Method for Maximizing the Number of Detected Keypoints on Homogeneous Underlying Surfaces

The developing area of unmanned aerial vehicle (UAV) applications states an urgent problem of positioning in the case of Global Satellite Positioning System (GNSS) or Gyro-Peak Navigation System (GNS) absence. The widespread visual SLAM positioning algorithms are based on the based on keypoints detection algorithms that process in most cases RGB images. However, in the case of processing images of homogeneous underlying surfaces (e.g., forests, deserts, water spaces, etc.), RGB data could be insufficient in terms of the small number of possible detecting keypoints. In this case, multispectral data that contain more electromagnetic radiation wavelengths bands, as well as spectral indices computed using combinations of spectral bands could be essential for effective processing of homogenous areas of the underlying surfaces in terms of keypoints detection. This paper is devoted to developing a novel method of search for a combination of spectral channels that provides the maximum number of detecting keypoints using pretrained DARKFEAT deep neural network on the image of homogeneous underlying surface. The results of the experiments have shown that the number of detected keypoints on average increased 1.44 times in the calculated index images compared to RGB images and, in turn, resulted in an average 1.592-fold increase in the number of keypoint matches in the current and previous images.