IEEE - Aerospace and Electronic Systems - July 2021 - 17

D€astner et al.
Figure 21.
Principle for the assignment between routes and a radar track
object.
Figure 20.
Ship routes in the Baltic Sea based on one-month AIS data.
ROUTE ASSIGNMENT, PREDICTION, AND ANOMALY
DETECTION
Extracted routes are an enabling technology and allow
numerous applications. The most significant are long-term
kinematical predictions and anomaly detection [64].
The use of a route map converted from a geographical
position into a tiled raster index representation enables
a quick assignment of any position, e.g., a radar
measurement to an individual route, see Figure 21. In
reality, the route picture is denser so that not only one
route matches to a track. So a comparison between current
track heading and route direction and including the
track history gives the possible routes the track might
travel on. The route strength, i.e., how many individual
trajectories have contributed to this route, should also
be taken into account.
A track that cannot be assigned to any route may be
classified as anomaly, whereby predictions about destination,
arrival time, and source can be made for an assigned
track. In the case of multiple assignments, the strength
and difference between the heading vector and the route
direction can determine a priority or weight.
With the labels of the AIS or ADS-B messages, e.g.,
platform type, routes can also be classified. Tracks that
are assigned to a route in real time can inherit the classification
of the assigned route.
CLASSIFICATION
Modern surveillance systems already use a number of sensors
that help to classify or even identify objects. On the
one hand, these are sensors for cooperating targets, such
as IFF, AIS, ADS-B, etc., and second, for noncooperating
targets, such as ESMs. Combined with radar measurements,
this information can contribute to target classification.
Furthermore, certain kinematic features, such as
JULY 2021
altitude and speed, or the direction of the target can provide
further information about a classification. Nevertheless,
the results are always subject to uncertainty and
further sources of information would be desirable. The use
of big data and machine learning opens up new opportunities
here.
Labeled data, e.g., AIS or ADS-B data, are typically
used for supervised machine learning, especially classification
topics. Each trajectory point has one or more labels,
e.g., the ship type and navigation status for AIS and military
flag and aircraft type for ADS-B [8], which can be
learned to be associated with given features of the trajectory.
These features are derived from the complete or windowed
trajectories and are based on the object's positional
and kinematical history. After training of a supervised
machine learning algorithm in the serving layer, labels are
predicted based on the positional or kinematical features
of a trajectory. This is schematically depicted in Figure 22
that shows a classification algorithm that is trained with
given labels. The trained model is saved and serves as
autonomous classifier in future applications where unlabeled
data (e.g., radar data) are classified (i.e., the corresponding
labels are predicted).
For classification tasks, a lot of classical machine
learning algorithms are known, such as: Logistic regression,
supported vector machine, decision trees, ensembles
of decision trees, e.g., random forest, gradient
boost tree, and multilayer perceptron, which is the simplest
form of a neural network [45], [46], [65]-[69].
The choice of a classification algorithm depends on the
implementation constraints, e.g., integration environment,
learning time, number of labels, etc., as well as
on the data itself. Classification of objects can be used
in following different ways.
Online: Flag objects for real-time surveillance, e.g.,
directly or for further processing in DBN or other
classifier, which combine different methods.
Offline: Flag objects in big data and use them for
new labels, e.g., classify routes, areas of interest,
etc. This in turn canbe usedtotrain anew
classifier.
IEEE A&E SYSTEMS MAGAZINE
17
IEEE - Aerospace and Electronic Systems - July 2021

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - July 2021
