Database Reference
In-Depth Information
Figure 10.10. An example of the raw trajectory observations overlaid on a (given)
road network. Incorporating extra structure, we are able to provide a more accurate
path over which the object is likely to have traveled.
point mass is propagated through the model dynamics to predict future
states, and is then re-weighted according to the observation likelihood
given the predicted value for each. As the number of point masses tends
toward infinity, this representation will tend toward the underlying den-
sity function and thus can provide a very accurate representation of the
system. Several excellent and practical introductions to particle filtering
can be found in [20, 10, 3].
3.3 Tracking with Road Networks
Incorporating road network structure into tracking algorithms to im-
prove accuracy has been a popular area of research lately [86, 18, 56,
44]. One such example is the work by Agate and Sullivan [2] in which
the authors develop a model for tracking mobile objects that utilizes a
road network to constrain object movement and hence improve tracking
accuracy. The authors focus on tracking when both ground moving tar-
get indicator (GMTI) and high-range resolution (HRR) radar readings
are available and thus their observation likelihood models are specific
to these measurements. The dynamic model encodes the restriction of
the object to only move along the known road segments. Given the
road segment upon which the object is currently located, the probabil-
ity of the next state is a function of the structure of the network since
the object is limited to transition to adjacent roads. The state variable
Search WWH ::




Custom Search