Intelligent Systems Laboratory

beaverswimmingAI and Robotics

Nov 14, 2013 (3 years and 8 months ago)

52 views

Technion
-

Israel Institute of Technology

Computer Science Department



Intelligent Systems Laboratory












Students:


Alex Sherman



salex
she@t2.technion.ac.il





Valentin Rozental


svalentr@t2.technion.ac.il


Instructor
:

Igor Katzman


igork@cs.technion.ac.il



Supervisor:

Ehud Rivlin


ehudr@cs.technion.ac.il







Technion


2002






2



Vision / Robotics projects: 236754



Mobile Model Car Navigation Based on
Wireless Video Camera and Laser Sensors











Projec
t

VR
-
3:

Vehicle tracker: In this project we would like to develop
and implement a simple automatic tracking algorithm for a mobile model,
which would utilize an on
-
board wireless video camera to follow another
leading mobile model.













3





Content
s:


1 Introduction

………………………………………………
4


2 Problem presentation
…………………………………..
6


3
Ways to solution

………………………………………..
7


4
Experiments

and progress steps
……………………
11


5 Results
…………………………………………………...…
13


6
Conclusions
……………………………………………….
13


7 References
………………………
…………………………
14




















4





Chapter 1


Introduction





Mobile model navigation in unknown environment is an interesting and
challenging task. It consists of several computer
vision

based su
btasks, the
combination of which can provide the mobile model with the ability to avoid
obstacles, follow tracks, explore areas, etc. In the following project we would
like to develop and implement basic sets of functionalities, which together
will allow t
he mobile vehicle to behave smart.


In real life tracking is one of the most useful parts of human activity.

Tracking is everywhere. Almost all human activity in some way involve
s
tracking objects .For example w
ill take any different application areas:


S
pace exploring


1.

Stars interaction

2.

Clouds motion

3.

Space ships con
n
ection


Medical application:


1.

Remote

surgeon operation

2.

Patient

control


Military

application


1.

Weapons direction

2.

Target detection

3.

Aims classification

4.

Enemy identification into transport

5.

Missile

launching

6.

Guide a missile

7.

Nuclear experiments detection


5




Hostage situation & Security


1.

Terrorists detection (we must separate bad guys from good ones)

2.

Motion detection in guarded areas (secret or closed areas)

3.

Bank security (fast motion and unusual mov
ements of actors)

4.

Suspect tracking in public places

5.

Computer monitoring

6.

Prison observation (cells control)


Civil engineering



1.

Traffic control (junction, highways, traffic jams)

2.

Electronic surveillance

3.

Automatic driver

4.

Transport navigation


Science

1.

Lab
experiment analisys

2.

Video data analisys



Work inside human dangerous
environment


1.

Nuclear station

2.

Underwater works


Robotics

1.

Robots motion control

2.

Conveyer work & equipment verification


In all these activities a human mind takes responsibilities for tra
cking control
and it hides from us very complicated methods

that it uses. Developing simple

6

software for tracking objects is complicated but interesting challenge that is
even more interesting when it is become a game.





Chapter 2.

Problem presentation





The whole project is built from several parts. Here is an illustrating
picture of RC Car navigation:




server





frame











target



Turn Left









A module we developed is responsible for processing video input stream and
calculating a current coordinates of tracked
mobile and triggering a remote
module function

that is actually activate a RC Car and make it move.


7


Chapter 3
.

Ways to solution






Two of the basic methods for target tracking in real
-
time video
applications are temporal differencing (DT) and template correlation

matching. In the former appr
oach, video separated by a constant time DELTA
t are compared to find region which have changed. In the later approach each
video image is scanned for the region which best correlates
to an image
template. Independently, these methods have significant shor
tcomings.


DT tracking is impossible if there is significant camera motion, unless
an appropriate image stabilization algorithm is employed. It also fails if the
target becomes occluded or ceases its motion. In our case camera motion is
very significant, t
hus we didn’t use this method. Template correlation
matching generally requires that the object’s appearance remains constant.
The method is generally not robust to changes in object size, orientation or
even changing lighting conditions.

Nevertheless ther
e are solutions to these
obstacles and we have implemented them.


Template correlation

matching could

be done in several ways.

Here we
make a close look at some of them

(with their pros and
vs.
)
:


1.Countour



Advantages:



depends lightly on light conditi
ons.



good for the simple objects

Our
Module

Input
Video
Stream

Remote module
f
unction trigger


8


Disadvantages:



hard to detect if the noise level is high.



Complex forms in different visual angles.



A big variety of such forms.

2.Color


Advantages:



avoid the noise problem (for example with mean
filter)supposed to gi
ve good results.



Don’t depend on object form



Easier to detect the object location (center of mass)

Disadvantages:



depends heavily on light conditions.



The probability to misclassifications of the object is great,
since the color is a feature that is widel
y distributed.

3.Size


Advantages :



Easy to detect if the object is big.



Doesn’t depend on light conditions



Possibility to define the distance from two cams

Disadvantages:




Easy to misclassify the object if the object is small



The size of the object is
very sensitive to visual angle
(changes widely)



The size highly depends on the object distance, so its hard
to track the moving objects

4.Template:


Advantages:



Easy to scan the image (not necessary fast)



The matching could be done by partial match

(e
mpirically estimated) and if the match takes place the
classification is quite guaranteed.




Avoid the noise difficulties if they are not huge.

Disadvantages:


9



To detect object in motion a variety of templates must be
provided and (
sequent
/
parallel
) diagn
osed.



Not only
has a sight angled

complexes the tracking, but
also the movement forward/backward.


Taking in consideration the mentioned above pros and vs. we came to
conclusion that color matching is most
appropriate method

for us since real
-
time
performa
nce is

crucial.







When a new image becomes available, the tracker performs a local
search to find the best match by minimizing the normalized histogram
intersection between the color histogram of the searched area interior and a
previously stored, per
sonalized color histogram model. Histograms are almost
as flexible as memory
-
based methods but use a more compact representation.
Estimation of a histogram is also trivial. Color histograms for an image are
built from pixel values in one of color spaces. T
his method allows us to give a
different weight to different color ranges, thus increasing probability of quick
and precise target detection.


To make the tracking
even
more robust and efficient we extended color
matching method by
integration of
three

sel
f
-
learning features:

1.

Template updating

2.

Template history management

3.

Motion area prediction



Template updating



As mentioned above, biggest disadvantage of color template matching
is its great sensibility light conditions.
In our
i
mplementation, adaptiv
e
template updating is used to
e
nsure

that current template accurately represents
the new image of object. So the new template is generating by merging the
previous instance of template with current information

we have.


Template history management



10


Anot
her problem with adaptive template upd
ating where tracking itself
is performed

by color histogram recognition is when an object become
partially or fully occluded. In such case an adaptive template updating can
change a template to
something that is not a
tracked object but some obstacle
that occludes it. To overcome it we manage a template history by merging the
previous instance of template with current one.

For example:



R
(n) = a * M(n)

+ (1
-
a) *
R
(n
-
1)


where

R
(n
)




is a template at n
-
th stage

M(n
)




is a new motion detected

a





is a history parameter

Motion area prediction



Scanning a whole picture is computationally very expensive and
is

inapplicable to real
-
time applications, or
requires

specialized hardware to
operate in the real
-
time
domain. To overcome this obstacle we use
a motion
region prediction that reduces

search area.



The 2D image velocity vector of the target (u, v) (pixels/frame) can be
approximately determined by calculating the difference between centroid of
the previous
template
R
(n
-
1) and centroid of the new template
R
(n)
.









In
the area of centroid

exhaustive
search
is performed. In case
this
search fails the
n

area
is
enlarged

in such way that ensures object detection
.




11

















Chapter 4
.

Experiments and progress steps





N

Task

Duration

Notes

1

Background
acknowledgement

1 week

Done

2

Information gathering,
image processing
theoretical background
learning

1 week

Done

3

Studying existing tracking
algorit
hms and techniques

1 week

Done

4

Phase 1 document

1 week

Done

5

Tracking motion inside
tracked object issuing and
phase 2 document

2 week

Done

6

Color histogram and
tracked object centroid
detection studying

1 week

Done

7

Color histogram and
tracked o
bject centroid
detection implementation

2 week

Done


12

8

making test movies and
tracking algorithm
implementation

2 week

Done

9

Improvements analyses

3 week

Done


Remark:


First we have tried to classify the object according to the classification
method d
escribed by Lipton, Fujiyoshi and Patil, in

[1]
, which classifies a
vehicle by two parameters: Area and Perimeter. The exact formula is:


Dispersedness

=
Area
Perimeter
2
)
(

We tried this method on a large amount of different objects
as:



a.

different ca
r sizes

b.

different car forms

c.

different angle of sight


A majority of examples were misclassified. The reason for this mostly
is that a method doesn’t relate to the object size but only takes in the
consideration a fraction mentioned above. There is a logi
cal explanation why
object size shouldn’t be took in consideration: we don’t really know the size
of the tracked object since a car can be far from the camera


then it is a small
object ,and otherwise if a car is close to the camera


then it is a large
object.

Since this method didn’t show good results we decided to abandon it.

This
was another trigger for choosing color histogram object detection.













13













Chapter 5
.

Results





The
module

has been implemented on a Pentium 800 Mhz system under
Microsoft Windows 2000 in Matlab 6 ver 12
environment
. The module can
detect and track targets starting from 10 till 20 frame/seconds over a 320x240
pixel image. The module has been applied to l
arge amount of video films in
unstructured environments in which other activities is present.




Chapter 6
.

Conclusions





Using
template correlation matching

has three main advantages. It
allows continuous tracking despite occlusions and ces
s
ation
of tar
get motion,
it prevents template drifting onto background texture, and it provides robust
tracking. Target models are simple and based purely on target color so they
are applicable to a large number of real
-
word video applications.


Although a module show

good results it has several drawbacks.
Some of them
are:

1.

hard to detect if the noise level is high.


14

2.

The probability to misclassifications of the object is great, since the

color is a feature that is widely distributed.





Suggestions:


To over
come these obstacles gradient
-
based algorithm might be
implemented

as secondary stage recognition
.

This will
strongly
decrease an
probability for misclassification and “loosing” a target.





Chapter 7
.

References





[1] Alan J. Lipton, Hironobu Fujiyoshi and Raju S. Patil

"Moving Target Classification and Tracking from Real
-
time Video"


[
2
] T.Kanade, R. Collins, A.Lipton, P. Ananda
n, P
. Burt
“Cooperative

Multisensor Video Surveillance”

Proceeding of DAPRA Image
Understanding Workshop 1997, Vol.1, pp.3
-
10,1997.


[3] A.M. Baumberg and D.C. Hogg
.
An Efficient method for contour tracking
using active shape models.


[4] P. Fieguth a
nd D.Terzopoulos
. Color


based tracking of heads and other
mobile objects at video frame rates.


[5] M.Isard and A.
Blake

“Contour tracking by stochastic propagation of
conditional dencity”

Proceeding of European Conf. on Computer Vision 1996
pp. 343
-
356
, 1996



15

[6] D.Koller K.Danilidis H.
-
H. Nagel
“Model Based Object Tracking in
Monocular Image Sequences of Road Traffic Scenes”

International Journal of
Computer Vision; 10
-
3 pp. 257
-
281, 1993


[7]

Henry Schneiderman amd Takeo Kanade
“A Histogram based met
hod for
detection of faces and colors”.


[8] J. Davis and Bobick
.
“The representation and recofnition pf action using
Temporal Templates
.


MIT Media Lab Technical Report 402, 1997