1. Real-time background estimation for video surveillance applications
Mahfuzul Haque, Manzur Murshed and Manoranjan Paul
Gippsland School of Information Technology, Monash University, Victoria 3842, Australia
Email: {Mahfuzul.Haque, Manzur.Murshed, Manoranjan.Paul}@infotech.monash.edu.au
Abstract
Background subtraction is the initial step for extracting foreground information from a video sequence in surveillance systems for
object tracking and activity recognition. Basic Background Subtraction (BBS) technique is sufficient in controlled environments
where system initialization is possible with a clear background image and where background is uncluttered and illumination change
is rare. But for real world environments adaptive background model like Gaussian Mixture Model (GMM) is used to adapt with those
situations. This research attempts to estimate the background from a given video sequence at a particular instance of time using a
modified version of the existing Gaussian Mixture Model. Instead of pixel classification as foreground or background like other
methods, the estimated background is used in background subtraction from the current frame for foreground extraction. Several
experiments with real and synthesized videos show that the proposed method performs better in background estimation and
foreground extraction where clear background is rare or never present.
Motivation
Background model performance evaluation Toolkit
1. Basic Background Subtraction technique does not work in real world
environment with dynamic and cluttered background scenes.
2. Adaptive background models are used to adapt with the dynamics of the real
world with pixel level classification as background and foreground.
3. A real-time background estimation model will simplify the foreground
extraction process by employing the basic background subtraction.
Background model construction and estimation
Current frame
Background
Model
Input
Frames
Estimated
background
Background
Subtraction
Current frame
Sample experimental results
Foreground
Post
Processing
Technique
Input
frame
Results
1. Background estimation model is constructed using the basic formulation
introduced by [1][2].
2. The history of each pixel {X1,X2,âŠ,Xt} is modelled by a mixture of K Gaussian
distributions where Xt is the pixel value of {x,y} at time t.
3. The distribution parameters, âi,t, ”i,t and wi,t are learned and updated after each
pixel observation.
4. Unlike performing pixel classification as background and foreground like existing
methods [1][2][3][4], the proposed method first estimates the background and
extracts the foreground using the basic background subtraction.
5. The extracted foreground can be used in activity recognition and tracking in
surveillance applications.
Frame 1
Frame 2
..
Frame t
Input frames
Foreground
Frame 1
Frame 71
Frame 297
Conclusion
Proposed method simplifies foreground extraction from dynamic and cluttered
video scenes and shows better foreground extraction result than existing methods.
Gaussian Mixture Model (GMM)
for each pixel
P( X t ) ïœ ï„iïœ1 wi ,t ïȘïš ( X t , ïi ,t , ïi ,t )
K
ïš ( X t , ït , ï t ) ïœ
Estimated
background
1
(2ï° )
n/2
|ï|
1/ 2
e
1
ï ( X t ï ï t )T
2
ï„ ï1 ( X t ï ït )
References
[1] Grimson, W.E.L. and C. Stauffer. Adaptive background mixture models for real-time tracking. in IEEE Conference on Computer
Vision and Pattern Recognition. 1999.
[2] KaewTraKulPong, P. and R. Bowden, An Improved Adaptive Background Mixture Model for Realtime Tracking with Shadow
Detection, in 2nd European Workshop on Advanced Video Based Surveillance Systems. 2001, Kluwer Academic Publishers.
[3] Lee, D.-S., Effective Gaussian Mixture Learning for Video Background Subtraction. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 2005. 27(5): p. 827 - 832.
[4] Zhang, J. and C.H. Chen. Moving Objects Detection and Segmentation In Dynamic Video Backgrounds. in IEEE Conference on
Technologies for Homeland Security. 2007.
The sample video has been collected from http://www.gaitchallenge.org