IEEE - Aerospace and Electronic Systems - July 2021 - 36

ASurveyofArtificial Intelligence Approaches for Target Surveillance With Radar Sensors
Figure 5.
Taxonomy of target tracking and commonly applied AI methods in this area.
TARGET TRACKING SUPPORT THROUGH AI
A typical target tracking application comprises in particular
the following two major parts: data association and target
state estimation [94], [95]. Thereby, data association
can be further divided into the subparts measurement to
measurement association, measurement to track association,
track evaluation, and multiple hypothesis tracking
(MHT). Each of these parts and subparts offers the potential
to be supported or carried out completely by AI.
Hence, we investigated literature within each part of target
tracking applying AI algorithms for improvement. In the
following, we describe the past and current state of
research in each part of target tracking, deducing the
development of AI for tracking (cf. Figure 5). A survey
from Emami et al. [96] reviews data association for target
tracking from a more in-depth perspective regarding
machine learning approaches. Hence, we refer to their survey
as a reasonable supplement to our presentation of AI
in target tracking shown in Figure 5.
DATA ASSOCIATION
MEASUREMENTTOMEASUREMENTASSOCIATION
Measurement to measurement association or often called
track initiation can be described by the first step of tracking
where multiple measurements are analyzed in order to
assert the existence ofa target and as a result create a track
36
based on matching measurements [47], [97]. Traditional
track initiation is mainly based on either rule-based, logicbased,
Hough transform-based or modified Hough transform-based
approaches. More precisely, the Hough transform
aims at detecting the peak in the parameter space for
initiating tracks. However, this is only possible when the
track to be initiated is a straight line. To overcome this
limitation, the modified Hough transform has been
invented [47].
Beginning with the work of Lian-ji et al. [98], their
knowledge reasoning system aims to combine fuzzy reasoning,
neural networks, and expert systems to improve
the track initiation. However, their approach only incorporates
a neural network like architecture and hence, avoids
a training step where labeled data are required. Instead,
their fuzzy reasoning assigns each input feature (e.g.,
velocity, radar cross section, SCR) based on its value to
an interval. Further steps include weighting and normalizing
each feature. At last, all of the weighted and normalized
confidence assignments are combined by using a rule
set of 160 rules, resulting in a final confidence for the
measurement.
Further researchers, for instance, Zhang et al. [47],
applied CNNs to initiate tracks. Although convolutional
neural networks have been developed for image processing,
applying them to radar images in terms of the view of
a radar operator would lead to a loss of information (e.g.,
amplitude, signal-to-noise ratio or range). Nevertheless,
they tried to train a CNN on the mentioned type ofimages.
To do so, they prepared images of three consecutive radar
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
