Digital Signal Processing Reference
In-Depth Information
N
p
N
m
i
i
X
≈
w
X
i
,
X
≈
w
X
i
(10)
p
m
k
k
k
k
i
=
1
i
=
1
The Bhattacharyya likelihood between target candidate position
X
,
X
and the
target position can be calculated. If the likelihood in
X
is larger than
X
,
randomly remove the
low-ranking N
m
particles in the N
p
particles and add the
N
m
particles to the N
p
particles. Otherwise remain the primary N
p
particles.
5)
Resample
Compute an estimate of the effective number of particles as N
eff
N
p
i
k
2
N
=
1/
w
)
(11)
eff
i
=
1
If N
eff
<N
thr
, copy the particles with large weight and remove the particles with
small weight.
Let
i
k
=1/ N
p
, i=1... N
p
, we can obtain the new particles set {
X
j
k
i
k
w
,
w
,
j=1….N
p
} as the initial particles at time k+1, and then go back to step 2).
4
Experiment Results
The experiments are conducted on three video clips,
face.avi
,
hand.avi
, and
soccer.avi
. The former two videos are captured via a normal USB webcam at 25fps
with the resolution of 240×320 pixels. The last one is downloaded from public
database with the resolution of 480×360 pixels. The algorithms are implemented with
OpenCv2.0 and vs. 2008 on a desktop with 2G RAM and Core2 2.4GHz CPU. The
MS algorithm from Ref. 2, and the PF algorithm from Ref. 3 are used for the
comparison.
To compare the efficiency, the traditional PF method can achieve a running speed
of 75fps on the face and hand datasets. In comparison, the proposed method is more
efficient with a speed of 125fps. Since the particle number of 100 is required by PF,
only 70 particles are needed in the proposed method.
Fig. 4(a) shows the tracking results under illumination changing. The 1
st
row shows
the results by our method, the 2
nd
and 3
rd
rows are the results by MS and PF
algorithms respectively. As the result shows, when illumination suddenly changed
from bright to dim, the proposed algorithm can still locate the eyes precisely. But the
MS and PF algorithms are seriously affected. It strong robustness to illumination
change mainly comes from the use of LBP feature in the proposed algorithm.
Fig. 4(b) evaluates the performance of different algorithms with occlusions. The 1
st
row shows the result by the proposed algorithm, and the 2
nd
and 3
rd
rows show the
results by PF and MS algorithms respectively. When a partial of the hand is occluded,
the PF and our method still work but the MS algorithm failed. When occlusion is
seriously, both PF and MS algorithms loss the tracking target.
The dataset of
soccer.avi
is used to evaluate the performance of different
algorithms under the target with continuous and huge appearance change. The
