IEEE - Aerospace and Electronic Systems - March 2021 - 11

Wang et al.

Figure 5.
Categorization of vision-based approaches.

camera to detect the UAS and neural network to identify
the UAS [49]. An outstanding research is presented
in [50]. A lightweight and fast algorithm which could
operate on embedded system (Nvidia Jetson TX1) and
identify the UAS in movement.
A real-time vision-based UAS detection system is
designed which is based on two vision processing platforms: FPGA-based platform, which can operate below
10 W (i.e., power saving), and graphics processing unit
(GPU)-based platform, which is able to process more
frames. However, for FPGA, it is impossible to change
algorithms in real time [51]. Muhammad et al. compared
different CNNs' performance in detecting UAS, and their
results showed that the visual geometry group (VGG 16)
network with Faster R-CNN achieved outstanding performance [52]. An approach is proposed to combine different
pictures to generate synthetic images to extend the image
data set to train the CNN to enhance the performance of
the UAS detection [53].
Birds are a serious factor which downgrades the
identification of UAS from the images. Significant
research efforts have been made to use CNNs to
enhance the identification of UAS. Survey about the
challenges of detection of UAS and birds is presented
in [54]. The survey concluded that the neural network
algorithms are promising in identifying UAS and birds.
It compared policy-based approaches and neural network based algorithms for the recognition of birds and
UAS using datasets of videos and pictures. The results
showed that the neural network based approaches can
outperform the policy-based approaches over 100 times
in terms of accuracy and efficiency. A UAS detection
framework is presented in [55], which is based on video
streams and classified the objects into different types
with CNN. The work mainly focused on distinguishing
the birds and UAS in different scenarios.
At the same time, some attempts were made to apply
infrared cameras to identify the UAS. Infrared sensors are
leveraged to detect small variations of UAS in heat to
identify the UAS. The drawback of this approach is that
the heat from batteries has significant effects on result
detection [56]. Different from other research on classifying the frame of images, dynamic vision sensors are
MARCH 2021

applied to capture the rotating frequency of the propeller
to distinguish the UAS from birds efficiently [57].
Currently, the vision-based approaches can be implemented in some specific scenarios to recognize the features of UAS from the environment. The evolution of
deep neural networks stimulated the processing of image
processing which could have multiple positive effects on
the UAS detection in vision field. The real time attempts
showed that the vision approaches have the potentials of
efficiency. However, how to implement the recognized
algorithms in multiple and variable environments is challenging. The novel approaches are supposed to be robust,
adjustable, and precise. The vision-based approaches need
to be robust to the quick variation of the environment. The
image distortion caused by weather change could be mitigated by the multiple level image processors which capture the features of UAS in different spectrum. The
mobility of UAS poses a challenge to the vision-based
approaches, i.e., the images are supposed to be captured
and recognized in different levels of mobility of UAS.
The bioinspired robots are limited in the detection accuracy, which poses the risk of distinguishing of UAS and
birds mistakenly. The enhanced accuracy of the distinguishing can improve the efficiency of UAS detection and
mitigation greatly.

RADAR-BASED UAS DETECTION
Radars have several advantages in detecting airborne
objects compared with other sensors in terms of day and
night operating capability, weather independency, and
ability to measure range and velocity simultaneously.
However, regular radar systems focus on fighting air targets of medium and large size with radar cross-section
larger than 1 m2 , which makes it infeasible to detect
small-size and low-speed UAS in [18] and [58]. It is difficult to detect UAS due to slow speed, since Doppler processing is typically used. Therefore, efforts are in needed to
either develop new radar models or increase the detective
resolution of conventional systems. In this section, we
will discuss three categories of radar-based UAS detection
technologies: Active detection, passive detection, and posterior signal processing.

ACTIVE DETECTION
Typically, there are two ways to increase the resolution of
conventional radar detection systems for UAS surveillance:
utilizing higher frequency carriers and using multiple input-
multiple output (MIMO) beamforming radio front-ends, as
shown in Figure 6.
To utilize shorter wave length, in [59] and [60], Xband and W-band frequency modulated continuous
wave (FMCW) radars are designed for UAS detection.
Their solutions use bistatic antenna and finally convert

IEEE A&E SYSTEMS MAGAZINE

IEEE - Aerospace and Electronic Systems - March 2021

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - March 2021