Digital Signal Processing Reference
In-Depth Information
Tabl e 4. 4
Results of the system in the TRECVID-ED 2009 formal
evaluation
Event
#Ref
#Sys
#CorDet
#FA
#Miss
Act. DCR
PeopleMeet
449
125
7
118
442
1.023
PeopleSplitUp
187
198
7
191
180
1.025
Embrace
175
80
1
79
174
1.02
ElevatorNoEntry
3
4
2
2
1
0.334
the door is moving, the elevator region is detected as foreground. Thus, each eleva-
tor's states (open or closed) can be identified by using background subtraction. The
elevator's opening and closing moments are thus detected and recorded. The fore-
ground area size is related to the number of persons in front of an elevator. Given
an incoming frame, the
ElevatorNoEntry
event can be detected according to the
elevators' states and the size of foreground area. Moreover, the area ratio of detected
foreground regions before and after the detected elevator open-and-close action can
be computed to tell whether the number of persons around an elevator has changed.
When the ratio is less than a threshold, it is probable that some people enter the
elevator, and the frame interval is labeled as a potential event of
ElevatorEntry
.
4.4.2.3
Evaluation Results in TRECVID 2009
This system is evaluated over the surveillance event detection task in TRECVID
2009. Three runs were submitted, by using different human detection and tracking
modules. According to the comparative results in the TRECVID-ED formal evalua-
tion, the experimental results are promising. Among all submissions for the formal
evaluation, four detection results (e.g.,
PeopleMeet
,
PeopleSplitUp
,
Embrace
,and
ElevatorNoEntry
) ranked at the first place (Table
4.4
).
However, there are some problems yet. Regarding the results, the false alarms
have been reduced greatly by effective post-processing. Unfortunately, much correct
detection is wiped off at the same time. In other words, the system recall is too low.
Furthermore, the system should make a good tradeoff between false alarms and
system recall.
4.5
Summary
This chapter presents a tutorial on the problems and solutions of video scene
analysis from the perspectives of the learning components and tasks. Two major
categories of the state-of-the-art tasks are discussed in this chapter, based on their
application setup and learning targets: generic methods and genre-specific analy-
sis techniques. Many of the works discussed in this chapter are deemed by the
authors as good representatives of existing learning-based video scene analysis
