Cryptography Reference
In-Depth Information
In this chapter, we present solutions for video-based event detection sys-
tem in the spatial domain and the compressed domain, respectively. In the
spatial domain, we first adopt a mixture of Gaussian approach to determine
the background. Through this simple modelling, we can separate foreground
objects from the background e
ciently and accurately. To track foreground
objects, we use color blob-based tracking. However, the method is not perfect
due to the effect of self-occlusion or mutual occlusion among various body
parts [12, 13, 17]. Therefore, the tracking process is designed as a multi-blob
tracking process that generates multiple similar trajectories. We have also
designed an algorithm that merges these trajectories into one representative
trajectory and then apply the Douglas-Peucker algorithm to approximate it.
To compare a trajectory extracted from a real-time environment with those
stored in the database, we propose a translation and scaling invariant met-
ric to execute the matching task. Using the above procedure, we can detect
abnormal intrusion events in real time by pre-storing a number of possible in-
trusion trajectories in a local computer linked to a camcorder monitor. If the
path of an intruder is close to one of the pre-stored trajectories, the system
sends a signal directly to the control center. The contribution of this work is
twofold. First, it simplifies the real-time event detection problem by compar-
ing the degree of similarity between two arbitrary trajectories. Second, the
proposed method is translation and scaling invariant. It can therefore be ap-
plied to different application domains, such as home security systems, complex
surveillance systems, or parking lot management systems.
In the compressed domain, since stored surveillance videos are compressed,
we directly utilize the motion vectors embedded in MPEG bitstreams to de-
velop a motion descriptor. It is known that motion vectors only record the
direction and magnitude of movement between corresponding macroblocks of
two consecutive anchor frames, as they are only comprised of local data that
do not have much semantic meaning. In this study, we utilize the consistency
of motion direction, the color distribution, and the overlapping area between
macroblocks of two consecutive frames to link all neighboring motion vectors.
These linked motion vectors form so-called motion-flows, which contain more
semantic meaning than the original motion vector data. Since a large moving
object may occupy several macroblocks and produce multiple motion flows,
we also propose an algorithm to reduce motion flows that are similar in shape
to one or more representative motion flows. We approximate these representa-
tive motion flows by generating control points and storing them in a database
as models. When a user wants to query a specific motion, he/she can draw
a trajectory on a sketch-based interface, and the system will retrieve a set of
shots that contain similar motion content. Since the temporal information of
a trajectory is hard to express precisely, we only consider the spatial informa-
tion in a QBS (Query-By-Sketch) process. Initially, a user executes the process
to retrieve some candidate shots from the database. Then, he/she can choose
one of candidate shots and execute a QBE (Query-By-Example) process to
extract the video clips that are most similar to the query.
