ENBIS: European Network for Business and Industrial Statistics

Gaussian mixture model-based processing of video images

23 September 2009, 11:55 – 12:15

Abstract

Submitted by

András Zempléni

Authors

Zs. Robotka – A. Zempléni

Affiliation

Eötvös Loránd University, Budapest

Abstract

In our previous paper at the ENBIS 2008 meeting we have used Gaussian
mixture models (GMMs) in image clustering. To process video data is an
even harder, but at least as important task.

Different features, such as motion, color histogram, have been usually
utilized for video analysis. There are earlier papers which have used
Gaussian mixture models, but our approach differs from theirs in its
simplicity and parametrisation. Our model is based on a sample from
the pixels of the video. This allows for quick processing, while
keeping the key features. We introduce two possible applications of
this method. The first is anomaly detection which is an important
problem in manufacturing, where one is interested in detecting defects
in the products. Another use includes the compression of video data,
since the residual has a much simpler structure and the GMM is very
easy to store.

We illustrate our methods on real video data, and provide comparisons
with known methodologies.

View paper

Return to programme