IEEE - Aerospace and Electronic Systems - July 2021 - 16

Machine Learning Support for Radar-Based Surveillance Systems
Figure 18.
DTW barycenter and average trajectory based on AIS and ADS-B.
Figure 16.
Change of connections for military A/C in Europe, 2018.
These can be grouped into connections between two
endpoints, which can also be analyzed in terms of their
temporal development. It leads to conclusions about activities,
but more specifically, e.g., for particularly intensive
used connections, or particularly long distances. Figure 16
shows the monthwise change of military flight connections
over the year for distance flights greater than 500 km
and with more than one flight per month in Europe. Further
investigations relate to pattern recognition and anomaly
detection.
ROUTE EXTRACTION
Reliable and up-to-date geographic information is often
essential for military and civilian missions. The knowledge
available is often not sufficient, particularly in rapidly
developing or highly dynamic scenarios. Under these
circumstances, the information needs of traditional GIS
systems cannot always be met. A fast and automated generation
of activity and route patterns using trajectories
from available sensors can satisfy these needs.
ROUTE EXTRACTION BY TRAJECTORY CLUSTERING
If trajectories between two areas or points look similar,
this may be an indication that they follow the same
infrastructure or routes. So, the comparison of trajectories
enables an elegant way of traffic flow analysis. This may
be used to detect hidden routes, e.g., for smuggling or
other activities. Or the deviation from highly frequented
routes may be a reason to focus attention because ofpiracy
or kidnapping. To judge the similarity between trajectories
in a mathematical manner requires the definition of a distance
between them.
Dynamic time warping (DTW) is an algorithm that
calculates a metric for two independent time sequences
[54], [55]. It is known from the field ofspeech recognition,
but its properties also make it suitable for comparing trajectories,
see Figure 17. Having defined a distance
between trajectories, clustering (e.g., DBSCAN or related
methods) can be applied to find trajectory clusters, i.e., trajectories
that following similar motion patterns [56]-[61].
Moreover, averaging methods like the DTW-barycenter
algorithm can be used to calculate dedicated representatives
of these trajectory clusters [62], [63]. These iteratively
calculate the barycenter between all trajectories of a cluster
and are very practical for creating a route map. Figure 18
shows trajectory clusters between AIS-based ports and
ADS-B-based airports as well as the routes based on the
averaged trajectories. Figures 19 and 20 show some examples
for extracted ADS-B and AIS routes.
Figure 17.
DTW and warping path.
Figure 19.
Worldwide aircraft routes from Frankfurt a.m., Germany, based
on one-week ADS-B data.
16
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
