A. Human Leg Segmentation
The raw 3D point cloud from the depth sensor contains redundant information in the dynamic and uncertain environment.
For instance, the background and floor may cause failure to detect the users lower limbs. Therefore, we apply two filters to
extract the foreground based on the empirical principles. First,
the 3D points are filtered as the background if their depth values
are greater than 120 cm, since the foreground users will move
most likely within the region which is a half circle centered at
the walker with radius equal to 120 cm. Second, we adopt the
Random Sample Consensus (RANSAC) algorithm to subtract
the floor plane. The RANSAC algorithm is an iterative method
to estimate the parameter of the randomly generated 3D plane
model from the n observed raw 3D point cloud. To capture the
floor, we try to obtain the generated plane model which has the
maximum inliers (number of 3D points) which are closest with
distance to the plane model. After the estimated plane model is
obtained and subtracted, the left 3D data point set is extracted
and considered as foreground. Note that we assume the floor is
flat and clear, and thus floor capturing is proposed only when
the system initializes. Finally, the foreground 3D data are con-
sidered in association with the users lower limbs.
For separating the left and the right legs, Vertical Projection
Histogram (VPH) facilitates us to locate the center of the lower
limb. For deriving VPH in our work, the depth image is first
convertedtoabinaryimage,whichisthenmappedtoaone-di-
mensional histogram, where the value of the histograms corre-
spond to the sum of white pixels along the y-axis. Referring to
Fig. 2, where Fig. 2(a) is the binary image converted from depth
image, and is actually the location of the valley (middle)
between the first and the second peaks, as shown in Fig. 2(b),
namely, vertical projection histogramof the binary image. Thus,
the x-coordinate of center defines a vertical line in the bi-
nary image of lower limbs which clearly separates the left leg
and the right leg, as shown in Fig. 2(a).