Improving the noise immunity of visual navigation algorithms based on the use of semantic descriptions of observed scenes

Abstract. The problem of estimating the coordinates of unmanned aerial vehicles (UAVs) using visual navigation in the absence of satellite navigation signals is considered. A camera is installed on the UAV, pointing towards the underlying surface, to assess the position by comparing the current images received on board with a reference image – a map of the area prepared in advance. The aim of the study is to increase the noise immunity and computational performance of visual navigation algorithms by switching from comparing bitmap images to comparing the content of observed scenes. In this case, the content of the scenes is presented in the form of semantic descriptions, including classes of objects, their attributes and the relationships between them. A technique for forming semantic descriptions of observed scenes based on the use of a neural network of the U-net architecture and computer vision algorithms is presented. Identification of scenes observed in the video camera with reference images is carried out using the Jaccard function. It is shown that the use of semantic descriptions increases the noise immunity and computational performance of UAV position estimation algorithms.