V.V. Dudihin1, E.A. Kuzina2, I.Y. Mezhuev3, S.A. Rylov4, I.Y. Yakupov5

1 Moscow State University (Moscow, Russia)
2,4,5 JSC Distant Radiocommunication Scientific Research Institute (Moscow, Russia)
3 Plekhanov Russian University of Economics (Moscow, Russia)
1 dudikhin@spa.msu.ru, 2 ekkuzina@npodr.ru, 3 Mezhuev.IY@rea.ru, 4 srylov@npodr.ru, 5 iyakupov@npodr.ru

The article is devoted to the application of deep neural networks for predicting time series that characterize target trajectories in over-the-horizon radar images (OTH). Such neural networks are effective in cases where traditional methods precision is not sufficient.

Method, described in this work relies on the similarities of an area close to the target in the next radar image to the corresponding areas in the previous images along the same trajectory. This approach in similar to “optical flow” calculation, which is used for tracking object motion across an image sequence. However, existing neural network architectures, such as FlowNet, SpyNet and PWC-Net are based on the “dense” optical flow and thus require definition of sample flow vectors for the entire images. Sample radar images, available for model learning, usually have flow vectors available only for a set of points in the images. Besides, OTH radar images are typically represented in azimuth, range and speed dimensions, which makes it difficult to apply existing architectures designed for optical images.

OpenCV, a popular framework for image processing contains an implementation of “sparse” optical flow based on the Lucas-Kanade method and allows to track the motion of isolated areas in the image. However, this method is primarily designed for optical images and does not support adaptation to the specifics of radar images.

The article provides an example of choosing the type and characteristics of a neural network for the task of trajectory extrapolation. A comparative assessment of various variants of neural network architectures is provided. The possibility of solving the problem using a combination of convolutional and recurrent neural networks is shown.

A significant advantage of the proposed trajectory processing approach has been demonstrated. That is the ability of an adaptive system to effectively determine the most likely options for continuing the target trajectory by learning from data.

Dudihin V.V., Kuzina E.A., Mezhuev I.Y., Rylov S.A., Yakupov I.Y. Extrapolation of trajectories in radar images using recurrent neural networks. Dynamics of complex systems. 2024. V. 18. № 3. P. 14−22. DOI: 10.18127/j19997493-202403-02 (in Russian).

1. Anderson T.V. Statisticheskiy analiz vremennyh rydov, M.: Mir. 1976, 755 s. (in Russian).
2. Boks J., Jenkins G. Analiz vremennyh rydov, prognoz i uprevlenie. Kn. 1. M.: Mir. 1974. 406 s.
3. Matrenin P.V., Khalyasmaa A.I., Potachits Y.V. Recurrent Neural Network-Based Autoencoder for Problems of Automatic Time Series Analysis at Power Facilities. Problemele Energeticii Regionale. 2023. V. 2(58). P. 61–71. DOI: https://doi.org/10.52254/1857-0070.2023.2-58-06
4. Kazakov F.A., Shnaider A.V. Ispolzovanie neyronnyh setey s vremennymi ryadami dlya analiza potokov dannyh. Sovremennye naukoemkie tehnologii. 2021. № 6(2). S. 260–264 (in Russian).
5. Sholtanyuk S.V.  Numerical conditioning of neural network components in time series forecasting. Open Semantic Technologies for Intelligent Systems (OSTIS-2020), 2020. № 4. P. 273–276 (in Russian).
6. Kobzarenko D.N., Mustafaev A.G., Gasanova Z.A., Magomedova D.S. One-dimensional convolutional layers in a neural network for wind speed time series analysis. Computing, Telecommunications and Control. 2022. V. 15(4). P. 98–107. DOI: 10.18721/JCSTCS.15408
7. Veremiev V.I., Kovalev D.A. Modelirovanie algoritmov traektornogo soprovozhdenia radiolokatsionnoy tseli na osnove apparata neyronnyh setey. Izvestia vuzov Rossii. Radioelectronica. 2010. №4. S. 61–67 (in Russian).
8. Badanina N.D., Dudikhin V.V., Mezhuev I.Yu., Novikova N.S., Yakupov I.Yu. Metod verifikatsii ob’ektov na radiolokatsionnyh izobrazheniyah s ispolsovaniem ansambley neironnyh setey. Naukoemkie tehnologii, 2023. № 2. S. 79–85 (in Russian).
9. Savian S., Elahi M., Tillo T. Optical Flow Estimation with Deep Learning, a Survey on Recent Advances. Deep Biometrics. USA: Sprin­ger, 2020. P. 257-287.
10. Motion Analysis and Object Tracking. OpenCV [site] URL: https://docs.opencv.org/2.4/modules/video/doc/motion_analysis_and _object_tracking.html (2023, October 17).
11. Cho K., Van Merriënboer B. Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation. Procee­dings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2014. P. 1724–1734.
12. Hochreiter S., Schmidhuber J. Long short-term memory. Neural Computation. 1997, V. 8(9). P. 1735–1780.
13. Keras: Deep Learning for Humans [site] URL: https://keras.io/ (2023, October 20).
14. Kingma D.P., Ba J. Adam: A Method for Stochastic Optimization. International Conference on Learning Representations (ICLR). 2015.