IEEE - Aerospace and Electronic Systems - May 2022 - 17

Swinney and Woods
incidents involved multiple UAS, which were never identified,
flying into restricted airspace over 13 different
nuclear power plants [34]. In 2014, the malicious use of
small UASs also came to the attention of the British
media, with the first conviction in the U.K. made after a
recreational UAS flew within 50 m of a nuclear facility
[35]. Aside from an incident involving Greenpeace
who crashed a superman shaped UAS into a French
nuclear facility to expose vulnerabilities [36], there has
been little media reporting of incidents at nuclear facilities
between 2014 and 2019.
However, a recent freedom of information request
revealed there had been 57 incidents at nuclear power plants
in the U.S. between December 2014 and October 2019, with
85% ofthe incidents being unresolved in terms ofattributing
it to a perpetrator [37],[38]. In September 2019, a swarm of
UASs entered the restricted airspace of the Palo Verde
Nuclear Plant, the largest power plant in the United States
The report stated that five or six small UAS flew around the
protected area for 80 min, indicating that this was not conducted
by a popular consumer UAS, such as the DJI Phantom
that has a much lower flight time [39]. In June 2021,
one or more small quadcopters were reported to have caused
substantial damage to a nuclear facility in Iran [40].
DETECTION AND CLASSIFICATION
This section is split into the main research areas for small
UAS detection and classification including imagery-based
systems including electro-optical and thermal, acoustic,
radar, RF, light detection and ranging (LiDAR), and laser
detection and ranging (LADAR) based systems. For each
sensor type, the review is conducted chronologically ending
with the most recent literature.
IMAGERY
Saqib et al. [41] compared various convolutional neural
networks (CNNs) for UAS detection including pretrained
VGG-16 using transfer learning. VGG-16 is a CNN developed
by Simonyan and Zisserman [42], which is commonly
used for vision architectures. Schumann et al. [43]
presented a CNN that is trained on video, which has been
preprocessed with median background subtraction. The
classifier is able to distinguish the UAS from birds and
other background noise. Rozantsev et al. [44] used 3D histograms
of gradients and a CNN to perform detection
using one camera. They utilize a regression approach for
motion stabilization and produce an open source dataset
for the community. Aker and Kalkan [45] proposed a
CNN for small UAS object detection using background
subtracted images. They show that birds can be distinguished
from small UASs using this method.
Yoshihashi et al. [46] used the deep learning and a multiframe
approach for identifying small objects. They present a
MAY 2022
recurrent correlational network, which can perform detection
and tracking simultaneously using multiple frames. Peng et al.
[47] addressed the need for large datasets to train neural networks
by creating an artificial dataset of photorealistic UAS
images using a process called physically based rendering toolkit.
The new images rendered vary with orientation, camera
specs, background details, and UAS positions/size. The network
is trained using Faster R-DNNand achieves much higher
accuracy with the larger dataset than using a smaller one. Lee
et al. [48] used camera imagery and a CNN to classify the
UAS type. Their detection system is based on a second UAS
and the camera imagery collected from it. The dataset consists
ofgoogle images and the OpenCV library to identify its location
on the image and the make/model ofthe UAS. The system
has an accuracy of 89%. Unlu et al. [49] used general fourier
descriptors as features for UAS detection and classification
from birds. In their work, they show that using these features
with a CNN produces the highest accuracy for classification.
Nalamati et al. [50] considered various CNN architectures,
such as ResNet and Inception. Transfer learning is employed
due to limited data and they show that ResNet using the fasterRCNN
is preferable for higher accuracy results. Coluccia et al.
[51] evaluated the highest performing algorithms produced
for the 2020 drones versus birds competition. Challenges for
video detection and classification were observed when classifying
between birds and UASs at greater distances. The other
problem perceived was with moving cameras. Coluccia et al.
recommend training data to be updated to include greater distances
and for categories, such as real-time detection and computing
complexity be considered in the future.
We will now consider the use of thermal imagery for
small UAS detection and classification. Andrasi et al. [52]
investigated the use of thermal signatures for small UAS
detection. They find that quadcopter style UASs, which are
electrically powered do not produce a significant amount of
heat compared with UASs, which burn fuel consumption for
power. This is due to the efficient nature ofelectrically powered
motors and the air circulation within the quadcopter.
Diamantidou et al. [53] proposed a fusion process using a
neural network. The fused information includes thermal
imaging but does not expand on specific details, such as resolution
[54]. Wang et al. [55] produced a monitoring system
using visible and thermal imagery. They suggest that the
largest issue for this work using deep learning techniques is a
lack of data. They address this using augmentation techniques
and show that systems trained on synthetic data perform
well on real-world UAS images, even those incorporating
complex backgrounds. Svanstrom et al. [54] proposed a multisensor
detection system for UASs. They observe that the
thermal camera, which has a lower resolution has equal performance
to an electro-optical camera.
While research suggests that imagery techniques are
effective, it is worth noting that issues with visual recognition
systems are being noticed in other capacities, even
without the involvement of malicious actors. We only
IEEE A&E SYSTEMS MAGAZINE
17
IEEE - Aerospace and Electronic Systems - May 2022

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - May 2022
