Game Development Reference
In-Depth Information
Table 8.7
Memory cost (byte) for each pixel in background modeling
Item
RA
GMM-1
GMM-5
MS
SWRA
Buffered pixel
1(char)
1(char)
1(char)
M(char)
1(char)
Mean values
1(float)
2(double)
2(double)
1
2(float)
Weight
0
1(double)
1(double)
0
2(char)
Threshold
0
1(double)
1(double)
0
1(float)
Match points
0
1(char)
1(char)
0
0
Tot a l
5
34
345
=
170
M
=
120
14
8.4 Surveillance Coding in AVS
In AVS and AVS2 standard, the above segmentwise-weighted average background
model has been adopted (encoder) for background model-based surveillance video
coding methods. The related tools include background modeling, background pic-
ture (G-picture) based prediction, i.e., background-predictive picture (S-picture),
and background difference prediction (Zhang et al.
2014a
). Moreover, AVS2 also
supports the signalization of the regions of interest (ROI) in picture header. While
combining with the G/GB picture (MBG mechanism), surveillance video's ROI can
be easily detected from background subtraction. And then, following the signaliza-
tion method of ROI in AVS2, the decoder can obtain the ROI regions directly from
video stream without any complex operations. Therefore, ROI signalization com-
bined with MBG is one practical tool for fast ROI extraction directly from video
streams.
In this section, we firstly introduce the background model-based video codec
architecture. Secondly, we will introduce the novel tools for better utilizing theMBG,
i.e., background difference prediction and MVP related to MBG.
8.4.1 Video Codec Architecture
In AVS and AVS2, G-picture and S-picture are defined to further exploit the temporal
redundancy and suppress background noise. The G-picture is a special I-picture and
its reconstructed picture is stored in a separate background memory. The S-picture
is a special P-picture which can be only predicted from reconstructed G-picture. For
surveillance video coding, the G-picture can be a nondisplay MBGwhich is modeled
from input pictures and encoded into stream to guarantee the decoding match, and
then each P-picture can also utilize G-picture as prediction reference or be encoded
utilizing background difference prediction.
