IEEE - Aerospace and Electronic Systems - July 2021 - 38

ASurveyofArtificial Intelligence Approaches for Target Surveillance With Radar Sensors
the update rule in a competitive winner-take-all scheme
[102]. By this means, they could avoid determination of
specific weighting parameters, necessary to initialize the
Hopfield network.
TRACKEVALUATION
The purpose of track evaluation is to determine if a track
and its estimates are valid or not. In order to identify false
tracks in an automated manner without using AI, research
relies on the track characteristics over the lifetime of a
track. For instance, features as the track life time, proximity,
and dynamics are exploited within a rule-based system
to find false track [104]. In the context of AI, we identified
two papers aiming at the development offalse track identification
[28], [37]. In this respect, Dai et al. [28] defined
false tracks as unreal targets arising, for example, from
clutter. Thus, they further the goal of suppressing false
tracks as soon as possible. To do so, they rely on the logistic
regression algorithm using the features speed and
course stability as well as the number of plots in the correlation
gate. In contrast, Ghadaki et al. [37] focused on
research in the area of air surveillance radars and therefore,
consider false tracks as all nonaircraft tracks. To
identify this type of false tracks, they made use of a nonlinear
SVM classifiers trained on flight path trajectory,
radar cross section, and velocity related features. For
assessing the performance of SVM approach, they had
access to real-world data recorded over 364 scans while
including 557 aircraft and 542 nonaircraft tracks. The
labels of the data (aircraft/non-aircraft) are based on the
presence ofsecondary surveillance radar data. On average,
their SVM approach was able to distinguish the two classes
for 97.72% of all tracks. In line with our own investigations,
these results demonstrate a high potential in AI
based false track identification. Additionally, our own
findings revealed that data labeled based on received
information from secondary surveillance radar is already
satisfactory separable with AI approaches on plot basis. In
detail, we used seven hours of recordings from an air surveillance
radar and labeled plots similarly to [37] by the
presence of secondary surveillance radar information. As
a result, we obtained 3841 357 plots belonging to the nonaircraft
class as well as 205 200 plots belonging to the aircraft
class. Due to the high amount of nonaircraft samples,
we applied a random undersampling procedure, inspired
by [26]. With the aim of favoring the aircraft class, we
reached a 2:1 ratio for the distribution of aircraft and nonaircraft
samples. On this basis, we conducted a five-fold
cross validation comparing the performances of Naı¨ve
Bayes, kNN, decision tree, random forest, and MLP classifiers.
While all of the classifiers were able to correctly distinguish
the two classes for over 90% of the plots, the
decision tree outperformed the other classifiers with an
accuracy of over 96%. Moreover, the correct classification
38
rate for both classes diverged by less than 3% for all
classifiers.
MULTIPLEHYPOTHESISTRACKING
For tracking multiple targets, MHT became a well-established
approach particularly when data association is
impeded by a high target and clutter density as well as by
a low probability of detection. MHT can be described as
the formulation of association hypothesis for the origins
of all measurements and the computation oftheir probabilities
[105]. While Chong et al. [105] gave a great overview
ofMHT approaches over the last forty years, Emami
et al. [96] in particular covered machine learning methods
for MHT. They start by reviewing research solving MHT
from an " end-to-end " fashion so that a tracker would output
filtered tracks, combining data association and state
estimation, based on noisy measurements as input.
Thereby, they distinguish approaches based on data-driven
association, RNNs as well as deep generative models (e.g.,
variational autoencoders, GANs, and normalizing flows).
Further on, they discuss the use of machine and deep
learning in order to learn discriminative features for data
association. The idea behind learning features is that a
machine or deep learning model can output association
scores, which can be used in traditional algorithms to
compute assignment costs for data association. In the context
of machine learning, in particular, boosting
approaches, SVMs as well as metric learning is addressed.
Relying on deep learning approaches, the focus is especially
on CNNs or advancements as, for instance, Siamese
Networks. However, these deep learning approaches are
usually applied in the context of visual tracking, which
goes beyond the scope of our investigations for tracking
with radar sensors.
TARGET STATE ESTIMATION
Target state estimation aims at the approximation of position,
velocity, and acceleration for x, y, and z coordinates
of the tracked target. Conventionally, a target tracking
Kalman filter estimates that state [34]. In order to accurately
track maneuvering targets it is common to run multiple
filters in parallel as, for instance, within the
interacting multiple model (IMM) Kalman filter [106].
Research addressing this field with AI approaches can be
particularly classified into two streams. On the one hand,
authors aimed at enhancing conventional target state estimation
with the aid of neural network approaches by
reducing the error of conventional approaches [34], [35],
[44], [45], [105]. On the other hand, the literature proposes
AI approaches, which establish a complete target state
estimation [23], [24], [43], [107], [108].
Describing the so-called neural network aided Kalman
filters, Chin [35] represented one of the earliest
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
