Database Reference
In-Depth Information
assigned to the class whose model most likely generated it. Similarly to what we
have seen with clustering, HMMs are a common choice to do it. As compared
to clustering, the problem is now simplified, since the association trajectories
↔
classes is known a priori. Behind the probabilistic framework they operate in,
HMMs essentially aggregate trajectories based on their
overall
shape, again
assuming that similar trajectories have better chances of belonging to the same
class.
The final way to classify trajectories we will see is based on a traditional
two-step approach: first extract a set of discriminative features by a preliminary
analysis of the trajectories, then use such features - that can be expressed as a
database tuple or a vector - to train any existent standard classification model
for vector/relational data.
The first step requires one to understand which characteristics of the tra-
jectories appear to better predict which class each trajectory belongs to. One
straightforward approach might consist in calculating a predefined set of mea-
sures expected to be informative enough for the task. For instance, aggregates
such as average speed of the trajectory, its length, duration, average acceleration,
and diameter of the covered region might be used. Other, more sophisticated,
solutions might instead try to extract finer aspects of the movement, tuned to
calculate only the most useful ones. A proposal of this kind can be found in lit-
erature with the name
TraClass
, which heavily relies on a trajectory-clustering
step. TraClass is based on a fundamental observation: in many cases, the features
that best discriminate trajectory classes are related to a small part of the overall
trajectory. All approaches mentioned so far, on the contrary, uniquely consider
overall characteristics - that includes HMM-based solutions, since each model
must fit whole trajectories. Single, short-duration events hidden in the long life
of a trajectory might then be lost in the process. TraClass tries to fill in the
gap by extracting a set of trajectory behaviors (which, we recall, look for local
behaviors rather than overall descriptions of full trajectories). The basic tool
adopted is trajectory segmentation and the clustering of such segments to form
movement patterns.
TraClass works at two levels: regions and trajectory segments. At the first
one, it extracts higher-level features based on the regions of space that the
trajectories visited, without using movement patterns; at the second one, lower-
level trajectory-based features are computed, using movement patterns. The
extraction phase is made more effective by evaluating the discriminative power
of the regions and patterns under construction. For instance, a frequent movement
that is performed by trajectories of all classes will be not useful for classification
(knowing that a trajectory contains such a pattern does not help in guessing the
right class to associate to it); on the contrary, a slightly less frequent pattern that
is mostly followed by trajectories of a single class is a very promising feature.
In the proposed framework, trajectory partitioning makes discriminative parts
