Information Technology Reference
In-Depth Information
pattern describes the local shapes contained on a grey-scale image. Thus, in
order to extract the LBP value corresponding to a pixel with a grey level g c ,we
need to define a circle of radius R around that pixel, and from that circle select
P equidistant points. The grey level of these P equidistant points will be used
to get the LBP value:
P− 1
g c + a )2 p
LBP P,R =
s ( g p
(1)
p =0
where g p are the grey levels of P points located in the aforementioned circle,
a is constant value used to reduce the noise impact, and s ( x ) is a thresholding
function:
s ( x )= 1 ,x
0
0 ,x< 0
(2)
In our case in particular we have set P =8and R = 1, in order to reduce the
computational time required to calculate the LBP.
We have used a second texture descriptor aimed at discriminating the torsos
upon their edge density. This gives a good measure of how textured an image
is. This descriptor is based on the the edges of the image, obtained with the
canny edge detector. Then, a convolution with a smoothing kernel is applied to
include information about the separation among edges in the image. We would
like to remark that this descriptor showed a high discrimination power during
our tests.
Once the person-following behaviour starts, the target model is initialized
starting from the person detected in front of the robot. We now will describe
how we use the features we just described in order to find the target in the
image.
2.2 Target Discrimination
The target discrimination algorithm pursuits the discrimination of the target
from the distractors (other people in the scene different from the person being
followed). This algorithm takes as inputs the torsos extracted from the people
detected in the image, and outputs the dissimilarity of each torso with the target
model.
In order to obtain the dissimilarity value amongst each torso and the target
model, we need to get the histogram for the torso and each feature, h torso,f ,
and compare it against the histogram obtained for the same feature but for the
target model h ref,f :
h torso,f ( i ) h ref,f ( i )
b− 1
1
b 2 h torso,f h ref,f
d f
=
1
(3)
i =0
d f is the normalized Bhattacharyya distance, h ref and h are the average values
of the histograms, b is the number of bins of the histograms. Starting from these
Search WWH ::




Custom Search