Biomedical Engineering Reference
In-Depth Information
original point is supposed to have moved). Given a suitable threshold
ʘ
M
, the
discrepancy measure
X
2
Mp
ðÞﾼ
ð
I
t
p
ð
þ
q
Þ
I
tþ
1
p
ð
þ
v
þ
q
Þ
Þ
ð
6
:
37
Þ
q
∈
W
is then used to evaluate whether tracking was correctly performed (
M
(
p
)
< ʘ
M
)or
not (
M
(
p
)
ʘ
M
)
.
It is thus interesting to analyze empirically when the Lucas-
Kanade algorithm tends to fail and why. Conclusions from this investigation will
lead directly to a method to perform independent motion detection. The main
empirical circumstances in which errors in the evaluation process of the optical
flow arise are three:
• Speed. The instantaneous velocity of the point is too large with respect to the
window where motion is being considered. Hence, the computation of temporal
derivatives is difficult.
• Rotations. The motion around the point has a strong rotational component and
thus, even locally, the assumption regarding the similarity of velocities fails.
• Occlusions. The point is occluded by another entity and obviously it is impos-
sible to track it in the subsequent frame.
Tracking failures caused by high punctual speed depend exclusively on the scale
of the neighborhood where optical flow is computed. This issue is usually solved by
the so-called pyramidal approach which applies the Lucas-Kanade method at
multiple image scales. This allows evaluating iteratively larger velocities first and
then smaller ones. Instead we determined empirically that when rotations cause
failures in the tracking process, this is often a consequence of a movement inde-
pendent from that of the observer. The third situation in which Lucas-Kanade fails,
is caused by occlusions. In this context the main role in determining whether optical
flow has been successfully computed is played by the speed at which such occlusion
takes place.
We therefore look for points where tracking is likely to fail as soon as one of the
conditions discussed is met, i.e., flow inconsistencies due to rotations or occlusions.
In detail, we run Lucas-Kanade over a uniform grid on the image, perform the
comparison indicated in (
6.37
), and then filter for false positives (isolated failures).
The results are a set of independent moving blobs.
We tested the method both in controlled situations (a small robotic device
moving linearly in front of the iCub) and, more generally, in tracking people and
other moving objects in the laboratory. Figure
6.17
shows results of tracking with
both stationary and moving cameras (therefore without and with ego-motion,
respectively). In the configuration considered, a linear speed of 10 cm/s corresponds
to one pixel per frame in a 30 frames-per-second (fps) acquisition. Experiments
were conducted up to 100 cm/s and with the iCub head adding movement up to
40 deg/s.
The sequence of images in Fig.
6.18
is an example of a more naturalistic
tracking. In spite of the complexity of the background, it is evident from the images
Search WWH ::
Custom Search