This document discusses background subtraction methods for comparing current techniques. It provides an overview of common parametric and non-parametric methods from literature. It then describes specific methods implemented at SIG, including single Gaussian, mixture of Gaussians, and a hybrid approach. The document outlines the methodology used to generate ROC curves for evaluating methods on several datasets, including results. It concludes that choosing improvements on the background model affects performance more than the initial model choice.

1. Background Subtraction
Validation

2. Overview
●Problem Statement
●Background Subtraction Methods
●Methodology in creating ROC curve
●Data Sets Used
●ROC Results
●Conclusions/Discussion

3. Problem Statement
●Review and compare current methods of
Background Subtraction that have been
implemented, internally in SIG as well as in
Machine Vision community.
●Create a methodology to quantitatively compare
these methods of background subtraction.
●Summarize strengths and weakness of methods
●Conclusions / Discussion points going forward

4. Methods in Community
Gaussian Parametric Models
Mixture of Gaussians (MoG)
●Adaptive background mixture models for real-time tracking - C
Stauffer, WEL Grimson
●Understanding background mixture models for foreground
segmentation - PW Power, JA Schoonees
Bimodal
●W 4: real-time surveillance of people and theiractivities - I
Haritaoglu, D Harwood, LS Davis
Univariate Guassian
●Pfinder: Real-Time Tracking of the Human Body - CR Wren, A
Azarbayejani, T Darrell

5. Methods in Community
Non Parametric
Kernel Density Estimation
●Non-parametric model for background subtraction - A Elgammal, D
Harwood, L Davis
●Motion-based background subtraction using adaptive kernel density
estimation - A Mittal, N Paragios
●Background and Foreground Modeling Using Nonparametric Kernel
Density for Visual Surveillance - A Elgammal, R Duraiswami, D
Harwood, LS Davis
Codebook
●Background modeling and subtraction by codebook construction - K
Kim, TH Chalidabhongse, D Harwood, L Davis

6. Methods in Community
Many other Variations
Review Paper
●Background subtraction techniques: a review - M Piccardi
Eigenbackground
●Efficient mean-shift background subtraction - T.Jan
Conditional Random Fields
●A Dynamic Conditional Random Field Model for Object
Segmentation in Image Sequences - Y Wang, Q Ji

7. Methods – SIG Implemented
●Single Gaussian (full covariance in RGB)
(UNI) currently checked into SVN.
●Mixture of Gaussians (MoG) based on
Stauffer and Grimson (loosely).
●Hybrid – modification of UNI with ideas
from MOG model (YcRcB, median, clique)
●Multiple Mean Model – Previous SVN
check in

8. Methodology – Data Sets
Duke Data
●Hand ground truth key frames, adaBoost –
waiting for code
USF Gait Data
●Outdoor gait data, hand ground truth
VSSN 2005 Open Source Algorithm Competition
●Outdoor background, sprite foreground
Vienna University of Technology, WP2 Evaluation,
Integration, and Standards
●Outdoor background, sprite foreground of
varying intensities

9. Methodology - ROC
●Pd/Pfa calculated over all frames from one
clip
●Maximum likelihood of background
probability is found at various thresholds
●Probability detection: ratio of ML detected
pixels over all pixels in ground truth
●Probability of false alarm: ratio of ML
detected pixels not in ground truth over
(frame size – ground truth)

10. USF Gait Data

11. USF Gait Data

12. Results – WP2 Data

13. Results – WP2 Data

14. Results – VSSN Data

15. Results – VSSN Data

16. Comparison of Methods
UNI MoG Hybrid
Execution Time 20 fps (>20) 17 fps
Memory Requirement M*N*9 M*N*K*4 M*N*3
Variations in Lighting - Adapts
Background
-
Jitter of Bgnd Objects - Compares
Clique
Compares
Clique
Bgnd Initialization Average 30
Frames
Uses first frame
then adapts
Median 30
Frames
Temporal Correlation - - -

17. Conclusion / Discussion
●We have a methodology to compare
background subtraction methods.
●Choosing improvements on background
model affects performance more than the
choice of the initial model itself.
●Method going forward to choose and
improve Background Model.