Digital Signal Processing Reference
In-Depth Information
Given two detected people, their distance, coexisting duration, and motion
direction's correlation are combined to form a feature vector, which are extracted
in a sliding window of twelve consecutive frames. One-vs.-All SVM is trained to
classify this general category of events. To distinguish “
Embrace
”or“
PeopleMeet
”,
a backward search is applied to locate the beginning of an event. In contrast, forward
search is used to detect “
PeopleSplitUp
”.
Finally, the results are improved with post-processing in which a set of heuris-
tic rules are used. For instance, if two peoples' distance at the end of an event is
greater than a threshold for “
PeopleMeet
”and“
Embrace
”, or their distance at the
beginning of an event is greater than a threshold for “
PeopleSplitUp
”, those would
be considered as a false alarm.
4.4.2.2
Single-Actor Events
Speed and direction of movements are key characteristics of “
PersonRuns
”. It is
observed that a running person have a larger velocity than others, and the motion
direction would not change dramatically. According to the feature statistics, the con-
straints of object position and motion direction can be used in a SVM classifier to
identify
PersonRuns
. In the camera setting of TRECVID dataset, running people
always move from left-bottom to top in the view of camera one. So, by using the
post-processing step we may remove many false alarms caused by tracking drifting.
ElevatorNoEntry
is defined as “elevator doors open with a person waiting in
front of them, but the person does not get in before the doors close”. As illustrated
in Fig.
4.14
, an automaton is adopted to model the detection process of
Eleva-
torNoEntry
. As there is no elevator in the view of cameras one, two, and five
over TRECVID'09, the automaton is executed only in the views of camera three
or four.
As the elevator's position is fixed, the system can easily locate the elevator. When
the elevator door is closed, the elevator region is labeled as background. And when
Fig. 4.14
An automaton for detecting
ElevatorNoEntry
