Ground targets can be detected by multiple classes of sources in
military surveillance. There are two main challenges for the
acquisition of ground situation picture from data collected by
multiple sources. First, different sources provide different
information that describes military entities at different
granularities and accuracies. This makes processing of data in one
unified tracker difficult. Secondly, the data update rates of
these sources vary, some update rates could be very low (such as
hours), leading to greater difficulty for data association.
This paper presents our attempt in multi-source ground target
tracking, taking the above two issues into consideration. Targets
are tracked in groups, and multiple trackers are designed so that
data of different granularities are processed by the respective
trackers. Tracks from these trackers are then correlated to form
the common picture. Two strategies are proposed to handle the
problem of varying data update rate. The first strategy is to
exploit different approaches to calculate the beliefs of data
association according to update rates. When update rate is high,
the belief is calculated by a distance function based on estimated
kinematical states. When update rate is low, the belief of data
association is computed using Bayesian network. Bayesian network
infers the beliefs based on observed information and domain
knowledge. The second strategy is to exploit the complementary
information in different trackers to improve data association. The
first step is to find the correlation among tracks from different
trackers. This track-track correlation information is fed back to
modify the beliefs of data associations in the tracks.
Experiments demonstrated that such combination of multi-source
information not only produces more complete ground picture, but
also helps to improve the data association accuracy in the
respective trackers.
|