by illegal state transition. A fire engine interrupts the current
vertical flow. The red boxes highlight the abnormal agents.
In Figure 11, the traffic state is forced to change in an
illegal ordering due to the fire engine. The rightward flow
should not follow the vertical flow according to the learning
results by
HDP
-
HMM
models. Therefore, the clip was identified
as abnormal event, even though its appearance is definitely a
right flow state.
Table 1. Comparison of Classification results between our methods and
others popular methods for
QMUL
Junction Dataset: The results of
MCTM
,
LDA
and
HMM
are cited from
[8]
.
State
MCTM
LDA
HMM
Ours
L R V VT L R V VT L R V VT L R V VT
Left
.99
.00 .00 .01
.49
.44 .00 .06
.98
.00 .01 .01
1.0
.00 .00 .00
Right
.00
.94
.01 .05 .00
1.0
.00 .00 .00
.92
.08 .00 .00
.99
.00 .01
Vertical
.00 .00
.77
.22 .01 .17
.82
.00 .02 .01
.69
.28 .00 .00
.98
.00
Vertical-
Turn
.31 .05 .20
.43
.01 .21 .30
.46
.49 .04 .32
.15
.05 .00 .00
.95
Average
Accuracy
.78
.69
.69
.98
Figure 10. Examples of abnormal events caused by rarely occurring motions. In the training dataset such motions have rarely
or never occurred. They do not belong to any typical activities. The red boxes highlight the abnormal agents.
Figure 11. Example of abnormal event caused by illegal state transition. A fire engine interrupts the current vertical flow. The
red boxes highlight the abnormal agents.
Figure 9. Example of falsely classified by GP classifier.
Table 3. Classification
performance for the
MIT
dataset.
Manually label
Our Classification
a b c d
a 86 2 1 2
b 2 264 0 4
c 0 0 188 2
d 0 2 0 76
Table 2. Classification performance of the
MIT
dataset.
State
Dural-HDP
Ours
a b c d e a b c d e
Manually
label
a 149 0 2 0 0 610 4 5 0 3
b 8 74 4 2 11 3 402 0 2 0
c 10 3 60 1 2 3 2 304 2 0
d 4 0 2 88 11 7 8 10 222 0
e 4 2 6 5 92 6 5 4 8 102
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
April 2018
211
