IEEE - Aerospace and Electronic Systems - July 2021 - 8

Feature Article:
DOI. No. 10.1109/MAES.2020.3001966
Machine Learning Support for Radar-Based
Surveillance Systems
Kaeye D€astner , Steffen Haaga , Bastian von Haßler zu Roseneckh-K€ohler ,
Camilla Mohrdieck
, Felix Opitz , and Elke Schmid , AIRBUS, 89077 Ulm,
Germany
INTRODUCTION
Radar-based surveillance systems operate in different
environments to observe and track objects in the same or
different environments and present a unique operational
picture to operators or other connected systems, e.g., tactical
data links. There are airborne radar systems for ground
and maritime moving target indication (GMTI/MMTI),
airborne air surveillance systems, e.g., airborne early
warning and control systems (AWACS) such as groundand
ship-based air defense systems, civil air traffic control
(ATC), and border and coastal surveillance systems. They
all deliver object trajectories independent of any cooperation
and enable the extraction of mid- and long-term trajectories
ofany kind of objects.
On the other hand, big data and machine learning
(including deep learning) are currently among the most
influencing technologies. Companies that deal with a lot of
commercial data have gathered a lot of information and
therefore developed special techniques to deal with this
enormous amount ofdata-simply called big data. Machine
and deep learning turned out to be a technique that is suitable
for analyzing, grouping, classifying, detecting anomalies,
and predicting behavior. The so-called supervised machine
learning requires labeled data, which means that the data
used for training has a well-known and defined label, e.g.,
state, color, name, etc. The goal is then to predict the label of
data, which the system has never seen before.
For air and maritime surveillance, labeled training
data are available via automatic dependent surveillance-
Authors' current address: K. D€astner, S. Haaga, B. Von
Haßler Zu Roseneckh-K€ohler, C. Mohrdieck, F. Opitz,
E. Schmid, AIRBUS, W€orthstraße 85, 89077 Ulm,
Germany (e-mail: kaeye.daestner@airbus.com).
Manuscript received February 3, 2020, revised May 18,
2020, May 25, 2020, and ready for publication June 10,
2020.
Review handled by Dietrich Fraenken.
0885-8985/21/$26.00 ß 2021 IEEE
8
broadcast (ADS-B) and automatic information system
(AIS). The data are commercially available and cover all
areas in the world with a large history. Actually, the data
are pretty similar. They contain the position and speed of
an object at a measured point in time and form a trajectory
using a unique identifier. They only differ in the context
data, which, however, can be used for labeling in supervised
machine learning.
The real challenge is to put the trajectories in the right
context and generate situational awareness in order to correctly
estimate the intentions of the objects being tracked.
Activity-based intelligence and the determination of pattern
of life are new opportunities for surveillance systems
but even modern surveillance systems are not able to take
a real advantage of the gathered data. To address trajectories,
various techniques are useful: Trajectories are partitioned
into specific segments of interest using cluster
algorithms. This helps to decode their pattern of life based
on unsupervised machine learning. Modern hierarchical
cluster methods allow the extraction of highly frequented
routes, predictive analytics, and the identification of
anomalous behavior.
In this article, we present various methods and
approaches that show how big trajectory data and new
technologies can be used to improve surveillance systems.
This requires specific system architectures as well as a
completely new software and hardware landscape. In the
further course, the generation for heat maps, clustering,
route extraction, and classification algorithms are
explained and how they can be used. Different methods
and use cases for anomaly detection are shown and finally
how all these processes can be combined and integrated in
the next generation of surveillance and intelligence systems.
Since a variety of topics are covered, we cannot go
into the machine learning algorithms in great detail. These
methods are generally very popular because big datadriven
methods do not require detailed modeling knowledge
and are, therefore, easier to use as compared with
classic algorithms. Interested readers are, therefore, recommended
to use the given references.
IEEE A&E SYSTEMS MAGAZINE
JULY 2021
IEEE - Aerospace and Electronic Systems - July 2021

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