D'Apuzzo, Nicola
The least squares matching algorithm is then applied at that
position and the result can be considered the exact position of the
tracked point in the new frame (step 3).
Like explained before, this process is performed in parallel for the
three images of the different views. To test the individual results in
the three images, a spatial LSM is then executed at the positions
resulting from the temporal L$Ms (see flowchart in figure 7) and if
no significant differences occur between the two matches, the
point is considered tracked and the process can continue to the
FRAME i-1 FRAME i FRAME i+1
| ZOOM
next time step. If instead the differences are too large, the process EN mar com
goes back to step (2) by searching the value of best cross a STAKES
correlation in a bigger region around the predicted position. If the F gure 8. Tracking in image space: temporal LSM
result is rejected again, then the tracking process stops. 1s applied at the position of best cross correlation
The result of the tracking process are the coordinates of a point in
the three images through the sequence, thus the 3-D trajectory is determined by computing the 3-D coordinates of the
point for each time step by forward ray intersection. Velocities and accelerations are also computed.
This way of tracking points may produce errors which cannot be easily detected. In fact, the only control of the tracking
result is the test executed between the spatial LSM results and the temporal LSM results. There is no 3-D control of the
trajectories. Thus, false trajectories can be generated even if the tracking results seems good. A new test has to be
integrated in the process to detect the false trajectories.
This can be achieved by tracking part of surfaces and not only single points. In this case, the result of the tracking process
can be considered as a vector field of trajectories, which can be checked for consistency and local uniformity. Indeed,
since the human body can be considered as an articulated moving object, the resulting vector field of trajectories must be
locally uniform, i.e. the velocity vector must be nearly constant in sufficiently small regions at a particular time.
Therefore, filters can be defined to check these properties. The next paragraph describes the approach.
2.3.2 Surface tracking. Tracking surface parts means track simultaneously points
belonging to a common surface. Practically, the tracking process depicted in the previous
paragraph, is applied to all the points matched on the surface of the first frames. With this
approach, a new problem has to be considered: during the sequence, some surface parts can
get lost by occlusion and new parts of surface can appear (e.g. the legs which occlude each
other during a walk sequence). For this reason, a new functionality has to be integrated in
the tracking process. Before proceeding to the next time step, the data resulting from the
tracking process is checked for density (see flowchart in Figure 9). This operation is
executed with a defined frequency (which can be for example every two frames). In the
regions of low density (determined by a threshold), new points are integrated in the process,
TRACKING PROCESS
FOR ALL POINTS
INCLUDE
NEW POINTS
Figure 9. Flowchart of the
so that new appearing surface parts are also tracked. The new points come from data Surface tracking process
previously computed (surface measurement of the body for each frame).
Figure 10 shows 6 frames of a walking sequence and the results of the surface tracking process.
Figure 10. Left: 6 frames of a walk sequence (upper left to lower right)
Right: Tracked points displayed in image space for the 6 frames
168 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000.
AS €
can
clea
com
are «
Figi
(pos
now
unif
are
trun
con:
acce
filte
mot
righ
divi
time
vect
are
mea
diffe
con:
or re
As
with
traje
the
traje
filte:
2.3.
LSA
usec
sequ
cam
sing
proc
Fig
trac]
key
seqt
acqu
exar
LSN
Sim]
the 1
