EXISTING TECHNIQUES OF VISUAL PROCESSING - Video Object Extraction and Representation: Theory and Applications - page 37

Figure 2.7. Balancing smoothness and optical flow for Motion Estimation. Given a)

the intensities of a section of a frame and its following frame, we can rate a given
motion field with respect to Eq. 2.5 and Eq. 2.6. In b), we have a zero error with
respect to the optical flow, while creating an incoherent motion field. In c), as in
block matching, we have completely smooth motion field and incur optical flow cost.
In d), a balance between smoothness and optical flow result in the most accurate
motion field.

function of time. The simplest of these schemes is as follows: watch a
given spatial location on the frame, mark that pixel location as non-
background if the intensity of pixel changes beyond a threshold. For
non-stationary backgrounds, we can compensate the background mo-
tion through offsetting of the frames, and still use change detection to
find non-background regions. However, the reliance on motion compen-
sation compromises the robustness of change detection since the change
detection now depends on the accuracy of motion compensation. Some
movement is not equivalent to change: for instance, homogeneous re-
gions may not exhibit any change properties through temporal corre-
lation. Also, change detection is a binary operator that distinguishes
between only two different classes (background and non-background) of
objects and may misgroup objects in scenes with three or more object
classes. An example of a system that uses signal analysis upon ∂ to sep-
∂

I
arate background and foreground is a scheme based upon a high ordered
statistics (HOS) [Neri et al., 1998].

-