MOVING OBJECTS SEGMENTATION USING BOUNDARY

LINKING AND BACKGROUND

Jun Ki Kim, Ho Suk Lee

Multimedia Laboratory, Dept. of Computer Engineering,

Hoseo University, Asan, Chungnam, South Korea

Keywords: Moving object edge, Moving object boundary, Boundary linking, Background, Moving object segmentation

Abstract: Moving object boundary is very important for moving object segmentation. We extract the moving object

boundary from the moving object edge. But the object boundary shows broken boundary and we use a

boundary linking to link the broken boundary. The boundary linking algorithm forms a quadrant around the

terminating pixel in the broken boundary and searches forward other terminating pixel to link within a

radius. The linking algorithm guarantees shortest distance linking. We register the background from image

sequence. We construct two object masks, one from boundary linking and the other from the background,

and use these two complementary object masks for moving object segmentation. We also filter out the

moving cast shadow using gradient operator. The major characteristics of the proposed algorithms are

accurate moving object segmentation, multiple moving objects segmentation, and the segmentation of an

object which has holes in its region using these two object masks. We experiment the algorithms using the

standard MPEG-4 test sequences and real video sequence. The proposed algorithms are very efficient and

can process QCIF image more than 48 fps and CIF image more than 19 fps in a 2GHz Pentium-IV computer.

1 INTRODUCTION

The MPEG-4 video coding standard introduces the

concept of Video Object Plane(VOP) to support

content-based applications. The moving object

segmentation has become an important research

subject and has been receiving considerable amount

of attention from many researchers for successful

MPEG-4 content-based applications

(Kim,2002)(Meier,1999)(Chien,2002)(Chien,2003).

One of the most important elements for moving

object segmentation is to obtain moving object

boundary. The object boundary, however, shows

broken boundary caused by the edge detection failure

due to instantaneous halt of the moving object. The

object or the part of object often halts temporarily

between consecutive frames and this causes the edge

detection failure. Illumination variation and camera

noise can also cause edge detection failure. The

broken boundary makes it difficult to obtain an

accurate moving object segmentation. The broken

boundary should be supplemented with boundaries

from other boundary sources and boundary linking

and further using object mask. The other boundary

sources include previous moving object edge, Canny

edge of current frame, and background edge of

current frame. The boundary linking is defined as the

linking of two terminating pixels in the broken

boundary. The boundary linking algorithm forms a

quadrant around the terminating pixel and searches

forward other terminating pixel to link in a clockwise

concentric circle within a radius. The algorithm does

not create a cycle and guarantees the shortest

distance linking. The linking algorithm tries to link

the broken boundary robustly within a search radius.

But the linking algorithm does not try to link the

broken boundary which is wider than the search

radius, so for some images, there might remain

broken boundaries after boundary linking. But for

most images, we can obtain a completely linked

moving object boundary after boundary linking. We

use the object mask to deal with the case of wide

broken boundary.

The main characteristics of the proposed

algorithm are accurate moving object segmentation

by boundary linking, the segmentation of an object

which has holes in its region, and real-time

performance. We use the standard MPEG-4 test

sequences and a real video sequence for experiment.

We assume that the camera and background are

stationary.

The organization of paper is as follows. Section 2

269

Suk Lee H. and Ki Kim J. (2004).

MOVING OBJECTS SEGMENTATION USING BOUNDARY.

In Proceedings of the First International Conference on E-Business and Telecommunication Networks, pages 269-277

DOI: 10.5220/0001401302690277

 SciTePress

gives the detailed introduction of the proposed

moving object segmentation algorithms. Section 3

shows the experimental results of the algorithm and

section 4 draws the conclusion.

2 MOVING OBJECT

SEGMENTATION

2.1 Overall architecture

We consider the moving object segmentation as a

process of development from initial moving object

to final moving object segmentation. The initial

moving object is actually the background change

detection mask. We use the background change

detection mask computed from the current frame and

the registered background because it can yield more

accurate moving object edge. The initial moving

object is processed in two complementary ways. The

two complementary ways are moving object

boundary construction from moving object edge and

boundary linking and noise elimination of initial

moving object for object segmentation. The

constructed object boundary, however, shows broken

boundary so we try to link the broken boundary to

obtain a completely linked boundary for most

images. And finally we combine the results from two

complementary object masks for accurate moving

object segmentation and the segmentation of an

object which has holes in its region. We use the

connected component algorithm and morphological

operation to eliminate the noise and to enhance the

object mask boundary. Fig. 1 shows the overall

system architecture

Figure 1: Overall architecture of moving object

segmentation

FD means the pixel-based frame difference between

the current frame and previous frame and

generates the current frame change detection mask.

BD means the frame difference between the current

frame and background and generates the background

change detection mask. BB means the background

buffer which is used to register the background. The

count value of FD

frame difference is sent to BB to

constructed using the Canny edge of current frame

and the initial moving object. We scan the

constructed moving object edge horizontally and

vertically and extract the moving object boundary.

Then, we examine the object boundary and set the

terminating pixels in the broken boundary regions

and perform the space-oriented boundary linking

robustly to obtain a completely linked moving object

boundary. We construct the object mask of first type

using this result. We also eliminate noise from the

initial moving object and construct the object mask

of second type. Then, we extract the VOP of the

current frame using these two object masks.

2.2 Moving object edge construction

We construct the moving object edge as a process

from initial moving object edge through

intermediate and to final moving object edge. Fig. 2

shows the block diagram of the moving object edge

construction.

Figure 2: Moving object edge construction

The initial moving object edge is constructed

using Canny edge (Canny, 1986) of current frame

and the initial moving object. The intermediate

moving object edge is constructed using Canny edge

of current frame and the previous moving object

edge to link the broken boundary region which is

mostly caused by instantaneous halt of the object

between the consecutive frames. After this process,

the moving object edge construction is divided into

two modes. In mode one, we regard the intermediate

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

270

MOEyxEyxE −= ),(),(

},|&{

2121

IMOeEeeeeMOE

initial

∈∈==

},|&{

122121 −

∈∈∈=

MOEeMOEeEeeee

MOE

final

MOE

MOEeeee ∈=

121

|||{

EeEe ∉∈

moving object edge as the final result. We can

process the video sequences such as “Weather” and

“Hall monitor” in mode one and can obtain a

sufficient moving object edge for further processing.

In mode two, we construct the final moving object

edge using edge tracing. The edge tracing is

implemented using Canny edge of current frame and

background edge of current frame. We can process

the “Claire”

which does not have background and

the “Akiyo” which has a high

quality in mode two

using edge tracing. The user can select the mode

using a parameter

2.2.1 Change detection and background

We use the absolute pixel difference as the test

statistics for change detection. Under the

hypothesis that there is no change in the current

pixel, the pixel difference can be modeled using a

zero-mean Gaussian distribution N(0,

) with

variance

(Aach,1993).

We register the background using stationary

background filtering (Meier,1999). We use the

background to generate the background change

detection mask and it is used as the initial moving

object.

2.2.2 Moving object edge

We define the edge as the pixels constituting the

edge in the image. The following is the equation for

initial moving object edge computation.

(1)

Here,

E is the Canny edge of the current frame

and IMO is the initial moving object. e, e

, and e

are

the edges. The operation & is the logical AND

operation.

The intermediate moving object edge is as follows.

intermediate

(2)

Here, MOE

is the current moving object edge and

MOE

n-1

is the previous moving object edge

The ||

operation means the logical OR operation. It means

the union of the edges of MOE

with the edges of

MOE

n-1

to link the broken boundaries in MOE

caused by the instantaneous halt of the object. The

final moving object edge is computed as follows.

intermediate

(3)

Here, E

is the background edge. The logical OR

operation means the union of the edges of Canny

edge of current frame excluding background edges

with the intermediate moving object edges in order

to obtain the more linked final moving object edge.

We call this edge tracing. We find edge tracing quite

effective when the background has no noise or when

the image has a high quality such as the “Akiyo” or

the “Claire”. We can see the edge tracing results in

Fig. 4 (c1) through (c4)

2.2.3 Edge tracing

Edge tracing is performed in mode two of moving

object edge construction. Edge tracing uses Canny

edge of current frame E

and the background edge

of current frame E

. The followings are edge

tracing steps.

(step1) Construct the background edge E

(x,y)

using the following equation automatically.

intermedia

(4)

But the background edge contains some still

remaining uneliminated object pixels as shown Fig.

3 (a). We use the connected component algorithm to

eliminate these remaining object pixels in the

background edge and obtain a clear background

edge as shown in Fig. 3 (b).

(step2) Scan MOE

intermediate

(x, y) by horizontal

scanline and locate the first pixel encountered.

(step3) Search around the neighboring pixels of the

pixel found in (step2) in 8-connected way to start

the edge tracing in MOE

intermediate

(step4) Check the presence of a pixel in E

(x,y) at

the corresponding coordinate position during edge

tracing in MOE

intermediate

until the edge tracing

comes to the terminating pixel in MOE

intermediate

(x,y).

If there exists a pixel at the corresponding

coordinate, perform edge tracing along the E

(x,y) edge as long as the pixel of E

(x,y)

under edge tracing does not encounter the

background edge E

(x,y) constructed in (step1).

And include the traced edge of E

(x,y) into

MOE

final

. If it encounters the background edge,

edge tracing stops.

(step5) Return to (step2) and continue.

MOVING OBJECTS SEGMENTATION USING BOUNDARY

271

(a)

(b)

Figure 3: Background edge (Akiyo 11

)

(a) Background edge with remaining object pixels,

(b) Background edge with no object pixels

Figure 4 shows the moving object edge construction

results.

(a1) MOE

intermediate

(a2) MOE

intermediate

(a) Hall monitor 45

(b1) MOE

intermediate

(b2) MOE

intermediate

(b) Akiyo 29

(c1) MOE

initial

(c2) MOE

intermediate

(c3) MOE

final

(c4) MOE

final

Figure 4: Moving object edge

(a) Hall monitor, (b) Akiyo, (c) Claire

Fig. 4 shows the moving object edge of (a) “Hall

monitor”, (b) “Akiyo”, and (c) “Claire”. The left

column of (a) and (b) were produced using the

current frame change detection mask as in

(Kim,2002) and the right column of (a) and (b) using

the background change detection mask. We can see

the more accurate moving object edge in the right

column. For (a) and (b), we use the intermediate

moving object edge as the final result. The (c)

“Claire” shows the process of moving object edge

construction using edge tracing. Fig. 4 (c4) was

shown for more demonstration.

2.3 Moving object boundary

2.3.1 Moving object boundary extraction

First, we eliminate small isolated noise pixels from

the moving object edge. Then, we scan the moving

object edge by horizontal and vertical scanline and

obtain the moving object boundary. The following

Fig. 5 shows the moving object boundary of “Claire”.

(a) (b)

Figure 5: Moving object boundary

(a) Moving object edge, (b) Moving object boundary

2.3.2 Moving object boundary linking

We examine the moving object boundary image

and set terminating pixels at the final end pixels of

the broken boundary line. Fig. 6 (a1) and (a2) show

the “Hall monitor” with terminating pixels, (b)

shows the “Akiyo” with terminating pixels, and (c)

shows the “Claire” with terminating pixels. The

actual terminating pixel is just a single pixel but we

enlarge it for clear vision.

(a1) Hall monitor 54

(a2) Hall monitor 69

(a) Hall monitor

(b) Akiyo 18

Figure 6: Terminating pixel setting

(a) Hall monitor, (b) Akiyo, (c) Claire

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

272

180/*cos

−

)

θπθ

θπ

++≤≤−+

vsv

Various edge linking algorithms considering the

pixel gradient magnitude are explained using

examples (Gonzalez,2002). An edge linking

algorithm considering three types of break points

using window is explained in (Hajjar,1999). The

proposed boundary linking algorithm is a space-

oriented geometric algorithm. The boundary linking

algorithm links the broken

boundary by inserting

pixels from

one terminating pixel to the other

terminating pixel. The algorithm considers the

linking direction first. The algorithm forms a

quadrant around the terminating pixel and searches

forward other terminating pixel to link within a

radius in concentric circle. The algorithm always

guarantees shortest distance linking and does not

create a cycle. We see that the boundary linking

algorithm forms the object boundary in the broken

boundary region accurately because the terminating

pixels used for boundary linking in the broken

boundary region exist on the object boundary.

We give a more detailed explanation. We assume a

virtual pixel around the terminating pixel. The

virtual pixel is assumed in the forward direction of

the terminating pixel by considering the slope of the

terminating pixel. The purpose of virtual pixel is to

predetermine the general direction toward which the

broken boundary is going to be linked from the

terminating pixel.

However, it is usually seldom the case that every

terminating pixel ought to be linked only in the

direction of the virtual pixel, because the pixel

distribution varies significantly according to the

characteristics of the input image. Thus, we form a

quadrant for a more general search range around the

terminating pixel

)0,0(

=T in the direction of

the virtual pixel

),(

VVV = . The algorithm

begins by positioning the terminating pixel at the

coordinate origin. The boundary linking operation is

now carried out within the quadrant from the

terminating pixel. First, we compute the angle

between the terminating pixel and the virtual pixel.

If we let

as the angle between

T and

we compute

(5)

and we search another terminating pixel

T to link

around the terminating pixel

T within the

quadrant in clockwise concentric circle within a

given radius. Now, if we find another terminating

pixel

),(

TTT = around the current

terminating pixel

T . We compute the angle

between

T and

T .

(6)

Then the condition and operation of boundary

linking can be summarized as follows.

procedure boundary_linking( )

{

then {

link

)0,0(

T to ),(

TTT =

} else expand the quadrant to half plane

and perform boundary linking from

)0,0(

}

If the angle

of the terminating pixel

T is

within the range as computed above, the boundary

linking is carried out from the terminating pixel

to the terminating pixel

T . If the angle

is not

within the range, we expand the quadrant to half

plane and perform the boundary linking iteratively

from

)0,0(

T .

We find other terminating pixel to link within the

radius around the starting terminating pixel within 5

iterations most of the time in the experiment. In

other cases, the boundary linking algorithm searches

the pixel in the expanded half plane. We continue

boundary linking until there remain no more

terminating pixels to link. Fig. 7 shows the

constructed moving object boundary of Fig. 6 after

boundary linking.

The linking algorithm tries to link the broken

boundary robustly within a search radius which is

determined experimentally. The linking algorithm

does not try to link the terminating pixel which is

located farther than the search radius, because it can

link the wrong terminating pixel in trying to do so.

For example, the search radius for “Hall monitor” is

set at 8 and “Akiyo” at 12.

The algorithm sets 20 terminating pixels in Fig. 6

(a1) and performed 11 boundary linking in Fig. 7

(a1) “Hall monitor”. In Fig. 6 (b) “Akiyo” 18

, the

algorithm sets 39 terminating pixels and performed

18 boundary linking. We can see several broken

boundaries in Fig. 6 (a1), (a2), (b), and (c) and can

see their corresponding completely linked object

boundary in Fig. 7 (a1), (a2), (b), and (c)

respectively. Currently, we can see two object

boundaries in some boundary, one by moving object

edge itself, and the other by boundary linking, but

these seemingly double boundaries do not cause a

problem, because we map the object texture to the

outermost boundary.

180/*cos

−

MOVING OBJECTS SEGMENTATION USING BOUNDARY

273

(a1) Hall monitor 54

(a2) Hall monitor 69

(a) Hall monitor

(b) Akiyo 18

Figure 7: Moving object boundary linking

(a) Hall monitor, (b) Akiyo, (c) Claire

2.3.3 Moving object linking at the bottom

We construct the moving object boundary as shown

in Fig. 7. But the moving object shows boundary

disconnection at the bottom. Fig. 8 (a) and (b) show

the connected moving object boundary of Fig. 7 (b)

and (c).

(a) Akiyo 18

(b) Claire 11

Figure 8: Object boundary connection at the bottom

(a) Akiyo, (b) Claire

2.4 Video object plane extraction

We perform moving object extraction with two

object masks. We use the object mask made from

boundary linking and the object mask from initial

moving object. We call the former the object mask

of first type and the latter the object mask of second

type. The advantage of the first type is the accurate

moving object boundary. But it cannot handle the

case of an unlinked broken boundary which could

exist in the wide broken boundary and the

segmentation of an object which has holes in its

region. The second type can handle these situations

but the boundary it shows is somewhat coarse.

Therefore, we use these two object masks in

combination.

2.4.1 Using the object mask of first type

Fig. 9 shows moving object extraction using the

object mask of first type. Fig. 9 (a) is the moving

object mask and (b) is the object segmentation result.

Fig. 9 (b) shows an accurate object boundary

(a) (b)

Figure 9: Object extraction with first mask (Akiyo)

(a) Moving object mask, (b) Segmentation result

2.4.2 Using the object mask of second type

Fig. 10 shows the block diagram of the second type

object mask construction.

Figure 10: Construction of object mask of second type

We show in Fig. 11 the moving object extraction

using the object mask of second type. Fig. 11 (a)

shows the initial moving object with noise, (b)

shows the object mask of second type with clear

background and object images, and (c) shows the

object segmentation result.

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

274

⎪

⎩

⎪

⎨

⎧

),(_

yxframeCurrent

yxobjectMoving

otherwise

yxondmask

yxfirstmaskif

1),(sec_&

1),(_

%100

)),(),((

_ ×

⊕

∑

yxOMyxa

PercentageError

(a) (b)

(c)

Figure 11: Moving object extraction 45

(a) Initial moving object, (b) Object mask of second type,

Fig. 12 shows moving object extraction result of

“Akiyo” using the object mask of second type. The

“Akiyo” object shows somewhat coarse object

boundary.

(a) (b)

Fig. 12 Object extraction with second mask (Akiyo)

(a) Moving object mask, (b) Segmentation result

2.4.3 Using both object masks

Following is an expression for moving object

extraction using both object masks.

(7)

Moving_object

(x,y) means the extracted moving

object. The mask_first

(x,y) means the object mask

of first type and mask_second

(x,y) the object mask

of second type. The & is the logical AND operation.

The moving object is extracted from the current

frame if and only if both object masks are equal to 1.

Thus we can handle wide broken boundaries and can

segment the object which has holes using the object

mask of second type. Fig. 13 (a) shows

“Mother&Daughter” which has a wide broken

boundary and (b) shows the segmentation result

using expression (7). Fig. 13 (c) and (d) show the

segmentation of “Junki” real video which has holes

in its region using expression (7), too.

(a) (b)

Figure 13: Moving object extraction with both masks

(a) Mother&Daughter with a wide broken boundary,

(b) Mother&Daughter segmentation result,

2.4.4 Moving cast shadow filtering

The moving cast shadow is composed of a dark

region called umbra and a soft transition region

called penumbra (Stauder,1999). We apply the

Roberts gradient operator to filter out the moving

cast shadow. This shadow filtering operation can be

executed to the input image.

3 EXPERIMENT

We experiment the moving object segmentation

algorithm using the MPEG-4 standard test image

sequences such as “Hall monitor”, “Claire”,

“Akiyo”, “Weather” and “Junki” real video.

3.1 Objective evaluation

We use the error percentage measure (Chien,2002)

for the objective evaluation. The error percentage

measures the cumulative error between the original

image and the segmented image.

(8)

MOVING OBJECTS SEGMENTATION USING BOUNDARY

275

The a(x, y) is the alpha map of the original image

used for reference.

⊕

is the Exclusive-OR

operation. The OM(x, y) is the alpha map of the

segmented image. W is the width and H is the height.

Fig. 14 shows the error percentage graph for

“Claire”, “Akiyo”, and “Weather”.

Figure 14: Error percentage graph

The size of image sequence is CIF(352x288) and

it is input 25 frames/sec. We obtained the a(x, y)

reference image sequence by segmenting it manually

using the PHOTOSHOP tool in order to obtain the

accurate reference image. Also, we obtained the

OM(x, y) by applying the proposed object

segmentation algorithm.

If there is no difference, the error percentage is

0%. If the error percentage is above 1.5%, then the

segmented moving object could show some visually

recognizable difference from the input moving

object.

The error percentage values we obtained from our

experiments are less than 0.78% for these image

sequences. This value indicates that the segmented

moving object by our algorithm is visually very

similar to the manually segmented moving object

frames from the input image sequence and

validates that the proposed segmentation algorithm

can segment the objects very accurately. In Fig. 14,

frames 3 and 4 of “Akiyo” and “Weather” show

some high value of error percentage and frame 49,

79, and 94 of “Claire” show high value of error

percentage because the object moves somewhat

faster in these frames.

We can compare Fig. 14 with Fig. 7 of

(Chien,2002) and can see that the error percentage

results of “Akiyo” and “Weather” of our approach

are comparable with those results.

Table 1 shows the performance statistics of our

algorithm and other approaches for comparison. We

use the P-IV CPU with 2.0GHz as the experiment

CPU. We have optimized the implementation of

computation intensive algorithms to satisfy the real-

time requirement of various multimedia applications.

Table 1: Performance statistics

Table 1 shows the performance statistics. The

table shows the test sequences, total frames, total

processing time, and frames per second. The test

sequences are CIF(352x288) and QCIF(176x144)

standard MPEG-4 test sequences such as “Hall

monitor”, “Akiyo”, “Claire”, “Mother&Daughter”,

and “Weather”. For

example, the proposed system

can process 72.25 frames per second for “Hall

monitor” QCIF sequence, in other words, the system

spends 0.0138 second per frame for “Hall monitor”

QCIF sequence. We have included the performance

results of (Kim,2002),(Chien,2002), and

(Chien,2003) in the table for comparison. The

approach (Chien,2002) can process QCIF at 25 fps

and CIF at 10 fps using P-III 450MHz CPU. The

approach (Chien,2003) can process QCIF at 9.4 fps

using P-III 800MHz. The test CPUs are different but

we can conclude that the proposed algorithms are

more efficient than those algorithms.

3.2 Subjective evaluation

We show the moving object segmentation results of

“Hall monitor”, “Weather(360x243)” in Fig. 15. We

also use “Junki” real video sequence to demonstrate

the segmentation of an object which has holes in its

region.

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

276

(a1) Hall monitor 45

(a2) Hall monitor 152

and 219

(a) Hall monitor

(b)Weather 217

Figure 15: Moving object segmentation

(a) Hall monitor, (b) Weather, (c) Junki real video

We see that our algorithm can produce the

accurate moving object segmentation and can

segment the multiple moving objects and the object

which has holes in its region using these two

complementary object masks. We can see a small

hole in the upper part of Fig. 15 (b) between the hair

and forehead and also can see two holes in (c)

clearly.

We can compare these segmentation results with

the results in (Kim,2002),(Meier,1999),(Chien,2002),

and (Chien,2003). Our segmentation results are

better than those results in terms of the accuracy of

the segmentation result because we use the object

boundary linking for accuracy. The proposed

segmentation algorithm can segment an object

which has holes in its region as shown in (b) and (c)

in Fig. 15. But none of references demonstrate this.

Particularly, we can see in (d) and (e) of Fig. 9 of

(Chien,2002) that the approach does not segment the

object which has a hole in its region. We can also

compare our “Mother&Daughter” segmentation

results of Fig. 13 (b) with those of Fig. 16 and Fig.

17 of (Chien,2003) and we can see that our results

are better.

4 CONCLUSION

We propose a novel moving object edge construction

algorithm, a space-oriented geometric boundary

linking algorithm, and a segmentation algorithm

using two object masks. We can achieve more

accurate moving object segmentation, multiple

moving objects segmentation, and the segmentation

of an object which has holes in its region using these

algorithms.

We have shown the performance of our proposed

algorithms using standard MPEG-4 test image

sequence and a real video from camera. The

proposed algorithms are very efficient and can

process QCIF image more than 48 fps and CIF image

more than 19 fps in a personal computer for real-time

content-based applications.

REFERENCES

Kim, Changick, Hwang, Jenq-Neng, “Fast and Automatic

Video Object Segmentation and Tracking for Content-

Based Applications,” IEEE Trans. on Circuits and

Systems for Video Technology, Vol. 12, No. 2, pp.122-

129, Feb. 2002.

Meier, Thomas, Ngan, King N., “Video Segmentation for

Content-Based Coding,” IEEE Trans. on Circuits and

Systems for Video Technology, Vol. 9, No. 8, pp.1190-

1203, Dec. 1999.

Chien, Shao-Yi, Ma, Shyh-Yih, Chen, Liang-Gee,

“Efficient Moving Object Segmentation Algorithm

Using Background Registration Technique,” IEEE

Trans. on Circuits and Systems for Video Technology,

Vol. 12, No. 7, pp.577-586, July 2002.

Chien, Shao-Yi, Huang, Yu-Wen, Chen, Liang-Gee,

“Predictive Watershed : A Fast Watershed Algorithm

for Video Segmentation,” IEEE Trans. on Circuits and

Systems for Video Technology, Vol. 13, No. 5, pp.453-

461, May 2003.

Canny, John, “A computational approach to edge

detection,” IEEE Trans. on Pattern Analysis and

Machine Intelligence, Vol. 8, No. 6, pp. 679-698, Nov.

1986.

Aach, T., Kaup, A., Mester, R., “Statistical model-

based change detection in moving video,” Signal

Processing, Vol. 31, pp. 165-180, March 1993.

Stauder, Jürgen, Mech, Roland, Ostermann, Jörn,

“Detection of Moving Cast Shadows,” IEEE Trans. on

Multimedia, Vol. 1, No. 1, pp. 65-76, March 1999.

Gonzalez, Rafael C., Woods Richard E., Digital Image

Processing, Prentice-Hall, Inc., 2002.

Hajjar, Amjad, Chen, Tom, “A VLSI Architecture for

Real-Time Edge Linking,” IEEE Trans. on Pattern

Analysis and Machine Intelligence, Vol. 21, No. 1, pp.

89-94, Jan. 1999.

MOVING OBJECTS SEGMENTATION USING BOUNDARY

277