![](bg4.png)
(a) (b) (c) (d) (e)
Figure 6: A series of range images of a complex object was collated. (a) A scene image from the sequence, (b) the 3D blob
map recognizes that two objects were moved, their states will be updated, (c-e) ICP fitting of blob valuated points with the
3D object model, see Table 3.
poses, the set-point values at each spot and the esti-
mated poses corresponding to a single trial. Consid-
ering that in this experiment the camera positions are
biased by a human factor regarding the manual op-
eration of the robot, in Table 1 we also report as a
reference, the pose values that were obtained by mov-
ing the robot and keeping the same scene static. Con-
cerning the dynamic scene, since the estimation of a
transformation depends on the quantity as well as the
quality of the points, we also include in the table the
mean squared error (MSE) of each transformation as
a measured of the reliability or precision of the esti-
mation (MSE Tr), and in order to have a statistics of
the accuracy of the process, we show the MSE of the
Euclidean distance (MSE Eu) between the estimated
poses and the reference positions corresponding to 40
measurements in each pose. Because we aim at build-
ing a 3D map for robot interaction, only the objects
that lie closer to the stereo rig, inside a radius of 2m
from the cameras, are registered into the map, in our
example this corresponds to the first three boxes in
Fig. 5(a). This figure shows a textured 3D image at
the first state of the scene. Fig. 5(b) shows by color all
the static registrations along with the estimated cam-
era pose frames for each step of the sequence. The
(a) (b) (c)
Figure 5: Visual motion estimations. (a) Textured 3D image
of the scene at its initial state, (b) robot-pose frames and
static registrations, (c) detection of motion in two mapped
objects.
last two columns of Table 1 indicate that our motion
estimation system is more precise than accurate, i.e.,
we can not certainly determine the absolute pose of
each mapped object in the world but rather determine
that the geometric relations in the map measured ei-
ther between any two of them or locally to a single
blob are the closest values to the actual ones. In Ta-
Table 1: Results of the ego-motion estimation.
Pose (X[cm], Y[cm], angle[
◦
])
Set Static Dynamic MSE MSE
Point Scene Scene Tr(x
−3
) Eu
1
(0,0,0) (0,0,0) (0,0,0) —– —-
2
(40,0,10) (41.21,-0,9.7) (42.25,0,10.46) 1.019 6.9104
3
(0,-20,0) (-0,-19.54,0.88) (0,-20.12,0.15) 1.029 1.5678
4
(-45,0,15) (-46.19,-0,14.71) (-45.16,-0,14.0) 0.119 3.3030
5
(-24,0,14) (-23.62,-0,14,32) (-24.72,-2,13.78) 0.282 5.9464
6
(0,10,0) (-0,9.65.4,1.1) (1.1,8.32,0.44) 0.688 3.6485
7
(20,0,10) (20,0,9.7) (22.22,-0,11.35) 0.316 5.8045
Table 2: Results of the object-motion estimation.
Pose (X[cm], Y[cm], angle[
◦
])
Set Point Cereal Box Set Point Pop Corn Box
1
(0,120,0) (0,117,1.52) (-33,115,10) (-33.67,114,8.1)
2
(0,110,0) (0,106.3,0.28) (-33,115,10) (33.84,113.42,9.52)
3
(0,130,0) (0,126.3,1.92) (-33,115,10) (33.78,113.01,8.78)
4
(0,130,0) (0,126.52,2.39) (-27,107,10) (-27.42,104.95,7.40)
5
(-33,120,10) (-32.77,118.28,4.81) (0,100,0) (0,97.15,1.16)
6
(-33,120,10) (-32.74,118.27,5.46) (0,94,0) (1.38,92.84,1.24)
7
(-33,120,10) (-32.83,118.22,6.04) (0,94,20) (0,92.42,22.47)
ble 2 we report the estimated pose values that were
obtained with the moved objects. We now present the
results of collating a sequence of range data of a non-
simple geometric model in Fig. 6(a). The object and
the cameras were moved to different spots during the
sequence. In order to show how precise the different
sets of valuated points of a 3D-blob image the actual
mapped object, we present the results of fitting by ICP
each point set to a 3D model of the mapped object.The
confidence value assignments ranges from 0-7. Some
fittings can be visually observed in Fig. 6(c-e). We
also present the magnitude of the matrix rotation,
Eq. 6, that was needed for each fitting: {valuated pts}
→ {model pts}. Since the object-model frame and
the valuated-point frame were aligned before running
ICP, this value will give us a measure of the amount
of correction that was needed to obtain a correspond-
ing RMS error value of this fitting. The results are
shown in Table 3. Although the amount of correction
Dynamic3DMapping-VisualEstimationofIndependentMotionsfor3DStructuresinDynamicEnvironments
405