Computational Intelligence - February 2013 - 64

(a)

(b)

(c)

(d)

Figure 14 Behaviors detected by our model in QMUL dataset2. (a) Vertical and vertical-right
traffic flow, (b) only vertical traffic flow, (c) pedestrian across road, and (d) pedestrian across
road when there is traffic flow.
Table 4 Weighted average recall precision and accuracy of four algorithms in qmul
dataset1.
algoriThm

exhausTive

geneTic

covering

lem2

recaLL

0.610

0.599

0.565

0.702

Precision

0.625

0.626

0.574

0.697

accUracy

0.605

0.588

0.582

0.776

coVerage

0.962

0.955

0.658

0.283

Table 5 Weighted average recall precision and accuracy of four algorithms in qmul
dataset2.
algoriThm

exhausTive

geneTic

covering

lem2

recaLL

0.503

0.505

0.438

0.610

Precision

0.549

0.551

0.452

0.583

accUracy

0.5

0.503

0.441

0.617

coVerage

0.98

0.975

0.885

0.3

Table 6 Confusion matrix of exhaustive algorithm in qmul dataset1.
behavior 1

recall

coverage

0.388

0.706

0.988

105

0.754

0.943

0.273

0.977

0.715

0.948

0.55

0.85

0.35

0.67

Precision 0.38

IEEE ComputatIonal IntEllIgEnCE magazInE | FEbruary 2013

At first, we set the number of topics
to 4. As shown in Fig. 7, the 4 topics
indicate 4 directions. We also set the
number of topics to 4 in the original
LDA, as shown in Fig. 8, the topics
under our model are different from the
LDA's. Because moving directly up and
down co-occur frequently, they are
assigned to the same topic, which can
be seen in Fig. 8(a). Since there are
more than 4 behaviors in the video,
4 topics are not enough to represent
the behaviors. Therefore, some different behaviors will be assigned to one
topic though they are not "frequently"
co-occurring, which can be seen in
Fig. 8(b). We can see that the cutting
across traffic behavior and the vertical
traffic behavior are assigned to the
same topic in Fig. 8(c) and (d).
We then set the number to 5. From
the results in Fig. 9, we find that there
is a topic red that indicates cutting
across traffic. The 4 topics still indicate
4 directions.
When the number comes to 6, it is
shown in Fig. 10 that the vertical traffic
flow has been split into more fine granularities. From Fig. 10(a) and (c) we
note that the traffic flow toward left has
been split into two topics: blue and
cyan. The blue occurs when there is
cutting across traffic (from up to left) or
there is pedestrian on the left. The cyan
occurs when there is horizontal traffic.
In Fig. 10(d), the traffic flow downwards has split into two topics: pink
and red. The red occurs when the traffic flow will turn right, and the pink
just indicates that the traffic flow moves
directly down. The two remainder topics cyan and yellow indicate the horizontal traffic flow.
In Table 2, it can be seen that the
granularity of activities becomes better
with number of topics increasing.
2) Behavior Detection:
As we discussed in section I, the
abnor malities are combinations of
atomic activities that cannot co-occur,
and the experiments above show that
the more topics there are, the finer
activity granularity will be. As there
are 6 atomic activities, no more than

Table of Contents for the Digital Edition of Computational Intelligence - February 2013