EXISTING TECHNIQUES OF VISUAL PROCESSING - Video Object Extraction and Representation: Theory and Applications - page 33

Figure 2.4.
path in 3-D space of video sequence has a constant intensity.

The optical flow constraints assumes that a point that moves along a

continuity) by either modeling the correlation of intensity and position
(or integrating the known structure of the object into analysis).
First, we consider concepts such as locality and connectivity that do
not require a priori knowledge of the object. These types of analyses
enhance or reject the local information through the consideration of a
larger context. For a region-based approach, if regions share a large
proportion of their border, then differences in their intensities may be
overcome and these regions may be joined together. For an edge-based
representation, if we can find a set of edges that form a closed contour,
then these edges are more likely to belong to an object boundary. Such
analysis includes region growing algorithm [Zucker, 1976], snake algo-
rithms [Kass et al., 1988], edge following algorithms, pyramidal image
representation [Pappas, 1992] [Bouman and Liu, 1991] [Horowitz and
Pavlidis, 1976] and model fitting [Hotter and Thoma, 1988].
For the knowledge-based techniques that use a priori information, we
defer our discussion to section 4., since this type of analysis intertwines
both segmentation and representation.

3. MOTION ESTIMATION
Unlike images, video has the concept of motion and we can use it as
a strong discriminant of object membership in the video space. In some
cases, motion alone can separate a video sequence into video objects. In
the previous section, we studied techniques to exploit spatial locality;
in this section, we now exploit the spatio-temporal correlation. Using
the assumption of the existence of the physical object, we can trace
the movement of a physical object through the video space as a set of
separate parallel paths under certain conditions.