Information Technology Reference
In-Depth Information
In order to reduce optical noise and boundary matching in raw depth images, a
novel method was introduced for depth image enhancement [14]. For noise reduc-
tion, a newly-designed joint bilateral filtering method was presented. The bilateral
filter reduces the noise while preserving important sharp edges. By applying the
joint bilateral filter onto the raw depth image, we can reduce the artifact in it.
Formally, for some position
p
of the synchronized color image
I
′
and depth image
I
′′, the filtered value
J
p
is represented by Eq. 2.
1
(
) (
)
(
)
∑
∈
′
′
′
′
J
=
G
p
−
q
⋅
G
I
−
I
⋅
G
I
−
I
I
(2)
p
s
r
1
p
q
r
2
p
q
q
k
p
q
where
G
s
,
G
r1
,and
G
r2
are the space weight, color difference in the color image and
depth difference in the depth image at the position
p
and
q
, respectively. The term of
Ω
is the spatial support of the weight
G
s
, and the term of
K
p
is a normalizing factor.
For lost depth region recovery, especially, to recover the lost hair region in human
modeling, a novel modeling algorithm was developed using a series of methods in-
cluding detection of the hair region, recovery of the boundary, and estimation of the
hair shape [15]. In addition, in order to fix the boundary mismatches between color
and depth information, we compensate the boundary of a human actor using a digital
image matting technique considering color and depth information at the same time.
Figure 3 shows one of the results of the depth image enhancement.
Noise removal
Hair region recovery
Raw depth image
Boundary matching
Fig. 3
Depth image enhancement
Finally, in order to calibrate measured depth data into real ones, we check the
depth of the planar image pattern within the limited space by increasing the dis-
tance (10cm) from the image pattern to the depth camera [16]. Since we already
know the camera parameters of each camera, the real depth values are calculated by
