Myronenko M. Models and methods of information technology machine learning of an autonomous unmanned aerial vehicle for video monitoring of terrain

An important scientific and practical task of developing informational intelligent machine learning technology is solved in the dissertation of an autonomous UAV for video monitoring of the terrain under the condition of incomplete data certainty within the framework of a functional approach to modeling the cognitive processes of natural intelligence in the formation and adoption of classification decisions. According to the results of the analysis of the current state and development trends of unmanned aerial systems for video monitoring of the area, it is shown that UAVs are mainly used as relayers of the region images to the ground control station (GCS). It is shown that the main way to expand the functionality and increase the information and/or cyber security of the UAV for video monitoring of the terrain is to provide it with autonomy based on machine learning to recognize ground objects. The application of traditional methods of intelligent data analysis Data Mining, including artificial neural networks (ANN), for the information synthesis of autonomous ORS does not always ensure high functional efficiency of machine learning due to arbitrary initial conditions for the formation of digital images of terrestrial objects; intersection of recognition classes characterizing images of objects in the feature space; the multidimensionality of the dictionary of features and the recognition classes alphabet and the influence of uncontrollable factors. The research was carried out within the framework of the so-called information-extreme intelligent data analysis technology created at Sumy State University, which is based on maximizing the information capacity of the system in the machine learning process in order to adapt the input mathematical description to the maximum full probability of making correct classification decisions. This approach, in contrast to known methods, allows the system to be flexible during retraining due to the expansion of the recognition classes alphabet. At the same time, the decisive rules constructed within the framework of the geometric approach are practically invariant to the multidimensionality of the recognition features dictionary . In the dissertation, for the first time, a method of information-extreme machine learning of an autonomous UAV was developed for the recognition of a ground vehicle with optimization of the interest zone frame pixels quantization level, which allows detecting the contour of the vehicle in order to determine the center of the polar coordinate system on it for the formation of the training matrix. As a result, the decisive rules became invariant to the displacement and rotation of the ground object in the frame of the area of interest. The following scientific results were obtained in the dissertation work: 1) For the first time, a method of information-extreme machine learning of an autonomous UAV was developed for the recognition of ground objects with the optimization of the frame size of the region image, which allows to reduce the influence of uninformative and interfering features of the recognition of the surrounding environment of the ground object. 2) For the first time, a method of information-extreme machine learning of an autonomous UAV was developed for semantic segmentation of the image of the region by optimization according to the information criterion of the weighting coefficients of the RGB components of the ground objects images, which allows to increase the full probability of making the correct classification decisions. 3) The method of information-extreme machine learning of an autonomous UAV for video monitoring of terrain based on a hierarchical data structure in the form of a decursive binary tree has been improved, which makes it possible to build in the machine learning process with a given depth error-free decisive rules based on the training matrix. 4) The method of autonomous video navigation by terrestrial natural and infrastructural landmarks with known geographical coordinates has gained further development, which allows determining the location of an autonomous UAV without using a global positioning network. According to the results of computer simulations, it was confirmed that the developed methods of machine learning make it possible to build operative and error-free decisive rules based on the training matrix. In addition, information technology tools for designing a decision-making support system for the GCS operator, which carries out machine learning and retraining of the ORS of an autonomous UAV, have been developed.