Fuzzy Integral for Moving Object Detection

1
Fuzzy Integral for Moving Object Detection

• Fida EL BAF, Thierry BOUWMANS, Bertrand VACHON
• Laboratory MIA, University of La Rochelle, France
• WCCI08 Fuzzy Image Processing
• 3 June 2008

2
Summary

• System overview
• Challenges
• Features
• Choquet Integral as sources fusion
• Fusing of "color " and "texture " features
• Foreground/Background classification based on
  Choquet integral
• Experiments - Evaluation

3
The objective

• In video sequences, the goal is to classify
  pixels of the current image as
  or

Foreground (FG)
Background (BG).
Pets 2006 dataset picture298 (720 x 576 pixels)
4
System overview of the different steps
Background Maintenance
t gt N
t N
tt1
N
Foreground Detection
Background Initialization
BG(t)
t N
Foreground Mask
Video Frames
I(t1)
N1
5
Challenges

• Critical situations generate a false
  classification of pixels

• Shadow -

Illumination Changes
Pets 2006 dataset picture298 (720 x 576 pixels)
6
Features

• Characterize the pictures element
• Features commonly used
• Color Edge Stereo Motion (optical flow)
  Texture
• Allow to handle changes caused by the
  illumination, the motion and the structure
  background.

7
Choice of features

• Color (3 components)
• Texture (Local Binary Pattern Heikkila PAMI
  2006)
• For each pixel, a degree of Similarity (S) is
  measured by computing the ratio of the BG and the
  Current Images values

8
How to compute S for Color and Texture?




0 T,C 255
Background Image
Current Image
0 S 1
For Texture
For Color
kone of the color components
9
Aggregation of Color and Texture features by
Choquet integral
BG(t)
I(t1)
Color features extraction
Texture Feature extraction
SC,1 SC,2 SC,3 ST
Color Similarity Measures
Texture similarity Measure
Choquet Integral
BG/FG Classification
Foreground Mask
10
Choquet Integral for fusing the features

• To increase the insensitivity to Illumination
  changes and Shadows
• For Color, 2 components are chosen
• For Texture, the feature obtained by the LBP code

0.6 0.5 0.5 0.5 0.53 0.3 0.4 0.3 0.39 0.34 0.1 0.1 0.2 0.11 0.13 0.9 0.9 0.8 0.89 0.87 0.7 0.6 0.7 0.61 0.66 0.4 0.5 0.5 0.5 0.47 1 1 1 1 1
FUZZY MEASURE VALUES
11
Classification BG/FG based on Choquet integral

• If then otherwise
• Where Th is a constant threshold
• is the Choquet integral value at pixel (x,y)

12
Experimental results with Ohta color space

• Aquatheque dataset (384 x 288 pixels)

Integral Color Space Choquet Ohta Sugeno Ohta
Quantitative Evaluation (Li measure Li2003) 0.44 0.27
a) Current Image
b) Groundtruth Image
Roc Curve Comparison of the 2 detection
algorithms
c) Segmented Image Choquet
d) Segmented Image Sugeno Zhang2006
13
Testing Other Color Spaces YCrCb, Ohta, HSV

• Aquatheque dataset (384 x 288 pixels)

Choquet - YCrCb
Choquet - HSV
Choquet - Ohta
Choquet Integral Color Space YCrCb Ohta HSV
Li Measure Li2003 0.46 0.44 0.34
Roc Curve Evaluation of the algorithm for
different color spaces
14
Experimental results for the datasets

• VS-Pets 2003 (720 x 576) and Pets 2006 (384 x 288
  pixels)

Original Image Choquet - YCrCb
Sugeno - Ohta
15
Conclusion

• Choquet Integral is suitable for the detection of
  moving objects
• YCrCb improves the insensitivity to illumination
  changes
• Method simple to implement
• Perspectives Learning the Fuzzy measure and
• testing other features

16
Thanks for your attention!
17
System overview of the different steps
Background Maintenance
t gt N
t N
tt1
N
Foreground Detection
Background Initialization
BG(t) I(t1)
t N
Foreground Mask
Video Frames
N1