Attention-Based Road Registration for GPS-Denied UAS Navigation

Matching and registration between aerial images prestored road landmarks are critical techniques to enhance unmanned system (UAS) navigation in the global positioning (GPS)-denied urban environments. Current processes typically consist of two separate stages extraction registration. These two-stage approaches time-consuming less robust noise. To that end, this article, we, for first time, investigate problem end-to-end Aerial-Road Using deep learning, we develop a novel attention-based neural network architecture In model, construct two-branch networks with shared weights map input into common embedding space. Besides, considering features sparsely distributed images, incorporate multibranch attention module filter out false descriptor matches from indiscriminative background order improve accuracy. Finally, results extensive experiments show compared state-of-the-art approaches, mean absolute errors our approach rotation angle translations x- y-directions reduced down by factor 1.24, 1.38, 1.44, respectively. Furthermore, as byproduct, experimental prove feasibility multitask learning simultaneously achieve accurate matching registration, thus providing an efficient UAS geolocalization.