16
Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceed ings of the 17th International Conference on

Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

  • View
    223

  • Download
    1

Embed Size (px)

Citation preview

Page 1: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Improved Adaptive Gaussian Mixture Model for Background

Zoran Zivkovic

Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference on

Page 2: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Outline

Introduction Gaussian Mixture Model Select the number of components Experiments Conclusion

Page 3: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Introduction Background subtraction is the common

process for surveillance system

Gaussian mixture model (GMM) was proposed for background subtraction Like Gaussian Dist-s model

These GMM-s use a fixed number of components

Page 4: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Gaussian Mixture Model

are the estimate of the means are the estimate of the variance are mixing weight (non-negative an

d add up to one)

Page 5: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Gaussian Mixture Model

Update equation a

a

a

Page 6: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Gaussian Mixture Model

If the current pixel didn’t match with any distributions s

Decide pixel is in background/foreground d

sd

Page 7: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Select the number of components

Goal choose the proper number of component

Implement Use prior and likelihood to select

proper models for given data

Page 8: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Select the nmber of components Maximum Likelihood (ML)

Likelihood function:

Assume we have t data samples

a

Page 9: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Select the number of components

Maximum Likelihood (ML) a

a

Constraint: weights sum up to one

The prior update func.

Page 10: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Select the number of components

Dirichlet prior a presents the prior evidence for the cla

ss m – the number of samples that belong to that class a priori

Use negative coefficients means that accept class-m exist only if there i

s enough evidence from the data for the existence of this class

Cm

Page 11: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Select the number of components

Maximum Likelihood (ML) +Dirichlet prior a

a

Fixed Expect a few components M and is small

a

Page 12: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Experiments

New GMM with slight improvement

Page 13: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Experiments

Max 4 Dist.

1 Dist.

Page 14: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Experiments

In highly dynamic ‘tree’, the processing time is almost the same

Page 15: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

Conclusion

Present an improved GMM background subtraction scheme

The new algorithm can select the needed number of component

Page 16: Improved Adaptive Gaussian Mixture Model for Background Zoran Zivkovic Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference

The method of Stauffer and Grimson

is the learning rate that is defined by usesr