Database Reference
In-Depth Information
Figure
8.1
d demonstrates the space-time cube (STC) display where two
dimensions represent the space and the third dimension the time. The time axis
is oriented from the bottom of the cube, where the base map is shown, to the
top. When all trajectories are included in the STC, the view is illegible due to
overplotting. In our example, the STC shows 63 trajectories selected by means
of a spatial filter (Figure
8.1
e). For the filter, we have outlined on the map two
areas to the northwest and southeast of the city and set the filter so that only the
trajectories that visited both areas in the given order are visible. There are also
many other interactive techniques for data querying and filtering, for example,
the ones suggested by
Bouvier and Oates
(
2008
)and
Guo et al.
(
2011
).
8.2.2 Clustering of Trajectories
Clustering is a popular technique used in visual analytics for handling large
amounts of data. Clustering should not be considered as a standalone analysis
method whose outcomes can be immediately used for whatever purposes. An
essential part of the analysis is interpretation of the clusters by a human analyst;
only in this way do they acquire meaning and value. To enable the interpreta-
tion, the results of clustering need to be appropriately presented to the analyst.
Visual and interactive techniques play a key role here. Visual analytics usually
does not invent new clustering methods but wraps existing ones in interactive
visual interfaces supporting not only inspection and interpretation but often also
interactive refinement of clustering results.
Trajectories of moving objects are quite complex spatio-temporal constructs.
Their potentially relevant characteristics include the geometric shape of the path,
its position in space, the life span, and the dynamics, that is, the way in which the
spatial location, speed, direction and other point-related attributes of the move-
ment change over time. Clustering of trajectories requires appropriate distance
(dissimilarity) functions that can properly deal with these nontrivial properties.
However, creating a single function accounting for all properties would not be
reasonable. On the one hand, not all characteristics of trajectories may be simul-
taneously relevant in practical analysis tasks. On the other hand, clusters pro-
duced by means of such a universal function would be very difficult to interpret.
A more reasonable approach is to give the analyst a set of relatively simple
distance functions dealing with different properties of trajectories and provide
the possibility to combine them in the process of analysis. The simplest and
most intuitive way is to do the analysis in a sequence of steps. In each step,
clustering with a single distance function is applied either to the whole set of
trajectories or to one or more of the clusters obtained in the preceding steps. If
the purpose and work principle of each distance function is clear to the analyst,
the clusters obtained in each step are easy to interpret by tracking the history
of their derivation. Step by step, the analyst progressively refines his or her
