VISUAL NAVIGATION OF UNMANNED AERIAL VEHICLES USING SEMANTIC TERRAIN DESCRIPTIONS

The article addresses the problem of visual navigation for unmanned aerial vehicles (UAVs), whichinvolves the automatic determination of the current position of the UAV (coordinates in the ground (local)coordinate system) based on the comparison and identification of descriptions of the current images (CI)received on board with reference descriptions stored in the form of a digital map in the memory of theUAV's onboard computer. The aim of this work is to improve the efficiency of visual navigation methods interms of increasing computational performance, robustness, and accuracy of image identification algorithmsin complex and changing observation conditions by using semantic descriptions of observed scenes.In this work, semantic descriptions are understood as descriptions that include classes of objects observedin the scene, their attributes, and relationships between them. The preparation of semantic descriptions ofthe map is carried out at the pre-flight preparation stage of the UAV using pre-trained neural networks forsemantic segmentation. Semantic descriptions of the received CIs are generated on board the UAV. Theuse of neural network algorithms allows this process to be implemented in real-time for a wide range ofobservation conditions (different times of day and year). The use of semantic descriptions of the map andCI reduces computations compared to traditional pixel-by-pixel matching of raster images. Semantic descriptionsare compared by matching object classes, their attributes, and relationships. The work presentsa general algorithm for visual navigation, the main stages of the methodology for forming semantic descriptions,and the algorithm for comparing and identifying semantic descriptions of CIs and map descriptions.A hierarchical algorithm for comparing and identifying images based on the sequential applicationof semantic and raster descriptions of observed scenes is proposed. It is shown that the use of the procedurefor comparing semantic descriptions of CIs and maps by the classes of objects present significantlyreduces the computations necessary for image identification