Z 0 50 100 150 200 240 Y
=
50
100 +
150 +
200
240
Figure 4: Houses in a suburban area: original dense range im-
age (top), detected houses in the triangular mesh (bottom).
[120x120 points, sampled at 2 m.]
From the experimental results described here we demonstrate
that we could exploit our mesh topographic coarsening to re-
duce the computational cost of recognizing objects in com-
plex 3D scenes using generic models based on nearly-planar
patches. These results show that our triangular meshes, our
topographic mesh coarsening, together with our generic sym-
bolic object models, can be successfully used in the problem
of 3D object recognition from real sensory data. Such data,
gathered by remote sensing and photogrammetry techniques,
is typically imperfect. Our approach yields accurate results
while reducing the recognition complexity. This is possible
through intelligent data compression and filtering by exploit-
ing the triangular mesh topographic coarsening. If fast ob-
ject recognition and identification is desired, the coarsened
meshes are adequate. Their results can be used to initiate
objects seeds in higher resolution meshes. For accurate ob-
ject recognition, the high resolution meshes are used. If more
accuracy is desired, the obtained results can serve as very
good initial guesses for techniques using combined range and
intensity information. It is our future goal to investigate this
combination for precise 3D model synthesis from sensory data
and reverse engineering of CAD models.
EEE m
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
# | Patches | Outside Floor C.O.G.
surface area
1 51 618.9 409.9 (60.8,110.8,103.3)
2 59 825.7 512.5 (111.3,64.5,104.1)
3 57 818.9 511.6 (86.7,103.1,104.7)
4 76 835.0 514.0 (65.0,137.6,104.7)
5 47 860.1 535.1 (42.4,176.8,105.2)
6 78 929.3 596.1 (97.1,164.9,105.1)
7 80 932.6 634.2 (125.0,180.5,105.0)
8 90 895.4 590.0 (150.8,198.7,105.0)
9 85 821.3 534.7 (128.0,220.5,105.1)
10 77 915.7 604.4 (183.8,217.1,105.1)
11 63 885.7 588.4 (209.3,179.0,105.3)
12 58 631.2 436.3 (229.5,141.9,105.4)
13 98 609.7 400.3 (212.1,101.7,103.6)
14 91 617.3 415.9 (181.8,83.3,103.7)
15 165 1510.9 997.7 (149.7,79.7,104.4)
16 218 1599.0 1065.8 | (183.7,111.2,104.5)
17 81 533.6 366.2 (218.0,9.1,101.4)
Table 5: Extracted properties of houses in a suburban area.
# | Patches | Outside | Floor C.O.G.
: surface area
10 83 915.9 603.5 | (183.8,217.1,105.1)
11 62 884.3 592.1 | (209.3,178.9,105.3)
Figure 6: Extraction of houses: close-up in the original tri-
angular mesh (top mesh) and the same detected houses in a
coarser mesh with 82% less vertices (bottom mesh).
6 CONCLUSIONS AND FUTURE WORK
We presented a technique for the compression of huge sets of
3D sensory data. Unlike more general brute force data com-
pression approaches, our technique is specifically concerned
with the preservation of topographic features of the 3D scenes
for the particular problem of shape recognition. For this pur-
pose, we use an irregular triangular mesh representation of
the scenes. From these, we identify and extract important
topographic surface features which are preserved in coarser
meshes representing the same scenes albeit at lower reso-
lution levels. The topographic coarsening can be repeated
several times with controlled steps preserving the same fea-
tures. This method is both more computationally attractive
and more informed than mesh refinement techniques. It sim-
plifies the analysis of massive amounts of sensory range data
that would, otherwise, be difficult to use.
At any resolution level, the topographic features can be used
to segment the triangular mesh representing the scene into
nearly-plan:
models of .
sections of
est. Thes
to the coa
tions of th
sensing an
ing (DTM)
navigation,
tion systerr
In this con
of interacti
ening for tl
Although r
object moc
results are '
ibility of tl
intelligent
able range
best done
tional infor
phases by «
data which
(Boender «
and P
from coi
Design,
(Chen & S
Surface
tion. Cc
(Cheng et
1988. A
formati
nition, |
(Chew, 19
generat
sium or
