Information Technology Reference
In-Depth Information
are summed in absolute values. Before computing the image gradients, the
cuboid is slightly smoothed with a Gaussian filter in order to reduce noise.
LBP-TOP is then performed on the gradient or gabor cuboid and we
name this method Gradient LBP-TOP and Gabor LBP-TOP, respectively. In
Figure 7 the different cuboids are shown and the three slices XY, XT and
YT are highlighted. Figure 10 shows the three XY slices extracted at 1/4,
1/2 and 3/4 of the cuboid's length for the Extended LBP-TOP method.
The Extended LBP-TOP can be applied on the gradient cuboid and we
named this method as Extended Gradient LBP-TOP or on the gabor cuboid
and we named this method as Extended Gabor LBP-TOP.
If the CS-LBP operator is computed on each slice, we define the mehods
as Gradient CSLBP-TOP, Extended Gradient CSLBP-TOP etc. In the result
section, more figures are shown.
4 Experimental Results
The dataset we used for our evaluation is KTH human action dataset [16].
This dataset contains six types of human actions: walking, jogging, running,
boxing, hand waving and hand clapping. Each action class is performed sev-
eral times by 25 subjects in different scenarios of outdoor and indoor environ-
ment. The camera is not static and the videos present illumination variations,
slightly different viewpoint and contain scale changes. In total, the dataset
contains 600 sequences. We divide the dataset into two parts: 16 people for
training and 9 people for testing, as it has been done in [16] and in [11]. We
limit the length of all video sequences to 300 frames.
We extract the STIP and describe the corresponding cuboids with the pro-
cedure described in Sections 2.2 and 2.3. The detector parameters of formula
1aresetto
σ
=2
.
8
and
τ
=1
.
6
, which gave better results in our evaluations,
and 80 STIPs were detected for each sequence.
Once a cuboid is extracted from the original video sequence, it is described
with the proposed methods in order to be robust and discriminative. In the
left side of Figure 8 the original video patch extracted from a running se-
quence is shown, while on the right side each slice of this cuboid is shown
side by side. In this Section, we show the results yielded by several methods
of feature description.
Fig. 8 Original video patch
Search WWH ::
Custom Search