Game Development Reference
In-Depth Information
Fig. 10.2
The effects of video compression quality on feature analysis, object detection, and face
recognition.
a
SIFT features are extracted on video frames with different definitions and different
QPs, and then the average similarity is calculated between SIFT features of the original frames and
those of the corresponding transformed frames.
b
In the object detection experiment, the well-known
deformable part model (DPM) trained by PASCAL VOC 2009 dataset (Felzenszwalb et al.
2010
)is
used as the detector for pedestrians and vehicles.
c
The face recognition algorithm (Su et al.
2009
)
is used
same time, however, such a video surveillance systemwould produce a larger amount
of video data, leading to much more expensive costs for the network bandwidth and
the storage system. To solve this problem, most of current video surveillance sys-
tems often compress a high-definition video of 1.5Gbps to 10Mbps by using H.264
high profile (H.264 HP for short) (Ayers and Shah
2001
). Some systems even set
the compression rate to 300:1 or higher so as to further reduce the network band-
width and the storage capacity. Nevertheless, high compression ratio will inevitably
bring difficulties in visual feature extraction, consequently reducing the recognition
accuracy of video analysis and recognition more or less.
To evaluate the effects of video compression on typical analysis tasks, we con-
ducted several experiments on one 1080p uncompressed video selected from the
PKU-SVD-Adataset, whichwas then resized to different resolutions and compressed
with different quantization parameters (QPs). Generally, QP regulates howmuch spa-
tial detail is saved in the compression. When QP is very small, almost all that details
retain; As QP increases, some of that details are aggregated so that the bit rate drops,
at the price of increase in distortion and loss of quality. As shown in Fig.
10.2
,we
can see that for all the three tasks, the performance is gradually going down with the
increase of QPs. In this case, the advantages of introducing high-definition cameras
would be fully offset by the used video coding technology.
In order to overcome the limitations of centralized intelligent video coding system,
the distributed intelligent video coding system should at least have the following two
features:
1. Recognition-friendliness should be its first and most important objective. In such
a smart system, the majority of coding bits should be used to represent the objects
or regions-of-interest (ROIs). It even can identify the clear facial images, extract
the features of the objects and track their trajectories in real time in the coding
process. These pieces of information are highly useful for further analysis and
recognition, yet cannot be easily obtained in the traditional surveillance video
systems.
