Biomedical Engineering Reference
In-Depth Information
Fig. 2 Left to Right: anterior view of graph growing, extracted leg parts, lateral view graph
growing, and extracted body parts. The graphs are displayed over the mean shift segmentation
results. The
crosses
show the initially labeled body parts, the
crosses
segments are the skeletons
found in anterior and lateral views
4
Jumping Pattern Analysis
Six joints are used in the anterior view, i.e., right/left hip, knee, and ankle. In the lateral
view, key joints include the centroid of head, hip, knee, and ankle. The location of knee
is approximated by the middle point of the nearby sticks' end points. In order to get a
robust joint estimation, the
x
and
y
components of each joint sequence are smoothed
separately
based on a priori observation that the motion of different joints shares
similar patterns in
y
coordinate such as peak and turning points, even in noisy
measurements when the camera is shaken slightly. Before smoothing, outliers are
rejected using a typical median filter. Then, the smoothing spline [
13
]isutilizedto
smooth each joint sequence.
The key turning points are determined as follows: (a) compute the curvatures
k
of the
y
sequence for each joint, (b) find sharp turning points
p
s
that are both local
maxima of curvature and greater than a threshold
t
k
¼ gs
(
k
), where
s
(
k
) is the
standard deviation of the curvature, and (c) identify the three key turning points, i.
e., the sharp turning point that is closest to the global minimum of the original
sequence, and the other two to both sides of the first one. Accordingly, the sequence
can be divided into three sections: preparing, jumping, and landing. Figure
3
shows
an example of key turning point detection. Here, the initial contact of the foot with
the ground during landing corresponds to the right turning point in Fig.
3
.
5 Experimental Results
The proposed approach has been tested on videos captured in different
environments with different qualities. In mean shift segmentation, we set the spatial
bandwidth as 10, range bandwidth as 10, and the minimum allowed area size
as 50 pixels. In joints tracking, the parameters were set to
b ¼
1.2 and
g ¼
1.5.
Those parameters were kept the same in all test videos.
