Graphics Reference
In-Depth Information
Ta b l e 7 . 1
Tracking results compared to ground truth. RMSE values are in millimetres
Sequence
# images
Object
With velocity
Without velocity
RMSE
Tracked
fraction (%)
RMSE
Tracked
fraction (%)
Industrial 1
69
Person
265
100
.
0
383
84
.
8
Table
603
100
.
0
218
69
.
7
Robot
878
95
.
5
1118
98
.
5
Industrial 2
79
Person
427
100
.
0
318
94
.
8
Table
435
100
.
0
275
100
.
0
Robot
121
98
.
7
177
96
.
1
Industrial 3
24
Person
196
100
.
0
147
100
.
0
Table
249
100
.
0
225
90
.
9
Robot
171
100
.
0
293
100
.
0
Person
247
75
.
7
352
89
.
2
Industrial 4
39
Table
270
100
.
0
245
97
.
3
Robot
91
100
.
0
200
97
.
3
Person
208
90
.
9
254
81
.
8
Industrial 5
24
Table
219
100
.
0
329
100
.
0
Robot
86
77
.
3
331
100
.
0
chosen as
H
r,
max
=
1
.
88 m,
H
r,
min
=
0
.
135 m,
H
d
=
0
.
113 m, and
H
v
=
0
.
188 m
for the industrial scenes. While the object sizes as well as the overall size of the
scenes are far different, the kernel widths merely need to be scaled by an empirical
uniform factor, such that the relative parameter values remain constant. The value of
ρ
depends on the typical velocities encountered in the scene. Hence, we set for the
tabletop scene
H
r,
max
=
4
.
14 m,
H
r,
min
=
0
.
297 m,
H
d
=
0
.
249 m,
H
v
=
0
.
414 m,
and
ρ
3200 s. For each sequence, ground truth was generated manually by mark-
ing the centre of the objects of interest in each frame, i.e. the head of the person, and
transforming them into three-dimensional coordinates using the known geometry of
the scene and the objects, e.g. the height of the person and the position of the ground
plane. The trajectories of the tracked objects are compared to the ground truth based
on the corresponding value of the root-mean-square error (RMSE) in world coordi-
nates. Figure
7.2
shows a typical tracking result achieved by our system. The results
in Table
7.1
illustrate that objects can be tracked in a stable manner at reasonable
accuracy. Using epipolar velocity as an additional feature yields a more accurate lo-
calisation result for 10 of 16 detected objects, and detection is usually possible in a
larger fraction of the frames. For four other objects the RMSE but also the detection
rate are lower when the velocity information is neglected. The system is designed
to segment the point cloud into clusters of differing velocity. As a consequence,
the proposed system works best for objects with homogeneous velocity. For exam-
=
Search WWH ::
Custom Search