IV. DETECTION PHASE
TUD dataset [21] is used in this work for testing phase.
Figure 4 shows examples of images in this dataset. As you
can see in the figure 2, the dataset contains images of the
external environment and crowded.
Full body positive image
Full body’s Torso’s FP Left Foot’sFP Right Foot’s FP FPLfoot
positive
image
LFoot
positi
ve
imag
e
Torsos
positive
image
First SVM
Torso Left Foot Right Foot Full body
Tors
o’s
SV
M
LFoot’
s
SVM
RFoot’s
SVM
Full
Body’
s SVM
Torso
’s out
put
LFoot’s
output
RFoot’s
Output
Full
body’s
output
Detection
window
Normalize gamma and
color
Compute gradient
Accumulate weighted votes
for gradient orientation over
spatial cells
Normalize contrast within
overlapping blocks of cells
Collect HOGs for all blocks over
detection window
Figure 3 Second SVM with different negative example
Figure4 Example of testing image
Figure 5 overview of detection phase
A. construction image pyramid
In test phase, the pyramids of test images are
constructed at first. Pyramid construction makes
possibility of human detection in different size
steps are used to build the pyramid. The
image is determined by (3). sinit is initial size and the
number of pyramid construction is shown by
#$
B. feature extraction
After pyramid construction, the feature introduced in
the previous section (HOG), such as all image in the
Training phase and feature
extraction of training image
Pyramid construction on test
image
Convert pyramid size to
training image size
Feature extraction on
pyramid
Applying SVM on feature
Detection
Second SVM with different negative example for each class.
of testing image
overview of detection phase
s of test images are
constructed at first. Pyramid construction makes the
in different size. The eight
The size of each
is initial size and the
number of pyramid construction is shown by interval.
(3)
After pyramid construction, the feature introduced in
all image in the
training phase, is extracted in pyramid. Before feature
extraction of each pyramid, test images must be converted
to the training image size. Since the training segments
have [104 48] pixels, feature extraction in test image must
be converted to [104 48] pixel. The main problem with
such problem is that, it ignores people between
windows. To overcome this problem,
used. Instead of dividing the image
components, overlapping components
5, shows an overview of detection phase.