Information Technology Reference
In-Depth Information
Fig. 2. Overview of our system for video input
the last class is from each key-frame in RGB color space,see subsection 3.2), and
labels them based on corresponding features. For example, if three features are
used (color, texture and shape), each frame has at least three labels from Ω out ,
three labels from Ω in and three labels from key-frame.
This reduces the video as a sequence of labels containing the common features
between consecutive frames. The sequence of labels aim to preserve the semantic
content, while reducing the video into a simple form. It is apparent that the
amount of data needed to encode the labels is an order of magnitude lower
than the amount needed to encode the video itself. This simple form allows
the machine learning techniques such as Support Vector Machines to extract
high-level features.
Our method offer a way to combine low-level features wish enhances the sys-
tem performance. The high-level features extraction system according to our
toolkit provides an open framework that allows easy integration of new features.
In addition, the Toolbox can be integrated with traditional methods of video
analysis. Our system offers many functionalities at different granularity that can
be applied to applications with different requirements. The Toolbox also pro-
vides a flexible system for navigation and display using the low-level features
or their combinations. Finally, the feature extraction according to the Toolbox
can be performed in the compressed domain and preferably real-time system
performance such as the videosurveillance systems.
3 Moving Object Detection and Extraction
To detect and extract a moving object from a video dataset we use a region-
based active contours model where the designed objective function is composed
of a region-based term and optimize the curve position with respect to motion
and intensity properties. The main novelty of our approach is that we deal with
the motion estimation by optical flow computation and the tracking problem
simultaneously. Besides, the active contours model is implemented using a level
Search WWH ::




Custom Search