1
Lucia Maddalena and Alfredo Petrosino, Senior Member, IEEE
A Self-Organizing Approach to Background Subtraction for Visual Surveillance Applications
A Self-Organizing Approach to Background Subtraction for Visual Surveillance Applications
Adviser : Chih-Hung Lin Speaker : Kuan-Ju ChenDate : 2009/04/06
2
AuthorLucia Maddalena
received the Laurea degree (cum laude) in mathematics and the Ph.D. degree in applied mathematics and computer science from the University of Naples Federico II, Naples, Italy.
Alfredo Petrosino (SM’02) is an Associate Professor of
computer science at the University of Naples Parthenope, Naples, Italy.
3
OUTLINE
INTRODUCTION1
METHOD2
EXPERIMENTAL RESULTS3
CONCLUSION4
4
1.INTRODUCTION
VISUAL surveillance is a very active research area in computer vision
The main tasks in visual surveillance systems motion detection object classification Tracking activity understanding semantic description
5
1.INTRODUCTIONThe usual approach to moving object
detection is through background subtraction
Compared to other approaches, The main problem is its sensitivity to dynamic scene changes light changes moving background cast shadows
6
1.INTRODUCTIONBackground subtraction:
Unimodal versus multimodal: Recursive: Pixel-based :
7
1.INTRODUCTIONUnimodal and multimodal:
Basic background models assume that the intensity values of a pixel can be modeled
• low complexity• cannot handle moving backgrounds
8
1.INTRODUCTIONRecursive
recursively update a single background model based on each input frame.
• Space complexity is lower• Background model is carried out for a long time
period
9
1.INTRODUCTIONPixel-based :
assume that the time series of observations is independent at each pixel
10
1.INTRODUCTIONOur approach is based on the
background model automatically generated by a self-organizing method and can be broadly classified as multimodal,
recursive, and pixelbased.
11
2.METHOD
Initial Background Model1
Subtraction and Update of the Background Model2
12
2.1 Initial Background Model
a b c
d e f
a1 a2 a3
a4 a5 a6
a7 a8 a9
b1 b2 b3
b4 b5 b6
b7 b8 b9
c1 c2 c3
c4 c5 c6
c7 c8 c9
d1 d2 d3
d4 d5 d6
d7 d8 d9
e1 e2 e3
e4 e5 e6
e7 e8 e9
f1 f2 f3
f4 f5 f6
f7 f8 f9
Let be HSV components , ex: a1=(h,s,v)
13
2.2 Subtraction of the Background Model
Input pixel tp
Build pixel model C(2.1)
Find best Match Cm
Cm is found
B(Pt)=0
NO
Pt is shadow
YES
B(Pt)=1
YES
NO
LastFrameNot LastFrame
Current Frame+1
Over
LastFrame
Self-Organizing Background Subtraction
Update Background
Model
14
2.2 Subtraction of the Background Model
Use Euclidean distance to computeCn and Cother pixel distance
),(min),(
sinsincoscos),(
2,,1
222
tini
tm
jijjjiiijjjiiiji
pcdpcd
vvhsvhsvhsvhsvppd
15
2.2 Subtraction of the Background Model
16
2.2 Update of the Background Model
If best match cm Weight vector At to update in the
neighborhood cm
If best match cm isn`t found Not update
17
2.2 Update of the Background Model
a b c
d e f
a1 a2 a3
a4 a5 a6
a7 a8 a9
b1 b2 b3
b4 b5 b6
b7 b8 b9
c1 c2 c3
c4 c5 c6
c7 c8 c9
d1 d2 d3
d4 d5 d6
d7 d8 d9
e1 e2 e3
e4 e5 e6
e7 e8 e9
f1 f2 f3
f4 f5 f6
f7 f8 f9If best match cm
Computer the weight vector to update background
18
2.2 Update of the Background Model
a1 a2 a3
a4 a5 a6
a7 a8 a9
b1 b2 b3
b4 b5 b6
b7 b8 b9
c1 c2 c3
c4 c5 c6
c7 c8 c9
d1 d2 d3
d4 d5 d6
d7 d8 d9
e1 e2 e3
e4 e5 e6
e7 e8 e9
f1 f2 f3
f4 f5 f6
f7 f8 f9
),(),())(1(),( ,1, yxpjiAtjiA tjitjit
19
SHADOW DETECTION
Foreground
Ayalyze Hue-Saturation-Value(HSV) color space
shadow mask:
otherwise 0
) ( )( )( 1SP s H
Hi
Ht
Si
StV
i
Vt
t cpcpc
pif
p
define a darkening effect of shadowsidentifying as shadows those points
average image luminance
Following three condition to mask shadow
20
3.EXPERIMENTAL RESULTS
(a) original frame;(b) computed moving object detection mask(c) background model(d) background model change mask from previous frame
21
3.EXPERIMENTAL RESULTS
(a) original frame;(b) computed moving object detection mask
22
3.EXPERIMENTAL RESULTS
fpfntp
tp
F
fptp
tp
fntp
tp
Similarity
PrecisionRecall
PrecisionRecall2
Precision
Recall
1
23
3.EXPERIMENTAL RESULTS
(a) test image (b) ground truth(c) SOBS result (d) Pfinder result(e) VSAM result (f) CB result
24
4.CONCLUSIONThis paper also includes a
comprehensive accuracy testing, performed with both pixel-based and frame-based metrics Experimental results, using different sets of
data and comparing different methods, have demonstrated the effectiveness of the proposed approach
• illumination changes• cast shadows
ONGOING WORK improve detection results
25
www.themegallery.com