. Istanbul 2004
|
|
al components:
cloud
tudy applied to
[he dataset was
'rrestrial laser
of Gävle. The
IS institute has
spatial objects
extracted. Thus
ith the size of
ned in the way
section 2. The
some values of
] each of the
the point cloud
1 [1]; Then the
adjusted their
according to
or the training
pre-determined
finished with a
etain the initial
he component
Jor transitions
oser than those
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Table 2: Parameter settings for the SOM training
Parameter Value
Size (m) 1160
Dimensionality 2
Neighbourhood Gaussian
Learning rate ( Q )
a(t)=a, (1+100t/7)
Initial learning rate (@,)
0.5 for the coarse period
0.05 for the fine period
Training length in epochs
(T)
0.51 epochs for the coarse
period
2.05 epochs for the fine
period
Initial neighbourhood
radius ( o.)
20
Final neighbourhood radius
5 for the coarse period
1 for the fine period
Figure 5: The original point cloud consisting of 9072 points
Figure 6: Component visualizations of the SOM: (a) x
coordinate, (b) y coordinate, (c) z coordinate, and (d) intensity
à
© 286
©
Figure 7: Five clusters detected from Umatrix of the SOM
In order to detect different spatial objects, we derive a unified
distance matrix (U-matrix) between the adjacent neurons
(Ultsch and Siemon 1990). Figure 7 illustrates the distance from
each neuron to its neighbouring neurons. We can note those
neurons that are surrounded by darker colours tend to be
clusters. We tried to select those points that best match to the
clusters in the SOM, and it ends up with 5 meaningful clusters
as indicated in figure 7. The cluster 0 match to the stones quite
well, while the rest four clusters match to clay-road. Figure 8
illustrates those points associated with clusters 1-4 (a) and
points representing clay road (b). Visual inspection suggests the
model is a useful tool for filtering scanning datasets. In the
meantime, cautious should be taken for the model, as other
spatial objects such as spruce and ground are not clearly shown
with the clusters in the umatrix of the SOM. This suggests
further work is needed with the training process, probably by
introduction of a weight among xyz coordinates and return
intensity.
(a) (b)
Figure 8: Points associated with detected clusters 1-4 (a) and
points representing clay road
5. CONCLUSIONS
This paper explores a new approach to filtering laser-scanning
dataset for the extraction of spatial objects based unsupervised