Cryptography Reference
In-Depth Information
9.3.3 Experiment Results
To test the effectiveness and e
ciency of our method, we took a compressed
four-hour MPEG-2 test video from eight different locations as shown in
Fig. 9.6. From the four-hour video, we extracted 108 video clips that con-
tained motion. We call these clips with motion events. To simplify the QBS
process, we used one of the eight locations as the background to assist in
drawing a query sketch. The left-hand side of Fig. 9.14 shows a query sketch
of a man climbing a stairway. Among the 108 events, the top three retrieved
events were all of the man climbing the stairway at different times in the
sequence. The 4th retrieved sequence was a descending the stairway event,
which was retrieved because of its high degree of similarity to the top three
retrieved events. In another experiment, we drew a climbing-wall trajectory
with the help of the user interface (as shown in Fig. 9.15). From the top six
retrieved events, it is obvious that the 1st, 3rd, 4th, 5th, and 6th all relate
to the query. The 2nd retrieved event happened at another location; however,
its corresponding trajectory was very close to the query sketch. In sum, we
consider that the proposed method is especially useful in surveillance-related
systems. It is also useful for a general coarse search. Since the approach can
retrieve video clips from a large database in a very e
cient manner and most
of the retrieved results are close to the input query, we can use the approach
as a coarse search engine. It is obvious that video databases of the future will
be very large indeed. If we do not manage this kind of database e
ciently, it
will not be possible to enjoy to the full extent of the multimedia world of the
future.
9.4 Conclusion
In this chapter, we have presented solutions for video-based surveillance sys-
tems in the spatial domain and in the compressed domain, respectively. In spa-
tial domain event detections, a simple, but e
cient, color blob-based tracker
has been developed to accomplish the multi-object tracking task. For event
detection in compressed videos, we use the information of motion vectors in an
MPEG bitstream to generate some trajectory-like motion flows. Since motion
vectors are embedded in compressed videos, we can process multiple moving
objects in a shot. After applying the Douglas-Peucker algorithm to approx-
imate a trajectory, we are able to compare two arbitrary trajectories. The
above mechanism enables us to conduct both spatial domain and compressed
domain event detections if a number of wanted trajectories are pre-stored in
a video-based surveillance system.
