Game Development Reference
In-Depth Information
Fig. 8.1
The gap between the growth rate of surveillance videos and the video compression rate
in the recent three decades
Surveillance videos own one special data redundancy beyond the traditional tem-
poral, spatial, information entropy, visual and structure redundancies, the background
redundancy, which is one kind of knowledge redundancy. The knowledge redundancy
indicates the videos' regular textures which can be detected by people's experience
and knowledge. Knowledge redundancy is the major characteristics of model-based
coding wants to make use of, but there are still difficulties in how to express it. How-
ever, in surveillance videos, there are fewer difficulties due to the relatively fixed
background information. For example, the unchanged objects like buildings, the
regularity of light changes with time, and the seasonal changes of trees, leaves, and
flowers, etc. We call this regular changing or relatively fixed background information
as background redundancy.
There are two kinds of background redundancies in surveillance video, namely the
exposed-region prediction redundancy and block matching prediction redundancy.
Figure
8.2
identifies the exposed-region prediction redundancy. Some background
data (circled regions) of the current frame has no suitable prediction reference in the
recent reference frame and key frame, but has suitable reference in the modeled back-
ground picture. Therefore, such exposed background redundancy can be removed
by using better modeled background picture as reference. The block matching pre-
diction redundancy originates the structure of the block, which may be mixed with
background and foreground pixels. For such blocks, block matching methods cannot
have very good prediction efficiency. A typical example is shown in Fig.
8.3
.Inthe
figure, the hybrid block A and its best matched block A' has no similarity in the
background pixel part, whereas B and B' has no similarity in foreground part. More-
over, the interframe prediction structure can also be viewed as a kind of background
redundancy of surveillance videos, which means the similar prediction structure in
the background could be removed.
