Matching and Localization Based on Deep Learning for Unmanned Aerial Vehicle Images

With the increasing popularity of unmanned aerial vehicles, drone aerial images can be used for image target positioning in many fields. This paper proposes a target positioning method based on deep learning, which aims to determine the position of a specified target in drone aerial images and the world coordinate system. First, this method can be combined with a variety of feature detectors. After extracting feature points, a filter module is referenced to eliminate erroneous feature points with obvious errors. Then, a Graph Neural Network (GNN) is introduced to calculate matching descriptors by letting features communicate with each other to improve matching robustness. An optimal matching layer is used to improve matching accuracy and finally determine the position of the target in the aerial image. Combined with drone positioning, a trigonometric function matrix is defined to calculate the position of the target in the world coordinate system. The effectiveness, versatility and robustness of this method are verified through multiple simulation experiments.