Page 162 - ITU Journal Future and evolving technologies – Volume 2 (2021), Issue 2

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 2

The data in Table 1 is used for creating the training set mean and standard deviations provided in Table 1. The
that is utilized developing the ML models for classi ica‑ NA entries in Table 1 indicate that there are no values
tion. Another Table 2 is provided that contains the size, present for the training features of that particular target
maximum velocity and light altitude of popular rotary type.
and ixed‑wing UAVs. Fig. 8 shows the Probability Density Functions (PDFs) of
the length of the central section and maximum velocity of
5. CLASSIFICATION, LOCALIZATION, AND different types of training targets based on the parame‑
TARGET TRACKING ters provided in Table 1. Even though the PDFs of differ‑
ent training targets overlap, the used ML models take sev‑
In this section, details of the classi ication, localization, eral other features into account that yields to high classi‑
and target tracking models used in our approach are pro‑ ication accuracy. Visual boundaries among the different
vided. types of targets for length of the central section and max‑
imum velocity are given in Fig. 9. It is observed that even
5.1 Classi ication of a target using training with only two features, target classes become quite sepa‑
(Tr)
data rable. Let M represent the matrix containing training
data of eleven targets that is given as
Any lying object that disrupts the path of the laser beams
is considered as a potential target in the proposed ap‑ (c,C 1 ) (c,C 1 ) ℎ (c,C 1 ) (w,C 1 ) (w,C 1 ) (t,C 1 )
1
proach. These objects can be man‑made or birds. Flying ⎡ (c,C 1 ) 1 (c,C 1 ) 1 (c,C 1 ) 1 (w,C 1 ) 1 (w,C 1 ) 1 (t,C 1 )

objects are classi ied based on the features discussed in ⎢ 2 2 ℎ 2 2 2 2
Section 4, i.e., shape, maximum velocity, pitch and drift ⎢ ⋮ ⋮ ⋮ ⋯ ⋮ ⋮
⎢ (c,C 1 ) (c,C 1 ) (c,C 1 ) (w,C 1 ) (w,C 1 ) (t,C 1 )
angle characteristics, and maximum light altitude by de‑ ⎢ ℎ
veloping ML models using training data details provided ⎢ (c,C 2 ) (c,C 2 ) ℎ (c,C 2 ) (w,C 2 ) (w,C 2 ) (t,C 2 )
1
1
1
1
1
1
in Table 1. Target types are grouped into four categories ⎢ (c,C 2 ) (c,C 2 ) ℎ (c,C 2 ) (w,C 2 ) (w,C 2 ) (t,C 2 )

2
2
2
2
2
2
based on the 3D shape (Fig. 7), i.e., drone‑like objects, ⎢ ⋮ ⋮ ⋮ ⋯ ⋮ ⋮
⎢
chopper‑like objects, ixed‑wing typeobjects, and missile‑ ⎢ (c,C 2 ) (c,C 2 ) ℎ (c,C 2 ) (w,C 2 ) (w,C 2 ) (t,C 2 )
like objects. The dimensions of the 3D shapes are pro‑ ⎢ ⋮ ⋮ ⋮ ⋯ ⋮ ⋮
vided in Table 1. The irst category is assigned to multi‑ ⎢ ⋮ ⋮ ⋮ ⋯ ⋮ ⋮
rotor UAVs that have either a square or rectangular cen‑ ⎢ (c,C K ) (c,C K ) ℎ (c,C K ) (w,C K ) (w,C K ) (t,C K )
⎢
1
tral section from the mainframe and mounted rotors on‑ ⎢ (c,C K ) 1 (c,C K ) 1 (c,C K ) 1 (w,C K ) 1 (w,C K ) 1 (t,C K )

board the mainframe. The 3D shape of the irst category ⎢ 2 2 ℎ 2 2 2 2
⋮
⋮
⋯
⋮
⋮
is shown in Fig. 7(a). The second category for helicopters ⎢ ⋮ (c,C K ) (c,C K ) (w,C K ) (w,C K ) (t,C K )
(c,C K )
with a square central section due to the rotor blades fol‑ (Tr) = ⎣ ℎ
lowed by a tail section shown in Fig. 7(b). The tail sec‑ M (t,C 1 ) (C 1 ) (C 1 ) (C 1 ) ℎ (G,C 1 )
tion is not present in some cases for Category 2. Central 1 (t,C 1 ) 1 (C 1 ) 1 (C 1 ) 1 (C 1 ) ℎ 1 (G,C 1 ) ⎤
sections of objects in Category 2 are expected to be larger 2 ⋮ 2 ⋮ 2 ⋮ 2 ⋮ 2 ⋮ ⎥
⎥
than central sections of objects in Category 1. In Fig. 7(c), (t,C 1 ) (C 1 ) (C 1 ) (C 1 ) (G,C 1 )⎥
a third category is shown that covers all the aerial tar‑ (t,C 2 ) (C 2 ) (C 2 ) (C 2 ) ℎ ⎥
(G,C 2 )
gets with central, wings, and tail sections. These include 1 1 1 1 ℎ 1 ⎥
ixed‑wing UAVs and planes, cruise missiles, and birds. (t,C 2 ) (C 2 ) (C 2 ) (C 2 ) ℎ (G,C 2 )⎥
2
2
2
⎥
2
2
The fourth category represents targets with only a long ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
main central section without any signi icant wings and (t,C 2 ) (C 2 ) (C 2 ) (C 2 ) ℎ (G,C 2 ) ⎥

tail spans (Fig. 7(d)). The fourth category includes ballis‑ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
tic missiles, rockets and artillery shells, and Hypersonic ⋮ ⋮ ⋮ ⋮ ⋮ ⎥
⎥
Glide Vehicles (HGVs). (t,C K ) (C K ) (C K ) (C K ) ℎ (G,C K ) ⎥
The parameters of the training data in Table 1 consist 1 (t,C K ) 1 (C K ) 1 (C K ) 1 (C K ) ℎ 1 (G,C K )⎥
of the length, width, and height of the central section of 2 ⋮ 2 ⋮ 2 ⋮ 2 ⋮ 2 ⋮ ⎥
⎥
eleven different targets (i.e., classes), grouped into four (t,C K ) (C K ) (C K ) (C K ) (G,C K )
categories. The wingspan, wing width, tail span, and tail ℎ ⎦
width values are also provided. The height of the wings (c,C 1 ) (c,C 1 ) (c,C 1 ) (11)
and tail sections have small variances among different where 1 , 1 , and ℎ 1 are the irst points repre‑
types of targets, hence, these two features are not in‑ senting length, width, and height of the central section, re‑
(w,C 1 )
(w,C 1 )
cluded in the training set to decrease the statistical noise spectively, belonging to the irst class. 1 , 1 are
as low as possible. Other features used while creating the irst points of the wingspan and wing width of the irst
training set are maximum velocity, pitch and drift angles, class, respectively, and the irst points of the tail span
(t,C 1 ) (t,C 1 )
and maximum altitude values of eleven different types. and width are represented with 1 , 1 . Moreover,
Each and every feature that is used creating the training the irst points of maximum velocity, pitch and drift an‑
data is assumed to have a Gaussian distribution with the gles, and maximum light altitude are represented, re‑
spectively, with (C 1 ) , (C 1 ) , (C 1 ) , ℎ (G,C 1 ) , for the irst class.
1
1
1
1
148 © International Telecommunication Union, 2021