABMST 77.73 06.32 73.86 73.86 72.84 73.86

4.3.2 Target tracking results for dataset 2

The results of target tracking algorithm (ABMST) for video data

set 2 is shown in Figures 4.31 (a, b, c, d and e). The Figure 4.31 (a) shows the
sample model of transition of elements between two consecutive frames. The
video is high altitude data. The green color bounding box tracks the target in
consecutive frames. Four sample frames out of total 180 frames are shown
below in Figure 4.31 (b, c, d and e).

In Figure 4.31 (b), it can be observed that the target is at centre of

road. The size of the target is very small compared to the frame size. The
UAV is in hovering state. The target moves at regular velocity bur moves in
curvilinear direction. The UAV also starts following the target thereby
causing transition in every element of the frame in Figure 4.31 (a). Anyhow
the tracker is able to track the target with change in orientation, speed and
size. The results of performance metrics for three target tracking algorithms
(TMST, CAMST and ABMST) for video data 2 are shown in Table 4.14.
These values are obtained for video 2 with 20FPS for CAMST and TMST
and 30 FPS for ABMST. The DOST and TR are higher for ABMST. It is
achieved due to regular and non interfered movement of target with smooth
background.
 112

(a)

(b) (c)

(d) (e)
Figure 4.31 (a) Transition model between two sample frames (b)
ABMST result for video 2 sample frame 6 (c) ABMST result
for video 2 sample frame 92 (d) ABMST result for video 2
sample frame 115 (e) ABMST result for video 2 sample
frame 149
 113

Table 4.14 Target tracking metrics for video data set 2

CS FAR SS ACC DOST TR

T-MST 84.44 7.28 79.55 80.00 81.67 81.11
CAMST 85.36 5.92 81.66 82.2 83.33 82.23
ABMST 88.24 3.84 85.23 85.55 86..67 85.56

4.3.3 Target tracking results for dataset 3

The results of target tracking algorithm (ABMST) for video data

set 3 is shown in Figures 4.32 (a, b, c, d and e). The video is taken from high
altitude. The Figure 4.32 (a) shows the sample model of transition of elements
between two consecutive frames. The red color bounding box tracks the target
in consecutive frames. Four sample frames out of total 286 frames is shown
below in Figure 4.32 (b, c, d and e).

In Figure 4.32 (b), it can be observed that the target is moving in

highways at regular velocity. Since the altitude is very high, the size of the
target appears to be very small compared to the frame size. The size of the
target becomes still smaller and negligible in Figure 4.32 (e). The
background is smooth (black color road). The UAV is following the target
thereby causing transition in every element of the frame in Figure 4.32 (a).
the effectiveness in detection is higher due to clear and smooth background.

The results of performance metrics for three target tracking

algorithms (TMST, CAMST and ABMST) for video data 3 are shown in
Table 4.15. These values are obtained for video 3 with 20FPS for CAMST
and TMST and 30 FPS for ABMST.
 114

(a)

(b) (c)

(d) (e)
Figure 4.32 (a) Transition model between two sample frames (b)
ABMST result for video 3 sample frame 12 (c) ABMST
result for video 3 sample frame 116 (d) ABMST result for
video 3 sample frame 200 (e) ABMST result for video 3
sample frame 270
 115

Table 4.15 Target tracking metrics for video data set 3

CS FAR SS ACC DOST TR

T-MST 96.65 2.99 93.86 94.06 94.74 94.41
CAMST 97.12 2.13 95.34 95.26 95.16 95.23
ABMST 98.59 1.06 97.55 97.55 97.90 97.55

4.3.4 Target tracking results for dataset 4

The results of target tracking algorithm (ABMST) for video data

set 4 are shown in Figures 4.33 (a, b, c, d and e).

The Figure 4.33 (a) shows the sample model of transition of

elements between two consecutive frames. The red color bounding box tracks
the target in consecutive frames. Four sample frames out of total 212 frames
are shown below in Figure 4.33 (b, c, d and e). In Figure 4.33 (b), it can be
observed that the target is moving in a bridge at regular velocity. The altitude
is high and the view is angular. The UAV is following the target thereby
causing transition in every element of the frame in Figure 4.32 (a).

The target is intruded partially by a light post in Figure 4.33 (c).

The tracking algorithm neglects the intrusion by adapting the model from
next frame and continues tracking as in Figure 4.33 (d, e). The background is
cluttered. The results of performance metrics for three target tracking
algorithms (TMST, CAMST and ABMST) for video data 4 are shown in
Table 4.16. These values are obtained for video 4 with 20FPS for CAMST
and TMST and 30 FPS for ABMST. Lesser rate in tracking is sensed due to
cluttered nature of the scene. The metrics show that the performance of
adaptive background mean shift tracking is better than Traditional mean shift
tracking and Continuously adaptive mean shift tracking.
 116

(a)

(b) (c)

(d) (e)
Figure 4.33 (a) Transition model between two sample frames (b)
ABMST result for video 4 sample frame 2 (c) ABMST result
for video 4 sample frame 46 (d) ABMST result for video 4
sample frame 102 (e) ABMST result for video 4 sample
frame 129
 117

Table 4.16 Target tracking metrics for video data set 4

CS FAR SS ACC DOST TR

T-MST 93.27 9.09 85.31 85.38 84.51 84.91
CAMST 95.42 6.66 87.43 88.36 87.88 88.14
ABMST 97.02 4.88 92.42 92.46 91.54 91.98

4.3.5 Target tracking results for dataset 5

The results of target tracking algorithm (ABMST) for video data

set 5 is shown in Figures 4.34 (a, b, c, d and e). The Figure 4.34 (a) shows the
sample model of transition of elements between two consecutive frames. The
green color bounding box tracks the target in consecutive frames. Four
sample frames out of total 360 frames are shown below in Figure 4.34 (b, c, d
and e).

In Figure 4.34 (b), it can be observed that the target is moving in

ground at regular velocity. The altitude is low and the view is angular. In
Figure 4.34 (a), the UAV is following the target thereby causing transition in
every element of the frame. The background is smooth and regular. The target
changes its direction in a few frames at final stage of video in Figure 4.34 (e).
Thus there is a change in target size at later stages. The results of performance
metrics for three target tracking algorithms (TMST, CAMST and ABMST)
for video data 5 are shown in Table 4.17. These values are obtained for video
5 with 20FPS for CAMST and TMST and 30 FPS for ABMST. Regular and
defined motion, clear and smooth background and abstract structure of the
target helps in achieving higher DOST and TR.
 118

(a)

(b) (c)

(d) (e)
Figure 4.34 (a) Transition model between two sample frames (b)
ABMST result for video 5 sample frame 6 (c) ABMST result
for video 5 sample frame 66 (d) ABMST result for video 5
sample frame 146 (e) ABMST result for video 5 sample
frame 209
 119

Table 4.17 Target tracking metrics for video data set 5

CS FAR SS ACC DOST TR

T-MST 97.74 2.85 95.00 95.00 94.17 94.45
CAMST 98.12 1.12 96.26 96.33 95.66 95.72
ABMST 98.60 0.08 97.78 97.78 97.5 97.78

4.3.6 Target tracking results for dataset 6

The results of target tracking algorithm (ABMST) for video data

set 6 is shown in Figures 4.35 (a, b, c, d and e). The Figure 4.35 (a) shows the
sample model of transition of elements between two consecutive frames. The
red color bounding box tracks the target in consecutive frames. Four sample
frames out of total 242 frames is shown below in Figure 4.35 (b, c, d and e).

A car on the bridge is the target to be tracked. There are many cars
similar to the target. Initially the UAV captures from higher altitude. After
170 frames, the UAV lowers its altitude. It can be observed in Figure 4.35 (e).
The tracking algorithm still manages to track the target of interest amidst the
multiple objects and changing altitude. This disturbs the color model of the
target kernel, yet the algorithm tracks the target accurately. This is done by
adaptively compensating the background pixels based on the histogram
density spectrum. The results of performance metrics for three target
tracking algorithms (TMST, CAMST and ABMST) for video data 6 are
shown in Table 4.18. These values are obtained for video 1 with 20FPS for
CAMST and TMST and 30 FPS for ABMST. Regular motion, clear and
distinguishing background of the target helps in achieving higher DOST and
TR.
 120

(a)

(b) (c)

(d) (e)
Figure 4.35 (a) Transition model between two sample frames (b)
ABMST result for video 6 sample frame 6 (c) ABMST result
for video 6 sample frame 105 (d) ABMST result for video 6
sample frame 170 (e) ABMST result for video 6 sample
frame 201
 121

Table 4.18 Target tracking metrics for video data set 6

CS FAR SS ACC DOST TR

T-MST 92.61 0.04 88.75 88.84 88.89 88.84

CAMST 94.33 0.02 91.23 91.66 90.78 90.57

ABMST 96.21 0.01 95.00 95.05 95.06 95.05

4.3.7 Target tracking results for dataset 7

The results of target tracking algorithm (ABMST) for video data

set 7 are shown in Figures 4.36 (a, b, c, d and e). Four sample frames out of
total 451 frames are shown below in Figure 4.36 (b, c, d and e).

The Figure 4.36 (a) shows the sample model of transition of

elements between two consecutive frames. The green color bounding box
tracks the target in consecutive frames. It may be observed that the target to
be tracked is smaller in size and is moving in the sea shore. Thus the
background also changes as sea water and sand which can be observed at
frames of Figure 4.36 (b) and 4.36 (e) respectively. Thus there is change in
color density at irregular intervals of time. Multiple colors change rapidly in
the kernel like color of wave, color of sand, color of the wardrobe of the
target. The size of the target becomes bigger than the defined target model
size as the UAV moves towards the target. The ABMST technique still
manages to track the target efficiently. The results of performance metrics for
three target tracking algorithms for video data 7 with 20FPS for CAMST and
TMST and 30 FPS for ABMST are shown in Table 4.19. Regular and
distinguishing background from the target helps in achieving higher DOST
and TR.
 122

(a)

(b) (c)

(d) (e)
Figure 4.36 (a) Transition model between two sample frames (b)
ABMST result for video 7 sample frame 35 (c) ABMST
result for video 7 sample frame 86 (d) ABMST result for
video 7 sample frame 192 (e) ABMST result for video 7
sample frame 327
 123

Table 4.19 Target tracking metrics for video data set 7

CS FAR SS ACC DOST TR

T-MST 97.30 1.37 96.01 96.01 96.00 96.01

CAMST 97.88 1.00 97.44 97.66 97.33 97.50

ABMST 98.89 0.022 98.67 98.67 98.67 98.67

4.3.8 Target tracking results for dataset 8

The results of target tracking algorithm (ABMST) for video data

set 8 is shown in Figures 4.37 (a, b, c, d and e).

The Figure 4.37 (a) shows the sample model of transition of

elements between two consecutive frames. The red color bounding box tracks
the target in consecutive frames. Four sample frames out of total 105 frames
are shown below in Figure 4.37 (b, c, d and e).

The target is a boat sailing in the sea. The target possesses smooth
background and so detection rate is higher. The results of performance
metrics for TMST, CAMST and ABMST for video data 8 are shown in Table
4.20. These values are obtained for video 8 with 20 FPS for CAMST and
TMST and 30 FPS for ABMST.

Regular and defined motion, clear and smooth background and

abstract structure of the target helps in achieving higher DOST and TR.
 124

(a)

(b) (c)

(d) (e)
Figure 4.37 (a) Transition model between two sample frames (b)
ABMST result for video 8 sample frame 21 (c) ABMST
result for video 8 sample frame 71 (d) ABMST result for
video 8 sample frame 91 (e) ABMST result for video 8
sample frame 102
 125

Table 4.20 Target tracking metrics for video data set 8

CS FAR SS ACC DOST TR

T-MST 97.00 3.96 93.27 93.27 94.11 93.26

CAMST 98.33 1.85 95.34 94.88 94.32 93.39

ABMST 99.03 0.09 98.08 98.08 97.06 98.08

4.3.9 Target tracking results for dataset 9

The results of target tracking algorithm (ABMST) for video data

set 9 are shown in Figures 4.38 (a, b, c, d and e).

The Figure 4.38 (a) shows the sample model of transition of

elements between two consecutive frames. The green color bounding box
tracks the target in consecutive frames. Four sample frames out of total 541
frames are shown below in Figure 4.38 (b, c, d and e).

The target to be tracked is a man walking on a smooth background.

The target moves forward and then towards right side. This orientation can be
observed in frame 4.38 (b) and 4.38 (e) respectively. The ABMST algorithm
efficiently tracks the irregular motion pattern of the target. The results of
performance for three target tracking algorithms with 20FPS for CAMST and
TMST and 30 FPS for ABMST for video data 9 are shown in Table 4.21.
Irregular motion, cluttered and more information embedded background
results in relatively lesser DOST and TR.
 126

(a)

(b) (c)

(d) (e)
Figure 4.38 (a) Transition model between two sample frames (b)
ABMST result for video 9 sample frame 35 (c) ABMST
result for video 9 sample frame 135 (d) ABMST result for
video 9 sample frame 361 (e) ABMST result for video 9
sample frame 382
 127

Table 4.21 Target tracking metrics for video data set 9

CS FAR SS ACC DOST TR

T-MST 95.27 2.36 93.16 93.16 93.33 93.16

CAMST 95.00 2.88 93.45 93.67 93.67 93.88

ABMST 96.85 0.03 96.49 96.49 96.11 96.49

4.3.10 Target tracking results for dataset 10

The results of target tracking algorithm (ABMST) for video data

set 10 are shown in Figures 4.39 (a, b, c, d and e).

The Figure 4.39 (a) shows the sample model of transition of

elements between two consecutive frames. Four sample frames out of total
461 frames are shown below in Figure 4.39 (b, c, d and e). A bike over the
bridge is the target to be tracked. The size of the target is negligibly small
compared to the frame size. The video is taken from high altitude, so there is
a problem in detecting the target initially. The altitude of the scene increases
with increase in frame count, resulting in diminishing size of target from 63
pixels to 21 pixels thereby increasing the area of background in the window.
This disturbs the color model of the target kernel, yet the algorithm tracks the
target accurately. This is done by adaptively compensating the background
pixels based on the histogram density spectrum. The results of performance
metrics for three target tracking algorithms with 20FPS for CAMST and
TMST and 30 FPS for ABMST for video data 10 are shown in Table 4.22.
 128

(a)

(b) (c)

(d) (e)
Figure 4.39 (a) Transition model between two sample frames (b)
ABMST result for video 10 sample frame 16 (c) ABMST
result for video 10 sample frame 113 (d) ABMST result for
video 10 sample frame 233 (e) ABMST result for video 10
sample frame 400
 129

Table 4.22 Target tracking metrics for data set 10

CS FAR SS ACC DOST TR

T-MST 93.47 2.96 90.93 90.93 90.48 90.93
CAMST 94.12 1.12 91.33 91.67 91.67 91.44
ABMST 95.21 0.07 94.56 94.56 94.56 94.56

4.3.11 Plots for Target tracking techniques

Several performance metrics to evaluate the efficiency of the

proposed Adaptive background mean shift tracking with respect to
Traditional mean shift tracking and Continuously adaptive mean shift
tracking are considered. The metrics employed are frame based evaluations
and results of 10 video data sets are plotted in Figure 4.40 till Figure 4.45.

120

100

80
CS (%)

60 TMST
CAMST
40
ABMST
20

0
video 2

video 9
video 1

video 3

video 4

video 5

video 6

video 7

video 8

video 10

Figure 4.40 Completeness measure for target tracking techniques

 130

14

12

10
FAR (%)

8
TMST
6
CAMST
4 ABMST
2

Video 10
Video 6
Video 1

Video 2

Video 3

Video 4

Video 5

Video 7

Video 8

Figure 4.41 False alarm rates for target tracking techniques Video 9

120

100

80
SS (%)

60 CAMST
TMST
40
ABMST

20

0
Video 1

Video 2

Video 3

Video 4

Video 5

Video 6

Video 7

Video 8

Video 9

Video 10

Figure 4.42 Similarity measure for target tracking techniques

 131

120

100

80
ACC (%)

60 TMST
CAMST
40
ABMST
20

Video 10
Video 1

Video 2

Video 3

Video 4

Video 5

Video 6

Video 7

Video 8

Video 9
Figure 4.43 Accuracy measure for target tracking techniques

120

100

80
DOST ( %)

60 TMST
CAMST
40
ABMST
20

0
Video 10
Video 1

Video 2

Video 3

Video 4

Video 5

Video 6

Video 7

Video 8

Video 9

Figure 4.44 Duration of Successful tracking of target tracking techniques

 132

120

100

80
TR (%)

60 TMST
CAMST
40
ABMST
20

Video 10
Video 1

Video 2

Video 3

Video 4

Video 5

Video 6

Video 7

Video 8

Video 9
Figure 4.45 Tracking rate of target tracking techniques

In the above results, the performance metrics is higher for adaptive

background mean shift tracking technique than Traditional mean shift
tracking and Continuously adaptive mean shift tracking technique. The
efficiency of tracking in moving platform is higher than other algorithms.

The interpretations from the results are as follows

 The target tracking algorithm mainly depends on the

background model of the defined target.

 If the background is smooth then the tracking rate is higher

whereas it is lower for cluttered background.

 The illumination changes, partial intrusions have no effects on

the tracking algorithm

 The ABMST technique is robust against change in orientation,

size and altitude.