=Paper=
{{Paper
|id=Vol-3742/short3
|storemode=property
|title=Information system for detecting low-flying air targets and predicting support trajectory
|pdfUrl=https://ceur-ws.org/Vol-3742/short3.pdf
|volume=Vol-3742
|authors=Mykhaylo Palamar,Mykhaylo Strembitskyi,Vitalii Batiuk,Andrii Chaikovskyi,Iryna Plavutska
|dblpUrl=https://dblp.org/rec/conf/citi2/PalamarSBCP24
}}
==Information system for detecting low-flying air targets and predicting support trajectory==
<pdf width="1500px">https://ceur-ws.org/Vol-3742/short3.pdf</pdf>
<pre>
                                Information system for detecting low-flying air targets
                                and predicting support trajectory
                                Mykhaylo Palamar1,∗,†, Mykhaylo Strembitskyi1,†, Vitalii Batiuk1,†, Andrii
                                Chaikovskyi1,†, and Iryna Plavutska1,†

                                1 Ternopil Ivan Puluj National Technical University, Ruska str., 56, Ternopil, 46001, Ukraine




                                                 Abstract
                                                 The objective of the paper is to create the information system for computer perception of
                                                 scenes by developing models of computer perception, methods and architectures for adaptive
                                                 processing of video streams in computer vision systems aimed at intelligent data processing
                                                 and its parallelization. The main expected results are as follows: development of methodology
                                                 for recognizing poorly formalized objects in heterogeneous field of attention in real time;
                                                 method for separating partially overlapped objects; investigation of the methodology for quality
                                                 assessment and its development particularly, the selection of the most effective method and
                                                 improvement of the existing one for quality assessment in computer vision systems, as well as
                                                 the development of hardware-oriented method for parallelizing video streams based on
                                                 thermal representation of algorithms; method for synthesizing computer vision system
                                                 architectures based on the spatial-time algorithms display.

                                                 Keywords
                                                 trajectory, definition, image, identification, computer vision 1



                                1. Introduction
                                   The development of image recognition systems remains complex theoretical and
                                technical problem. Image recognition is used in various fields, including military, security,
                                and digitization of various analog signals (e.g., automobiles with image recognition).
                                   The perception of phenomena in the form of images plays significantly important role
                                in the processes of cognition of the external world. In the preces of biological evolution,
                                many animals solved the problem of image recognition quite well due to their visual and
                                auditory apparatus. As follows from the definition of “recognition” of new objects itself ,
                                the learning process precedes it. During learning, creatures get acquainted with a certain


                                CITI’2024: 2nd International Workshop on Computer Information Technologies in Industry 4.0, June 12–14, 2024,
                                Ternopil, Ukraine
                                ∗ Corresponding author.
                                † These authors contributed equally.

                                   palamar.m.i@gmail.com (M. Palamar); m.strembitskyy@gmail.com (M. Strembitskyi); vtl.btk@gmail.com
                                (V. Batiuk); chaikovskyi@gmail.com (A. Chaikovskyi); iradenysiuk.70@gmail.com (I. Plavutska)
                                   0000-0002-8255-8491 (M. Palamar); 0000-0002-5713-1672 (M. Strembitskyi); 0000-0003-3190-4101 (V.
                                Batiuk); 0000-0002-0684-2052 (A. Chaikovskyi); 0000-0002-9373-0330 (I. Plavutska)
                                          © 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).




CEUR
                  ceur-ws.org
Workshop      ISSN 1613-0073
Proceedings
number of objects and, in addition, receive information from the source about what image
each of these objects belongs to. This process is called “learning with a teacher” [1].
   More general is “learning without a teacher”, where the system learns spontaneously
and performs the assigned task without the intervention of the "teacher". Machine
learning without a teacher is formulated as the problem of cluster analysis [2]. For such
case, the sample of the objects is divided into clusters (sets having empty intersection) in
such a way that each cluster consists of “similar” objects, and different clusters are
“significantly” different from each other [3].
   Clustering is often used as an auxiliary means for solving classification and regression
analysis problems. Some algorithms of classification problems solution combine both
learning with a teacher and learning without a teacher [4].
   Traditionally, image recognition tasks are included in the range of artificial intelligence
tasks. Two main areas are distinguished:

   -        study of recognition abilities possessed by living beings, their explanation and
modeling [5];
   -        development of the theory and methods of constructing devices designed to
solve specific problems for applied purposes [6].
   The formal statement of the problem of image recognition is the assignment of the
original data to a certain class by means of the selection of significant features
characterizing this data from the total mass of non-essential data. While stating the
recognition problems, they try to use mathematical language
   For optical image recognition, the method of sorting the object view at different angles,
scales, shifting, etc can be applied. Another approach is related to finding the contour of
the object and investigating its properties (connection, presence of corners, etc.).
   Another approach is to use artificial neural networks (multilayer perceptrons,
quantization networks, Kohonen maps, recurrent networks [7-9]).
   It is known that information systems for the detection of low-flying air targets use high-
performance technologies that ensure prompt response to various objects in the air, and
the introduction of artificial neural networks [10] in the feedback loop improves the
coordination of trajectory tracking, the accuracy and speed of their detection.
   The article [11] presents the results of a qualitative study of a neural network,
including discrete and distributed time delays. A method for calculating the exponential
decay rate for a neural network model based on differential equations with a discrete
delay was developed and applied [12, 13].
   In the development of information systems for the detection of low-flying air targets,
the direction of using sensors [14, 15] is promising, in particular for tracking the
trajectory, accuracy, speed of detection of air targets and for assessing the health of
operators. An important characteristic of various types of biosensors is stability [16-18].
Scientific studies [19-21] provide examples of modeling sensor responses. Numerical
modeling in cyberphysical biosensor systems [22, 23] is important at the stage of their
design.
   Let's divide the recognition procedure into separate stages:
   1.       Image perception (obtaining the values of the object characteristic properties).
   2.     Pre-processing (removing noise, presenting images in black and white colors,
cropping unnecessary parts of the image).
   3.     Characteristics allocation (measurement of the object characteristic
properties).
   4.     Classification (decision-making).

2. Development of the image recognition system
   While developing the recognition system, the following stages are involved:
   1.      Obtaining a training sample (training collection, training sample).
   2.      2. Sample of the object representation model.
   3.      Selection of significant characteristics.
   4.      .Development of the classification rule.
   5.      Learning the recognition system (the learning algorithm “collects experience”
on the basis of the recognition sample, in order to set correctly the coefficients of the
recognition system, the learning algorithm is applied to the training sample, controlling
the result of the algorithm).
   6.      Checking the learning quality.
   7.      Optimization of the recognition system.

3. Mathematical representation
    Let's consider clustering images using the distance function. As the classification
criterion, we use the approach based on the classification of images by minimum distance
criterion. The case of one standard or the nearest neighbor method. In certain tasks, the
objects of several classes (images) tend to be grouped around the certain object that is
typical or representative of the corresponding image. A typical example is our case (when
the object has typical contours, boundaries, and scaled dimensions).
    Let us consider M classes that allow images using reference representatives z1,…,zM..
The Euclidean distance between arbitrary vector x and these images is calculated by the
following formula

                          𝐷𝑖 = ‖𝑥 − 𝑧𝑖 ‖ = √(𝑥 − 𝑧𝑖 )′ (𝑥 − 𝑧𝑖 )                                   (1)

   Vector x belongs to the class ɷi, if the condition Di<Dj is satisfied for all j≠i. Let's
transform the formula for calculating the distance

                                                                                       1
𝐷𝑖2 = ‖𝑥 − 𝑧𝑖 ‖2 = (𝑥 − 𝑧𝑖 )′ (𝑥 − 𝑧𝑖 ) = 𝑥 ′ 𝑥 − 2𝑥 ′ 𝑧𝑖 + 𝑧𝑖′ 𝑧𝑖 = 𝑥 ′ 𝑥 − 2(𝑥 ′ 𝑧𝑖 + 𝑧𝑖′ 𝑧𝑖 )   (2)
                                                                                       2

   The last formula shows that choosing the minimum distance to the class is equivalent
                                          1
to maximizing the value𝑥 ′ 𝑥 − 2(𝑥 ′ 𝑧𝑖 + 2 𝑧𝑖′ 𝑧𝑖 ). Therefore, we define the decision functions
as follows
                                                     1
                                𝑑𝑖 (𝑥) = 𝑥 ′ 𝑧𝑖 − 𝑧𝑖′ 𝑧𝑖 ,            i=1,2,…,М.
                                                     2


   where 𝑑𝑖 (𝑥) is linear decision functions. If we put

                                     ɷ𝑖𝑗 = z𝑖𝑗 ,                  j=1,2,…,n.

                                              1                                           (3)
                                    ɷ𝑖,𝑛+1 = − 𝑧𝑖′ z𝑖
                                              2
        .
   І x=(x1,…,xn, 1), then the decision functions are written in the following form

                          𝑑𝑖 (𝑥) = 𝑤 ′ 𝑥.                             i=1,2,…,М.

where wi=(wi1,…,win, wi,n+1).

   In the learning process, all points and their belonging to the corresponding images
а1,…,аm. are memorized
   Recognition means that х, which is to be recognized, is calculated by the potentials of
each image, i.e. the sum:

                                            1
                                Φ𝑖 (𝑥) =         ∑𝑎∈𝑎𝑖 𝜑𝑎 (𝑥),        і=1,2,…,m,
                                            𝑛𝑖


  Where m is number of different images, n is the number of points of the corresponding
                                                 1
image used in the learning process, 𝜑𝑎 (𝑥) =     2    is the is the potential created by
                                                         1+𝛼𝑅 (𝑎,𝑥)
point а at point х.
   We compare the values of Φ1 (𝑥), Φ2 (𝑥), … , Φ𝑚 (𝑥) and the point х refers to the image
with the highest potential.
                                                                                        ̅̅̅̅̅
   In general, case the shape code is written in the form x 1, x2,…, xn, де xi∊{0,1}, i=1, 𝑛, n is
the number of receptor field elements.
   The point in the receptor space is called the boundary point of the set if its code
contains at least one digit, which change transfers the point to another set. Otherwise, the
point is called the interior point.
   If the above mentioned conditions are met for a set of points, we assume that such
points form the compact set: the number of boundary points is small compared to their
total number; any two interior points can be connected by sufficiently smooth line that
passes only through the interior points.

4. Experimental results and discussion
   Graphical presentation of the results of finding the maximum value for the object’s
vicinity and contour definition.
Figure 1: Graphical representation of the results of determining the object contours by
maximum levels.

    Special software has been developed for receiving video data from the thermal imaging
camera via USB port to personal computer. During the process of work, the video data,
frame-by-frame video data analysis with the definition of the shape contour for each frame
is carried out, and the trajectory of the monitored object’s further movement is predicted.
If the coordinates of the predicted movement of the object differ from the actual value, the
trajectory is recalculated.
    The task is to distinguish in this binary image the connected boundaries of the target
object contour among the noise and boundaries of competing sets of pixels. To reduce the
noise effect and minimize the amount of data, the a grid with a certain fixed spacing is
superimposed on the binary image, the value of which is determined by the required
algorithm speed and approximation accuracy, as well as the relative sizes of the target
objects. This this makes it possible to get rid of boundary breaks and noise smaller than
the grid spacing.
    At the next stage, complete mathematical description of the connected elements of the
boundaries represented in the image is performed. In order to do this, we developed
graph structure and implemented by software the method of its dynamic replenishment
during exploratory search for connected local extremes of the image gradient. The search
was carried iterative in the 8-connected domain by graph optimization - straight elements
without branches are represented by single edge (Figure 2).
                      a)                                               b)

Figure 2: Recognition of object contour at different positions in space: a – object to the left
of the observer, b – object to the right of the observer.

   For each constructed connected boundary graph, the cycle search is carried out. This
cycle covers and is the closest one to the image area known as that one belonging to the
object. The speed of the algorithm operation is very significant, so it is inefficient to search
for all possible cycles and choose the one that meets the requirements. The search
procedure immediately moves along the graph in such a way that the first found cycle
covering the area of interest is also the closest one.
   The essence of the paper is to provide support by controlling the rotary platform where
the monitored device (such as thermal camera) is located. The task of the motion control
subsystem is to control the rotary device in order to keep the identified object in the
center of the screen.
   It is obvious from Fig. 3 the system includes:
   – control object (rotary platform);
   – technical vision system;
   – subsystem of optimal control mode selection;
   – system of visual information processing;
   – motion control subsystem.




Figure 3: Model of adaptive control system for rotary system
   The optimal mode is chosen by finding the minimum of the objective function Z(P,D)
for each moment of mode change. Thus, this research area makes it possible to automate
the process of detecting low-flying air targets and predict the supporting trajectory.

Conclusions
   The system for recognizing objects by thermal images has been developed. Special
software processes the data from the thermal imaging camera, calculates the deviation of
the image from the observation center, and sends control action to the motion control
subsystem. The effectiveness of the proposed system has been confirmed by a number of
experimental investigations regarding the support of dynamic objects with uncertain
motion trajectory.
   The scientific novelty of the given investigation is the decision-making in the intelligent
information system which takes into account the multistage nature of solving the
problems of manipulating objects and navigating the system. The decision-making system
perceives changes in the environment where the monitored object is located and produces
such schemes for solving problems that would correspond to the changes in the operating
environment. The above mentioned approach requires the occurrence of a new property
of decision-making systems for detecting and supporting the objects, namely, the ability to
rebuild (adapt) flexibly the operation depending on changes in the environment, changes
in the goal or individual subgoals, in the state of the information system itself. The
applicaton of visual control methods makes it possible to adapt the operation of the
devices due to the widespread use of visual information about the state of low-flying air
targets and prediction of the supporting trajectory.

References
[1]   Aurélien Géron. Hands-On Machine Learning with ScikitLearn and TensorFlow.
      O’Reilly Media, 2017, 751 p.
[2]   Nikhil Ketkar. Deep Learning with Python: A Hands-on Introduction. Apres, 2017,
      226 p.
[3]   T. Hastie et al., The Elements of Statistical Learning, Second Edition, Springer
      Science+Business Media, 2009, 745 p.
[4]   Paul Viola. Robut real-time object detection / Paul Viola, Michael J. Jone // Proc. of
      IEEE Workhop on tatitical and Computational Theorie of Viion. – 2001.
[5]   Richard S. Sutton, Andrew G. Barto. Reinforcement Learning : An Introduction MIT
      Press, Adaptive Computation and Machine Learning Ser.: 2018. - 552 p.
[6]   W. N. Venables, D. M. Smith. An Introduction to R. Notes on R: A Programming
      Environment for Data Analysis and Graphics - Електронна книга, адреса доступу:
      http://www.ievbras.ru/ecostat/Kiril/R/Biblio_N/R_Rus/Venables.pdf
[7]   M. Strembitskiy Selection of the efficient video data processing strategy based on the
      analysis of statistical digital images characteristics / Mykhailo Palamar, Myroslava
      Yavorska, Mykhailo Strembitskyi, Volodymyr Strembitskyi // Scientific journal of
      the Ternopil national technical university, - Ternopil, 2018. № 3 (91) – P.107-114.
[8]    M. Strembitskiy Training of convolutional neural networks for the construction of a
       computer vision system / Mykhailo Palamar, Mykhailo Strembitskyi, Andrii
       Chaikovskyi, Yuriy Pasternak, Volodymyr Kruglyov // Proceedings of the Ⅳ
       International Scientific and Technical Conference "Theoretical and Applied Aspects
       of Radio Engineering, Instrumentation and Computer Technologies "dedicated to
       the 80th anniversary of the birthday of Professor Yai Prots, - Ternopil, 2019. - P.
       210-213.
[9]    Yevseiev, S., Ponomarenko, V., Laptiev, O., Milov, O., Korol, O., Milevskyi, S. et al.;
       Yevseiev, S., Ponomarenko, V., Laptiev, O., Milov, O. (Eds.). Synergy of building
       cybersecurity systems. Kharkiv: РС ТЕСHNOLOGY СЕNTЕR, 188, 2021.
       doi: http://doi.org/10.15587/978-617-7319-31-2
[10]   P. Mandal, L. Roy, S. Das. Classification of flying object based on radar data using
       hybrid Convolutional Neural Network-Memetic Algorithm. In Computers and
       Electrical        Engineering        2023.       Vol.      107,     p.       108623.
       https://doi.org/10.1016/j.compeleceng.2023.108623.
[11]   V. Martsenyuk, I. Andrushchak, A. Sverstiuk, A. Klos-Witkowska. On investigation of
       stability and bifurcation of neural network with discrete and distributed delays.)
       Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial
       Intelligence and Lecture Notes in Bioinformatics). 2018. Vol. 11127. LNCS, pp. 300 –
       313. DOI: 10.1007/978-3-319-99954-8_26.
[12]   V. Martsenyuk, A. Sverstiuk. An exponential evaluation for recurrent neural network
       with discrete delays. System Research and Information Technologies, 2019. Vol. 2.
       pp. 83 – 93. DOI: 10.20535/SRIT.2308-8893.2019.2.07
[13]   O. Nakonechnyi, V. Martsenyuk, A. Sverstiuk. On Application of Kertesz Method for
       Exponential Estimation of Neural Network Model with Discrete Delays. Mechanisms
       and Machine Science. 2020. Vol. 70. pp. 165 – 176. DOI: 10.1007/978-3-030-13321-
       4_14.
[14]   O. Saiapina, K. Berketa, A. Sverstiuk, L. Fayura, A. Sibirny, S. Dzyadevych, O.
       Soldatkin. Adaptation of Conductometric Monoenzyme Biosensor for Rapid
       Quantitative Analysis of L-arginine in Dietary Supplements. In Sensors. 2024. Vol.
       24, Issue 14, p. 4672. https://doi.org/10.3390/s24144672.
[15]   V. Martsenyuk, O. Soldatkin, A. Klos-Witkowska, A. Sverstiuk, K. Berketa. Operational
       stability study of lactate biosensors: modeling, parameter identification, and
       stability analysis. In Frontiers in Bioengineering and Biotechnology. 2024. Vol. 12.
       Frontiers Media SA. https://doi.org/10.3389/fbioe.2024.1385459.
[16]   A. Sverstiuk. Research of global attractability of solutions and stability of the
       immunosensor model using difference equations on the hexagonal lattice.
       Innovative Biosystems and Bioengineering. 2019. Vol. 3 (1), pp. 17 – 26. DOI:
       10.20535/ibb.2019.3.1.157644.
[17]   V. Martsenyuk, A. Sverstiuk, I. Andrushchak. Approach to the study of global
       asymptotic stability of lattice differential equations with delay for modeling of
       immunosensors. Journal of Automation and Information Sciences. 2019. Vol. 51 (2),
       pp. 58 – 71. DOI: 10.1615/jautomatinfscien.v51.i2.70.
[18] Konovalenko, I.; Maruschak, P.; Brezinová, J.; Prentkovskis, O.; Brezina, J. Research of
     U-Net-Based        CNN       Architectures      for    Metal        Surface      Defect
     Detection. Machines 2022, 10, 327.
[19] A. Nakonechnyi, V. Martsenyuk, A. Sverstiuk, V. Arkhypova, S. Dzyadevych.
     Investigation of the mathematical model of the biosensor for the measurement of α-
     chaconine based on the impulsive differential system CEUR Workshop Proceedings.
     2020. Vol. 2762, pp. 209 – 217.
[20] V. Martsenyuk, O. Bahrii-Zaiats, A. Sverstiuk, S. Dzyadevych, B. Shelestovskyi.
     Mathematical and Computer Simulation of the Response of a Potentiometric
     Biosensor for the Determination of α-сhaconine. CEUR Workshop Proceedings.
     2023. Vol. 3468, pp. 1 – 11.
[21] O. H. Lypak, V. Lytvyn, O. Lozynska, R. Vovnyanka, Y. Bolyubash, A. Rzheuskyi, et al.,
     "Formation of Efficient Pipeline Operation Procedures Based on Ontological
     Approach", Advances in Intelligent Systems and Computing III: Selected Papers from
     the International Conference on Computer Science and Information Technologies
     CSIT 2018, pp. 571-581, September 11-14, 2018
[22] V. Martsenyuk, A. Sverstiuk, A. Klos-Witkowska, K. Nataliia, O. Bagriy-Zayats,
     I. Zubenko. Numerical analysis of results simulation of cyber-physical biosensor
     systems. CEUR Workshop Proceedings. 2019. Vol. 2516, pp. 149 – 164.
[23] V. Martsenyuk, A. Sverstiuk, O. Bahrii-Zaiats, A. Kłos-Witkowska. Qualitative and
     Quantitative Comparative Analysis of Results of Numerical Simulation of Cyber-
     Physical Biosensor Systems. CEUR Workshop Proceedings. 2022. Vol. 3309, pp. 134
     – 149.

</pre>