Digital Signal Processing Reference
In-Depth Information
Fig. 16.3 Key points (
left
) and their 3D coordinates
With the raw matching result, we further refine them by enforcing Epipolar
constraint and perform linear triangulation to get their 3D coordinates through
precalibrated camera matrices. We set the left camera's optical center as the
world origin, and then the z coordinate is the depth of each matched key point.
For each object hypothesis that we obtained, we collect all the matched key
points inside its bounding box and select one representative for that bounding box's
depth. Here we use a simple way to select the representative point by finding the
nearest feature point around the diagonals' intersection and take the depth as the
hypothesis' depth
d
i
.
Despite its simplicity, this solution performs reasonably well compared with
other approaches such as using mean-shift to directly find the coordinates of the
mass center. The reason may be that a lot of matched point is found around the
object's boundary, and the mean-shift stops at local maxima frequently.
16.5 Utilize a Prior Height Distribution
The probability for the imaged height of a pedestrian hypothesis
Ph
i
j
is
obtained by a product of the observed height of its bounding box h
i
and a distance-
conditioned height distribution
Ph
i
j
ð
o
i
;
d
i
Þ
. The later one is obtained using depth
d
i
.
and a prior distribution of human's actual height.
Given a class-conditioned object hypothesis
o
i
, its distance
d
i
;
ð
o
i
;
d
i
Þ
and the camera's
focal length
f
which we already know from the camera's calibration, we further
model the height of an adult human using a simple Gaussian. The parameters of this
Gaussian could be estimated from statistical data. We follow [
5
] to use a mean of
1.7 m and a standard derivation of 0.085 m for the pedestrian height distribution;
therefore, we have the height distribution as
H
085
2
.
Given the prior distribution of pedestrian's actual height H, by using similarity
relation, we can represent the imaged pedestrian's height as
h
N
1
:
7
;
0
:
¼
Hf
j
D
. Because of
, h is also a simple Gaussian with 1
085
2
H
N
1
:
7
;
0
:
:
7
f
=
d
i
as mean and
0
:
085
f
=
d
i
as standard derivation. Therefore, we get
Search WWH ::
Custom Search