Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 125

Fig. 2.14 Reprojected sparse
scene flow field

Fig. 2.15 Left : Ellipsoid
model used for tracking
arbitrary objects or object
parts in the three-dimensional
point cloud.
Right : Reprojection of four
tracked objects

determine the object motion in the image sequence. The optical flow field is com-
bined with the three-dimensional points from the blockmatching stereo algorithm
by Franke and Joos ( 2000 ) to obtain the scene flow field (cf. Fig. 2.14 ). The velocity
component parallel to the depth axis is not computed.
At each time step, a graph-based clustering stage extracts all moving objects
from the scene flow field, essentially separating moving objects from the (stationary
or differently moving) background. The computed clusters are approximated as el-
lipsoids, and the three-dimensional ellipsoid pose T (c x ,c y ,c z ,β) T

=

is determined

(c x ,c y ,c z ) T of the ellipsoid and the rotation an-
gle β around the depth axis are part of the pose vector T , since we assume that the
approximated objects are parallel to the image plane. The three-dimensional pose
update of all tracked objects is based on a two-stage three-dimensional extension of
the mean-shift algorithm according to Cheng ( 1995 ), Comaniciu et al. ( 2000 ). If a
tracked object is not moving for more than five time steps it is deleted.

=

(cf. Fig. 2.15 ). Only the centre c

2.3.2.2 Target Model

The target model q (id) with the object index id is computed based on the first three-

dimensional ellipsoid pose T (id) and is updated at every time step. It consists of
a one-dimensional histogram of greyscale values. To compute the histogram

q (id)
we place a grid on the surface of the ellipsoid, the resolution of which is equal