Enst/uob/lu@trecvid2007 High Level Feature Extraction Using 2-level Piecewise Gmm

We describe a high level feature extraction system for video. Video sequences are modeled using Gaussian Mixture Models. We have used those models in the past to segment video sequences into 2D+time objects. The segmentation result has been used with great success in a compression scheme. In the present work, the Gaussian components of the model are considered to completely model the corresponding objects in the video. Their parameters are used as low-level features for a high-level model used for the detection of a topic or high level feature. The system is not optimized for a particular feature and is thus scalable to any number of features. A threshold is manually selected for each feature after normalization. The only difference between runs was normalization. We tested two runs: B_ENST_1 uses znorm per topic per video. B_ENST_2 use znorm per video. The second system provided better results. This is an initial system that will be used to explore the effectiveness of the modeling of videos using GMMs.