Information Technology Reference
In-Depth Information
space time interest point detected in the previous feature detection phase is
located at the central point in all 3 image planes. In Figure 7 the three ex-
tracted planes can be seen and LBP will be applied to each of them. CSLBP-
TOP is the computation of CS-LBP operator on Three Orthogonal Planes.
Kellokumpu et al [9] have recently used LBP-TOP for human detection and
activity description. However, their approach is based on background subtrac-
tion using LBP-TOP and a bounding volume has to be built around the area
of motion. Their method can be categorize as holistic, since no space-time
interest points have to be detected and differs from our part-based approach.
3.1 Modifications on LBP-TOP
As we described previously, the original LBP-TOP descriptor is the compu-
tation of LBP on the gray-level values of 3 orthogonal slices of each cuboid
(XY, XT and YT planes) resulting in 3 LBP codes (XY-LBP, XT-LBP and
YT-LBP).
Extended (CS)LBP-TOP
We propose to extend the computation of LBP to 9 slices, 3 for each axis.
Therefore, on the XY dimension we have the original XY plane (centered in
the middle of the cuboid) plus other two XY planes located at 1/4 and 3/4 of
the cuboid's length. The same is done for XT and YT dimensions. We named
this method as Extended LBP-TOP.
In this manner, more dynamic information in the cuboid can be extracted,
as the 3 slices in one axis capture the motion at different times. We also
exploit more information from the cuboid, dealing with 6 slices on each axis,
located from
2
/
8
until
7
/
8
of the cuboid's length for each axis. In this case, a
dimensionality reduction technique has to be applied since the final dimension
of the descriptor vector would be too high.
If the CS-LBP operator is computed on each slice, we define the mehods
as Extended CSLBP-TOP.
Gradient and Gabor (CS)LBP-TOP
Another modification we introduced is the computation of LBP operator
on gradient or Gabor images of the orthogonal slices. The gradient image
contains information about the rapidity of pixel intensity changes along a
specific direction, while the gabor image better highlight the area of motion.
The gradient of an image, has large magnitude values at edges and it can
further increment LBP operator's performances, since LBP encodes local
primitives such as curved edges, spots, flat areas etc.
For each cuboid, the brightness gradient is calculated along
x
,
y
and
t
di-
rections, and the resulting 3 cuboids containing specific gradient information
Search WWH ::
Custom Search