IEEE - Aerospace and Electronic Systems - July 2021 - 28

ASurveyofArtificial Intelligence Approaches for Target Surveillance With Radar Sensors
Table 1.
Overview of Common AI Approaches (Clutter Identification (CI), Target Classification (TC), Target Tracking (TT))
Description
AI approach
Bayes Classifier (incl.
Naı¨ve Bayes)
Decision Tree
Ensemble Methods (e.g.,
Boosting, Random
Forests)
k-Nearest Neighbor
(kNN)
Logistic Regression
Neural Network based
approaches
Support Vector Machine
(SVM)
Recurrent neural
networks (RNN)
Probabilistic classifier based on the Bayes
theorem [11].
Assigns classes by sequentially checking
simple conditions concerning the feature
values [14], [15].
Set of classifiers whose prediction results are
combined in order to achieve better results
[17]-[19].
Classifies objects according to the most
frequently occurring class considering its k
nearest neighbors among the training samples
[25].
Probabilistic classifier modeling a binary,
categorical variable applying a logistic
function [11].
Network of units (neurons) connected by
weighted edges according to their
interdependencies [29], [8].
Constructs hyper-planes that best separate the
classes from the training data [36].
Specific type of neural networks especially
used for processing sequential data [40]. RNNs
also include Long short-term memory (LSTM)
neural networks representing RNNs with
advanced internal memory capabilities [41].
Convolutional neural
networks (CNN)
Specific type of neural networks especially
used for image processing [40].
of the forward and backward prediction error and
include all spectral information that can be measured
(e.g., information about the shape and variability of
the Doppler spectrum as well as the Doppler frequency
itself). Since the Doppler frequency is influenced by
the relative direction of clutter movement with respect
to the radar, it can vary widely within the considered
clutter types (except for ground clutter). Consequently,
the mean Doppler shift has to be removed from all
coefficients by heterodyning, which represents a normalization
of the Doppler frequency. Besides the
reflection coefficients, two amplitude related features
were added in order to differentiate between point targets
and distributed clutter, resulting in 11 features
[12], [48], [54]. More in-depth feature analysis has
been performed by [30], who examined three different
feature vectors. The first was composed of the nine
reflection coefficients related to phase variations, the
second was based on six features related to the
28
Exemplary applications
CI [12], TT (own results, cf.
Chapter " Track
Evaluation " ), TC [13]
TC [16], TT (own results, cf.
Chapter " Track Evaluation " )
CI [20], TC [21], TT [22]-[24]
CI [26], TT (own results, cf.
Chapter " Track
Evaluation " ), TC [27]
TT [28]
CI [30]-[32], TC [33], TT [34],
[35]
CI [26], TT [37], TC [38], [39]
TC [42], TT [43]-[45]
TC [46], TT [47]
magnitude variations, including skewness, kurtosis,
two amplitude related features, signal-to-noise ratio,
and normalized magnitude,, and the last considered all
14 different features.
For their classification, the authors investigated two
different types of AI classifiers: Bayes and multilayer
perceptrons (MLPs). In [12], they applied a parametric
Bayes classifier under the assumption of a multivariate
Gaussian distribution of the features. Subsequent
research exploited the Bayes classifier as a baseline for
MLP based approaches, which rely on a modified backpropagation
algorithm [30], [48], [54]. To train and evaluate
the approaches, different compositions of training
and test sets have been examined comprising four main
classes of radar returns: weather, birds, and targets representing
moving object classes as well as ground clutter
representing the only stationary class. However, due to
difficulties arising from the resemblance of heterodyned
reflection coefficients of ground clutter and aircrafts, the
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
