UAV Visual Localization Based on SFA-Resnet and Contrastive Learning

At present, UAVs mainly rely on global navigation satellite system to achieve high-precision positioning. However, the positioning signals of UAVs often face the threat of electromagnetic interference as well as spoofing attacks. Visual localization, as an important complement to traditional UAV localization methods, has attracted increasing attention from both industry and academia, in the wake of the booming development of deep learning technologies and computer vision. Most current UAV visual localization methods focus on extracting local features or edge features that are helpful for the localization task. These geometric metric-based features are overly concerned with the details of the gain to the localization task or classification task, while often ignoring the macro-semantic information that is more useful for solving the localization task. The rapid development of cities and the fact that the satellite images of the cities that UAVs rely on for visual localization cannot be updated fast enough have posed a great challenge to UAV visual localization methods that simply use geometric feature information. Current deep convolutional neural networks often lose the contextual and spatial information of images while acquiring semantic information. To address the aforementioned issues, a spatial feature enhancement module is proposed to enable the model to retain spatial information that is helpful for solving visual localization tasks while acquiring high-dimensional semantic information. The proposed spatial feature enhancement module utilizes a set of learnable parameters to flexibly assign different weights to different channels within the feature maps. The experimental results indicate that the proposed method achieves an accuracy of 91.88% on the Recall@1 and 92.13% on the AP.