a robust background subtraction and shadow detection
DESCRIPTION
A Robust Background Subtraction and Shadow Detection. Proc. ACCV'2000 , Taipei, Taiwan, January 2000. 井民全. Outline. Introduction Background Modeling Pixel Classification or Subtraction Operation Automatic threshold Selection Experimental result. Introduction. - PowerPoint PPT PresentationTRANSCRIPT
A Robust Background Subtraction and Shadow Detection
Proc. ACCV'2000 , Taipei, Taiwan, January 2000.
井民全
Outline
• Introduction
• Background Modeling• Pixel Classification or Subtraction Operation• Automatic threshold Selection
• Experimental result
Introduction
Extracting moving objects from a video sequences
• Application• What’s problem with before?• Requirements• The purpose
Current Image
Background images
Training
Background
Moving object
The purpose1. Static background2. Using color information3. New color model
Ei (expected color)
R
G
B
Cdi=color distortion
Ii ( current color)
αi Ei
Color model
α>1 <1 =1
2)()( iiii EI || Ei ||
iii ICD
Brightness distortion
Background Modeling(training)
-A pixel is modeled by a 4-tuple <Ei,si, i ,bi> Ei = arithmetic means rgb value over n frame si = standard deviation of rgb value over n frame i = variation of the brightness distortion bi = variation of the chromaticity distortion
N
CDCDRMSbi
NRMSai
i
N
ii
i
N
ii
2
0
2
0
)()(
)1()(
)](),(),([
)]( ),( ),([
iiiSi
iiiEi
BGR
BGR
N=background frames
ai
-normalized color bands in the brightness distortion and chromaticity distortion.
222
2B
2G
2R
,,
2
)()(
)()(
)()(
)(
)()(I
)(
)()(I
)(
)()(I
)(
)()(min
ii
ii
ii
i
ii
i
ii
i
ii
i
iiI
B
B
G
G
R
R
B
B
G
G
R
R
BGRC c
ciCi
BGRc c
ciCi i
iiICD
,,
2
)(
)()(
Pixel Classification or Subtraction Operation
• Original background (B): brightness and chromaticity similar to the trained background.• Shaded background or shadow(S): similar chromaticity but lower brightness.• Highlighted background(H): similar chromaticity but lower brightness.• Moving foreground object(F): chromaticity different from from the expected values in trained background.
Different pixels yield different distributions of illumination and chromaticity distortion.
Using single threshold, we must do normalization
i
ii a
1
distortion brightness theofvariation
distortion brightnessˆ
i
i
b
CDCD
distortion chromacity theofvariation
distortion chromacity
otherwise :H
else 0, ˆ :
else ,ˆ and ˆ : B
else ,CD :
)( 12
i
ii
CD
S
iF
iM
Ei (expected color)
R
G
B
What’s problem of the dark pixel ?
distortion brightness
normalized for the bound low
otherwise :H
else 0, ˆ :
else ,ˆ and ˆ : B
else ˆor CD :
)(
lo
22
lo
where
S
iF
iMi
ii
iCD
Automatic threshold Selection
Total sample=NXYN=background frames
i
Freq.
Normalized Ahpha value0 +-
Fig. The normalized brightness distortion histogram
•The thresholds are selected according to the desired detection rate r
2 1
Automatic threshold Selection
Normalized CD iCD
Freq.
0 +
Fig. The normalized chromaticity distortion histogram
•The thresholds are selected according to the desired detection rate r
CD
Experimental result
Images size= 360 x 240Detection rate= 0.9999Lower bound of the normalized brightness distortion = 0.4
Fig. A sequence of an outdoor scene contain a person walking across the street
Fig. An application of the background subtraction in a motion capture system
Fig. game
Fig. An application of background subtraction in video editing
Conclusion
•Presented a background subtraction algorithm•Accurate, robust, reliable and efficiently computed•Real-time applications