Database Reference
In-Depth Information
Figure 10.5.
An example of turning noisy location observations into a trajectory. The
filtering model attempts to identify the most likely path that would have generated
the noisy observations given a specific movement model (linear in this case:
X
t
=
X
t−
1
+
X
t−
1
(Δ
t
)).
most traveled roadways (i.e. highways) and only expanding extra edges
when they have historically exhibited improved performance.
3. Probabilistic Models for Tracking
Processing location updates from mobile objects is a crucial compo-
nent of managing spatiotemporal data because the raw locations ob-
tained from a sensor are often noisy. Even GPS has been shown to
contain errors on the order tens or hundreds of meters [6]. Because lo-
cations at adjacent time steps are not independent (see
figure 3
), it is
possible to incorporate information about the dynamics of the mobile
object in order to improve upon its current position. This is exactly
what
tracking
accomplishes. By explicitly modeling the dependencies
between locations observed at adjacent time steps, we can filter the
raw data to produce a cleaner estimate of the object's trajectory which
accounts for noise in the sensor as well as the system dynamics (e.g. fric-
tion). Additional constraints over an object's possible movements (e.g.
road network) can also be incorporated to further improve the filtered
trajectory.
Due to the inherent uncertainty in the problem of tracking mobile ob-
jects, probabilistic models such as dynamic Bayesian networks (DBNs)
have been applied to several tracking scenarios with great success [50,
60]. In this section, we will first introduce the basic problems involved
in tracking mobile objects. We pose the tracking problem as Bayesian
