Computational Intelligence - February 2013 - 50

where, TP is the number True Positives,
regions containing target motions being correctly labelled as positive. TN is the number of
True Negatives, regions with no motions or
containing uninteresting motions being classified as negative. TOTAL is the total number of
examples. This evaluation criterion is basically
the classification accuracy, or detection accuracy. The weights on TP and TN are the same
so there is no particular emphasis on true positive or true negative.
Table 1 shows the runtime parameters for the training. The
minimum depth of GP tree is 2 because it is impossible for a
single-node to classify correctly on a reasonably complex problem. The maximum tree depth in previous studies was set to 9.
However it is set as 5 here to encourage small programs.
Strongly-typed GP is used. The population size is 200 in this
study so the chance of finding a small but good detector would
be greater. Our previous studies used a population size of 30,
which was sufficient for generating an accurate detectors for
those tasks. The rest of the parameters are consistent in all of
our studies. The choice of low mutation rate is also consistent
with other GP studies related to image and vision. These
parameter settings are not necessarily the best combination;
finding the optimal setting is not our goal.

One pixel may be sampled multiple times by the window
at different positions. Hence this pixel is marked multiple
times. Based on that, the final step is to integrate these
multiple labels on each pixel into one by voting strategy.
A pixel would be marked as positive if the majority
marks it has received are positive.
computational resource and increase the search space. On the
other hand a small n may not be sufficient for accurate detection. In our previous studies, most of the experiments used 20.
In this study, n is set as 5 to facilitate the analysis.
The labeled data set for each task is randomly divided into
two sets: the training set and the test set. The evolution process
is based on the training set.The selection process is based on the
test set.The best detectors from the last generation of the evolution are evaluated against the test set.The best achiever on test is
then chosen for the application phase as shown in Figure 4.
Due to the nature of motion detection problem, regions where
genuine motions occur usually only occupy a small proportion
of a video frame compared to regions of negative cases. Therefore negative cases are the majority. Most of the them are
removed by a random process so the training data and test data
would not have severe imbalance.
The fitness function for the training process is shown below:
Fitness = TP + TN # 100 ,
TOTAL

(1)

Table 1 GP runtime parameters.
MiniMuM Depth

MaxiMuM Depth

pOpulatiOn Size

1 200

GeneratiOnS

200

MutatiOn rate

CrOSSOver rate

85%

elitiSM rate

10%

Original
Video

Application Phase

Marked
Video

MF-A
Pseudo
Frame

Sliding Window

Create Mask

GP
Classifier

Voting
Counter

For All Windows

Figure 5 Overview of application phase.

IEEE ComputatIonal IntEllIgEnCE magazInE | FEbruary 2013

B. Application Phase

The application phase applies the best program from the evolution phase onto unlabeled video streams to perform the detection. However the n # n region size used in the evolution
phase is likely too small for video frames. It is also too coarse to
mark the contour of objects in motion. Instead a detector is
applied by the process shown in Figure 5.
The first step is to transform consecutive frames into
pseudo-frames or converted frames by the multi-frame representation discussed in Section III-C. The second step is to
sample sub-images on a pseudo-frame by a sliding window of
which the size is the same as the region size in training and test
sets, n # n. This sliding window moves from the top-left corner
to the bottom-right corner of the pseudo-frame. At each sampling position, the sub-image is fed into the GP detector
selected from the evolution phase. If the output of the detector
on this sub-image is positive, then all the pixels under that sliding window position are marked as positive. Otherwise all pixels are marked as negative.
The sliding window moves at a certain interval.When this step
size is smaller than n, there are overlaps between these positions. As
a result, one pixel may be sampled multiple times by the window
at different positions. Hence this pixel is marked multiple times.
Based on that, the final step is to integrate these multiple labels on
each pixel into one by voting strategy. A pixel would be marked
as positive if the majority marks it has received are positive.
Otherwise it would be considered a pixel belong to non-motion
category. Then a marked video can be produced as the detection
output. Our previous studies often set the step size as 8.

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